Normal view

There are new articles available, click to refresh the page.
Yesterday — 3 April 2025Reading

How the Drive for Autarchy Caused WWII

3 April 2025 at 18:08

In light of “Liberation Day,” we’re taking this show out from the archives — about 1930’s trade policy and the dangerous search for national self-sufficiency with Nick Mulder, where we discussed his book, The Economic Weapon. (You can listen to part one here.)

In this episode, we’re going to get into the juicy stuff around the late 1930s, the leadup to World War II, and interesting parallels that you might see today with what the US is doing with respect to China, trade and technology.

We address:

  • Why countries yearn for autarchy aka rohstoff freiheit, “raw materials freedom”

  • Why states start wars because of “temporal claustrophobia,” and what it has to do with Japan ultimately siding with the Axis;

  • Parallels between the “ABCD circle” (America, Britain, China, Dutch East Indies) and the semiconductor export controls today;

  • Why having an empire was a liability for Britain;

  • What sanctions had to do with the Czechoslovaks — even with a larger army — falling to the Nazis.

Have a listen on Spotify or iTunes.

ChinaTalk is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Hitler during his visit to the VW factory in 1938
Autarchy fun and games until someone goes and starts a world war

Sanctionomics

Jordan Schneider: So, Mussolini invades Ethiopia, the the League of Nations lowers the economic boom on him, but too slowly to make him lose the war.

Watching the world really economically squeeze a middling power focued minds in Nazi Germany and Imperial Japan.

What was the reaction that you saw in the sources of the leading Axis powers as they saw Italy struggling under the weight of global economic sanctions?

Nick Mulder: Italy is put under sanctions about a month after it invades Ethiopia — so November 1935. And what you can see from that point onward in discussions — in Nazi Germany particularly, but also in Japan in the months and years following that — is an increasing focus on what sanctions would mean for them.

In German, in fact, this already precedes the sanctions against Italy. Of course, they had been exposed to a particularly nasty blockade in World War I — they have that memory; the Nazi ideology is very concerned with food security. So they have plenty of reasons to be focused on that.

In Germany — which at that time also was operating really on a shoestring amount of foreign exchange reserves — this was very worrisome. Germany, of course, was engaged in one of the largest armament efforts ever seen in a capitalist economy in peacetime, as Adam Tooze shows in Wages of Destruction. And their external dependence was massive: n order to run all those steel industries, all the energy, the coal, the oil — you need to import it. And interestingly, because of the Great Depression, imports of these commodities had actually become cheaper because there had been a huge commodity downturn.

So you think that the Great Depression causes trade to collapse and the sanctions will no longer work. Actually, the opposite is true for a number of key commodities. The commodity downturn is so severe that it becomes extremely cheap to source oil, coal, iron ore, textiles, raw inputs for a variety of industries, scrap metal from abroad.

So that is the weak Achilles’ heel that Nazi Germany and Japan (with a very similar industrial structure) both have — and that is what they choose to then start protecting.

And in Germany, there’s a really direct effect of the League’s sanctions against Italy on [Germany’s] thinking. The main body that’s in charge of national defense planning — it’s called the Reich Defense Council (Reichsverteidungsrat) — gets together in early December 1935. And it has all the key people there: Hjalmar Schacht (the Reichsbank chairman and Finance Minister), the heads of the General Staff and planners, [Alfred] Jodl, [Wilhelm] Keitel; and soon, in the spring of 1936, Hitler joins them, too. And at each of those meetings they emphasize, “We need to look at what’s happening to Italy, we need blockade resilience, and we need to figure out how we move not just from a kind of trade-commercial protectionism, but to an economic model that is immune to sanctions.” And what they mean by that: it’s immune to having raw material imports severed; they call that rohstoff freiheit, or “raw materials freedom” — they have total autonomy because they have all the raw materials they need for war.

That becomes the aim of their planning going forward, and the main thing that it manifests in is the famous Four Year Plan that is announced in the spring of 1936 — while the League’s sanctions against Italy are still in effect. And it’s then given a particularly powerful head: Hermann Göring becomes the head of the organization running that, and they have a goal: “within eighteen months we want to be independent in terms of fuel from the rest of the world economy, and within four years we want to be totally ready for an aggressive war of conquest.”

Jordan Schneider: What’s the difference between autarchy and autarky?

Nick Mulder: Yeah — so we spell it different ways. Some people write it with a k — that’s the most common. But you also see it sometimes with ch. And there’s an interesting etymological difference between them.

The older version is autarchy, which comes from autos — so it means to rule oneself, to be independent, to be in command of oneself in one’s own position. And that basically just means that you have autonomy, political or otherwise.

Autarky actually comes from the verb arkhein, which is to suffice or to subsist. And that means that you could actually survive off of your own resources. So that’s a narrower definition.

And the interesting thing at the time (one of the famous Italian economists, Luigi Einaudi, notes this): some states that try and become autarkic — actually have access to all the resources that they need within their own territory — lose the ability to have full independence, because they need to engage in policies that are so radical that they effectively close off lots of options for themselves politically. And that’s actually what I think ends up happening in the 1930s: the road to full self-sufficiency is a road that goes through conquest. And that ends up accelerating this war — that had already been in the air was already very possible — but it ends up bringing on a kind of war that is particularly virulent and aggressive and even genocidal, I would argue, because of some of these dynamics.

Jordan Schneider: Let’s talk a little bit about import substitution and the role of German industry in not only becoming self-sufficient by conquering other people’s lands that had coal and cotton, but also changing the way that they consumed those raw materials in order to make themselves more self-sufficient.

Nick Mulder: Yes — there’s a number of schemes that they launched in order to become self-sufficient, and some of them had already been pioneered in World War I and in the 1920s.

So World War I has this big scientific breakthrough that today still powers a lot of global agriculture and sustains a huge part of the world population: the synthetic fixation of nitrogen, which allows you to make fertilizer using simply oxygen. And it means that you no longer need to use saltpeter and some nitrates that you get out of the ground. You can actually use atmospheric components. That’s one thing.

The other thing is fuel hydrogenation. And IG Farben, which becomes infamous for creating Zyklon B — the gas used in the Holocaust — also is one of the main huge German chemical corporations that pioneers a technique for turning coal into oil. [Coal and oil are] both forms of carbon energy, but coal is a lot harder — it’s kind of as if you imagine (even for people who don’t do chemistry) that you add a ton of water to coal, and you put it under an enormous amount of pressure, and you heat it and actually you get something approximating some sort of oily substance; you don’t need to refine it. It’s extremely energy intensive, very wasteful, and very inefficient — so you need an enormous amount of feedstock and fuel in order to get this reaction going. But it is possible for countries that have only coal to turn it into gasoline and a variety of other things, particularly aviation fuels, through fuel hydrogenation.

So that’s the technology that the Nazis hope is going to make them ultimately independent of imports of oil. And as the League of Nations considers this sanctions measure — extending the sanctions against Italy with an embargo on oil imports — [that technology] becomes very important.

The Japanese also take [hydrogenation technology] over. IG Farben — the chemical corporation in Nazi Germany that is accelerating this with huge subsidies from the Nazi government — also sends people to Japan. And in Manchuria and North Korea, the Japanese have access to huge coal reservoirs, so they build a number of plants — both the Imperial Japanese navy and the Imperial Japanese army have their own competing fuel-hydrogenation projects.

And apparently, the North Korean regime today still has some of these plants in the same places that the Japanese Navy built them in the 1930s; and North Korea has massive coal reserves. So there’s speculation that Kim Jong-un might be able to make it through a full fuel embargo by using, basically, Nazi-era technology.

Jordan Schneider: Wild.

Nick Mulder: The other really wild thing, by the way, is that the main postwar user of fuel hydrogenation is apartheid South Africa. They, too, have the same thing: massive coal reserves. They get put under an oil embargo, and they use [hydrogenation] in order to circumvent that [embargo], partially. And today, actually, the largest fuel hydrogenation plot is in South Africa, owned by the South African state-owned company Sasol. So this is a really interesting afterlife with technology.

Economic Asphyxiation and Pearl Harbor

Jordan Schneider: You started taking us to East Asia, so let’s stay there. How does Imperial Japan’s thinking change post-Ethiopia?

Nick Mulder: Japan is — even more than Italy, I would say — a country that is really on the fence for a long time about what its posture toward the West should be. And this is the general thing that I try to emphasize in the book about the interwar period: the war with Nazi Germany was, to some degree, probably inevitable at some point (it was just in the nature of the Nazi regime that they were going to try and use violence); but the fact that Mussolini ended up fighting on the side of the fascists was already less necessary; the fact that Japan sided with the Axis is even more remarkable.

And there was a much bigger split within Japan about who the opponent should be. It’s equally imaginable that they would have gone to war against the Soviet Union and  remained on the side of the Western allies, Britain and the United States, who they rightly saw as much bigger adversaries.

But one of the things that ends up derailing the Japanese liberals’ (so to speak) or the more pro-Western camp’s plans is the war in China. They, of course, are partially themselves to blame for this, because Japan has already invaded Manchuria with a false-flag operation — in 1931, the Manchukuo Imperial Army becomes a sort of state within a state that ends up undermining the central government and essentially running its own foreign policy.

But by the time it’s 1937 or so — we’re now in the immediate aftermath of the Ethiopia sanctions — the Japanese state is in a situation where it still could go either way. And what ultimately ends up happening is that one camp of its officers in China ends up in a fight with Chiang Kai-shek and the Nationalists. And actually, it seems now (according to most historians of China) that the Nationalists, too, were actually kind of pining for a confrontation at that point: in 1931 China didn’t want war — but in 1937 the calculation seems to have been, on the part of some people in the KMT ruling elite, that Japan was going to get stronger every year; if they were going to fight Japan, better do it now than later.

So interestingly, you can see a whole number of countries and groups in the spirit have this sense of temporal claustrophobia (a term from Chris Clark in his book Sleepwalkers). And it’s not just the Japanese and the Germans — it’s also the Chinese, actually, that want to have a confrontation with Japan sooner rather than later. So that ends up triggering a war in 1937 that is arguably the start of World War II, because it directly carries on into the Second World War.

So that complicates the picture dramatically, and it ends up triggering a slow drift, essentially, of Japan into an anti-Western alliance — because the West ends up siding with the Chinese resistance, because of course, [the West wants] to make sure that Japan doesn’t take over China entirely.

Jordan Schneider: “Temporal claustrophobia” — what’s your take? The causes of World War II are multivariate, and your book is front and center of my mind. So I’m curious: thinking back, in the final moments when Japan is thinking about starting Pearl Harbor, when Hitler is thinking about invading Poland and then invading the Soviet Union — this very human fear that even if our odds are bad now, they’re going to keep getting worse, seems very clear. And instead of reevaluating whether or not you want to play the game that gave you these bad odds in the first place, you decide to take the plunge and roll the iron dice, as the case may be.

Let’s talk about what the US ended up doing after 1937, as we get into 1939, 1940, and 1941 when it comes to economic sanctions on Japan.

Nick Mulder: The US has already been considering economic sanctions on Japan since 1931, since the original invasion of Manchuria and the creation of Manchukuo. But at that time, Herbert Hoover is the president, and he holds back on it. It takes quite a while into the [Franklin Delano] Roosevelt administration, really Roosevelt’s second term, before he decides to start getting tougher on Japan — and he has his famous Quarantine Speech in the fall of 1937.

And after that there are a number of incidents. By the summer of 1938, he for the first time begins to call on American companies to institute what he calls “moral embargoes” — voluntary restrictions by American firms. One of the reasons that he’s doing this is because there are Neutrality Acts in effect, which make it impossible for the US president to discriminate by cutting off trade with one country that’s party to a conflict and not with the other; the Neutrality Acts actually obliged the US government to break off arms trade with both parties to a conflict. So this is very tricky for Roosevelt — he has to negotiate these neutrality acts. And if he declares there is a war going on in East Asia, then China also loses access to American arms. So this is why he needs to first go through the private sector and try and have them do it voluntarily.

Now, at some point, they find ways around it. And by 1939, world war has broken out in Europe, too, with the invasion of Poland, and that makes it a lot easier. And that summer, Japan keeps pushing further and further, not just in northern China against British diplomatic presence, but also into Indochina.

And the other thing is that Japan at that point is even more dependent on US trade and on US exports of these commodities than it was in 1935 and 1936 — because the British empire is totally focused on producing for its own war effort, because it’s fighting against the Nazis and it has prioritized its own colonies. So the British empire goes into essentially full economic lockdown mode; Japan can’t really trade that much with them anymore. So Burma, India — those markets become a lot more difficult to access. So [Japan] becomes more and more dependent on trade with the US.

And then Roosevelt steadily ratchets up the pressure in 1940: he lets his commercial treaty expire, so trade becomes more onerous between the US and Japan. And by the summer of 1940 — after the Nazis have taken over all of Europe, and Japan also pushes into Indochina and is now trying to take over French colonial possessions there, because Paris has fallen to the Nazis — he decides it’s really time to start putting a limit on this. He begins to openly restrict a lot of iron-ore and scrap-metal shipments to Japan.

So these are actually the first full discriminatory economic sanctions. He’s targeting Japan openly. He’s trying to throttle this key raw material, making sure they just cannot produce enough to sustain their war in East Asia. And in the summer of 1941, the situation had escalated further still because Lend-Lease has gone into effect — so the US is now also bankrolling the war effort of the British Empire, of Chiang Kai-shek of the Nationalists, and of a number of other countries. And [Lend-Lease] actually needs to prioritize raw materials for [the US].

So part of the story of economic sanctions against Japan that end up triggering the Japanese attack is that the US cannot simultaneously mobilize its own war industry and keep exporting at the same rate with Japan — there’s just a limited amount of North American raw materials. And this, then, means that even if there hadn’t been really severe restrictions, Japan would have seen some dip in what it would have been able to obtain from the US.

But what Roosevelt ends up doing: he puts restrictions in place in July 1941, and then he leaves on a trip to meet Churchill on this big cruiser where they draft the Atlantic Charter together in August 1941. And while he’s away, Dean Acheson and [Henry] Morgenthau (at the Treasury) actually end up — on their own initiative — wrapping up and increasing the sanctions, making them very hard to take off. They freeze all Japanese foreign assets. The US has not declared war on Japan at all — so these are actions where they openly target Japan’s foreign financial reserves. And they cut off oil supplies — and that’s really the thing that sort of sets the final stage of the temporal claustrophobia in motion.

Jordan Schneider: So let’s play the counterfactual game. There are two fantastic books on this: Eri Hotta’s Japan 1941 and Michael Barnhart’s Japan Prepares for Total War, both of whom hint at the idea that perhaps America could have nudged Japan to go invade the Soviet Union instead by not putting on these sanctions.

I’m curious if you think there was a way in which these sanctions could have been rolled out more deftly, giving Japan more of an exit ramp than they felt they had.

Nick Mulder: I think it’s a very interesting suggestion that they could have pushed them toward invading the Soviets — but ultimately I don’t think it would have made a difference, and it wouldn’t have been a feasible solution for the Japanese leadership. And here’s why.

The core commodity that they are extremely anxious about is oil. They have none of it on the Japanese aisles. They have some technology to turn Korean and Manchurian coal into oil, but it’s still not the full amount they need. And what they really can do: they can import from the US, they can import from Mexico, and then there are a number of other places like Iran, Venezuela — those are the main producers in the world. And finally, there’s only one that’s within a reasonable distance of their own territory: the Dutch East Indies.

And the key factor, I would argue (and not just because I’m Dutch), is the fact that this oil embargo is a three-country embargo — it’s a British-American-Dutch embargo. So that is important because the Japanese have simultaneously been negotiating with the Dutch East Indies over preferential access to oil production from Indonesia — and that would have given them maybe as much as 60% of the entire Dutch East Indies’ oil production, which would have taken care of their basic needs.

But the two important things that happen: one is that the Dutch East Indies government is by that time isolated, because the Netherlands has already fallen to the Nazis; the actual Dutch government is in exile in London. So that means, essentially, the Dutch don’t have a lot of independence anymore because they’re now hosted by Churchill — effectively, the Anglo-American leaders can determine what the Dutch do. The second thing is that the Japanese are so desperate for commercial expansion that they end up over-egging the demands they make to the Dutch East Indies government — and the trade treaty goes nowhere; they don’t get that access. And that ultimately is what makes them realize, “Look, this is an encirclement.” It’s an ABCD encirclement, as the Japanese nationalists call it: America, Britain, China, and the Dutch East Indies — that’s the box that they think that they’re in.

And it bears a really interesting parallel with the current [chip export controls]: I’m not saying we’re in the same situation yet, but the current restrictions on chips (including ASML and the Japanese) are an American-Dutch-Japanese-English embargo, now against China. So Japan and China have just switched roles here — the other three countries are actually the same.

Jordan Schneider: What was the Dutch political economy? How are they thinking about managing their negotiations with Japan in 1940, 1941?

Nick Mulder: They are traditionally neutral, and they have been trying to play that role for a long time. They didn’t participate in World War I. They had no desire to enter World War II. They weren’t really planning to enter on the side of the Franco-British Expeditionary Force. They would have opened their territory if there was a need to, but they were trying to do what Switzerland, and Denmark, and the Scandinavian countries were doing — but those also got  invaded by Hitler, so Switzerland is really the only one to get away safely.

And so this traditional Dutch idea of neutrality was already under threat — and this is what ends up pushing them into joining the Anglo-American oil embargo.

And if you also think about it: Shell — which is one of the main oil producers that is not American in this period, and controls most of the Venezuelan and also a lot of the Indonesian oil production — it’s an Anglo-Dutch firm. It was called Royal Dutch Shell for a reason. It’s like Unilever, one of these Anglo-Dutch capitalist enterprises.

So they’re increasingly drifting into the Anglo-American camp and losing this traditional middle position that they had between Britain and Germany.

When Sanctions Actually Work

Jordan Schneider: Can we say FDR had temporal claustrophobia, too, in starting Lend-Lease? Is this a unified theory of everything?

Nick Mulder: It’s an interesting question, and if you read the accounts of people who’ve recently written about this shift in thinking — like Stephen Wertheim in his book Tomorrow, the World — I think that there is a kind of sense that the whole world had changed for FDR after the fall of Paris.

The summer of 1940 is this moment where, for the first time, one of the original three victors of World War I, a beacon of liberalism in the eighteenth, nineteenth, and twentieth centuries, is under the rule of a new kind of totalitarianism — and that’s when even fairly neutralist Americans, for the first time, become amenable to this idea that “Nazism is really a threat to Western civilization,” and they need to do something. And it’s from the summer of 1940 onward that you, I think, start to see in the American elite an increasing preparedness to use these measures.

And the first place where they do it is Franco in Spain. The experience of using coordinated oil sanctions between Britain and the US also starts in that summer. And one of the reasons, I think, that the US goes into the oil sanctions against Japan in the summer of 1941 so blithely is that the oil sanctions in the summer of 1940 against Franco work really well.

They were extremely small. They imposed them for only a few weeks, and then they lifted them again. And they did it just to prove the point that Spain is entirely dependent on American oil: they only have to stop two tanker ships in the Port of Houston (Spain’s entire oil supply can be provided by six vessels a month) — that was enough of a demonstration to show that Franco better not join the Axis.

And it’s the confidence bestowed by that sanctions’ success in the summer of 1940 — a kind of “almost” deterrence, very light usage that issues a clear threat — that, I think, makes [the US] think that they can do the same with Japan.

And of course, racial attitudes play a role here: they just think that the Japanese, ultimately, are easier to manage than the hot-headed Spaniards, and that ultimately they won’t do [something like Pearl Harbor].

But Japan is much further away, and it actually does have a major oil producer right next to it (the Dutch East Indies) that it hopes it can secure. So the main objective of the Japanese campaign in the winter of 1941, 1942 is the Dutch East Indies’ oil fields. There’s a lot of other useful stuff for them, but the general staff is extremely clear that that needs to be the priority. And in order to get there, you need to conquer the Philippines, you need to boot the US naval bases out of that part of Asia — so a lot of these other things become necessary as a way of getting to Sumatra.

Jordan Schneider: The story of Franco being scared off joining the Axis is illustrative of my biggest takeaway from your book: sanctions are great when you go all in. The half measures — say, “Oh, we’ll do this cute thing and have it be financial sanctions,” or, “We’ll just have this nice little escalatory ladder to slowly try to make our adversary realize that we mean business” — don’t work nearly as well as the times where the countries just say, “No, you’re not allowed to import any stuff, and the stuff you’re not going to be allowed to import is going to be the most important thing for your economy, and we won’t let you get it again until you do what we want you to do.”

And there are a number of moments — particularly with Nazi Germany in the mid-1930s — where that really could have stopped rearmament. And we talked last time about Bulgaria, we talked about Paraguay, potentially Japan in 1931 as well: if the pain spigot was turned all the way on early enough, then maybe you end up not having these horrific, world-shattering eventualities of what World War II brought us.

First, am I wrong? And second, what was it about the 1930s that stopped more aggressive economic actions from being taken earlier on as the tides of revanchism ended up ebbing?

Nick Mulder: So to your first question: you are not wrong — but you are right that maximum sanctions are the best only under certain highly specific conditions.

The two particular factors that are key to expanding the success of oil sanctions against Spain in the summer of 1940: [firstly] Franco has just come out of this really grueling civil war — so reconstruction is paramount, and he’s ruling over a devastated society, and he needs all the resources he can get for reconstruction. So [the sanctions] get him and threaten him at a moment of weakness.

Secondly, he has an alternative — the Axis — and he goes to Hitler and asks Hitler what resources Hitler has for him. And of course, Hitler himself is extremely anxious about his access to these things and has nothing to spare, and says, “Sorry, but you’re going to have to fend for yourself and conquer Morocco or something.” So that’s not exactly an attractive proposition. And ultimately that’s one of the reasons why the Axis is a weak alliance — because they cannot meaningfully compensate for each other’s weaknesses. So this is one of the things that makes the situation for Spain and Japan different.

And Japan still has hopes that they can win in China. It’s the classic story of: you are committed to a war that’s not going anywhere; it’s devolved into a guerrilla war; it’s gobbling up ever more resources (like Afghanistan in 2010 or something) — and the Japanese military keeps telling the leadership, “No, but we need one more surge, and then we can win in China.”

And surely they want to go back to peace in East Asia. They don’t want to have to fight the British Empire, the Royal Navy in Singapore, and the US Navy across the Pacific. But they end up being in this war against an opponent that is now receiving steadily more aid from the West — and the spring of 1941 [brought] Lend-Lease, and then it became clear to the Japanese, “Look, the Chinese are going to be in this confrontation for as long as the Allies want them to be. So we need to get to a deal with the Allies. But they now also are not only funding our opponents, but also turning the screws on us. How is this not already a war against us, essentially?” And that accounts for one of these crazy things — that they declare a war, that they actually know that they stand a very small chance of winning and they’re almost certainly going to lose in the long-run. So that’s one aspect.

The other thing you asked about — what are the factors that are holding back tougher sanctions? Part of it has to do with the states in question needing to come up with these sanctions plans on the fly. They have some studies — I used a bunch of them as source material for my book, and they were really interesting to read because they give you great analysis of different import vulnerabilities, and they’re very useful as inside intelligence accounts of the economic history of the 1930s. But they do not always have a good understanding of the world economy. That’s one thing.

They also have large amounts of interest involved in these international, intercontinental sanctions campaigns. And the main trading partners of Japan are the colonies and the dominions of the British Empire — and they actually are not in favor of sanctions on Japan. Australia, Canada, India, New Zealand all have an enormous amount of trade at stake with Japan, because Japan is the only rich, industrialized country in Asia that can buy their exports. So they either trade with Europeans who are now all at war — and then the only other place for them is the Japanese Empire. So for Britain, having an empire is actually a liability. It prevents them from being able to put tougher pressure on Japan early.

Jordan Schneider: Let’s come back to Nazi Germany. Hitler wasn’t doing so incredible when he invaded the Rhineland, right? And plenty of historians have written the hypotheticals of, “If France just decided to fight, then they would have won the war.”

Do you think the same would have happened if the Germans weren’t confronted militarily but in a more aggressive economic fashion in that time period?

Nick Mulder: Yes, I think it would have caused huge problems for the Nazi regime — and so it is an important counterfactual to ask at certain moments what it would have done.

Some of the vulnerabilities were compensated for: Germany got more raw materials from Southeastern Europe and from Eastern Europe, where they got these preferential trade agreements and were able to bully Balkan states into giving up their resources. But they certainly remained very vulnerable.

The other thing is, just militarily, Czechoslovakia could have decided to fight in 1938, and there’s a very good chance that Germany would have lost. The Czechoslovaks had a larger army than Nazi Germany in the fall of 1938, [but] they are persuaded — and actually forced — by Britain and France to dismantle their border defenses and stand down with their army.

And a role, I think, should be accorded to the just Munich crisis — there, too, you already have backup plans for an economic blockade if there is a war that breaks out, but ultimately, the appeasement argument wins. So it’s not even sanctions — it would have been just basic alliance integrity: if they had just upheld the French, particularly their pact with Czechoslovakia and with the Soviet Union, Germany would have faced a three-front conflict, and it would have been over pretty quickly.

In 1939, if France had invaded on the western front when Hitler invaded Poland — same thing. The German General Staff would have probably refused to implement Hitler’s plan. 

So there are many, many moments where Hitler rolls the dice, and he keeps winning — but every time he does it again, he has to wager everything he has gained up to then. And that’s the story of the radicalization of Nazi Germany.

Heading to Sanctions School

Jordan Schneider: All right. World War II — it started. How do the Allies take what they’ve learned over the course of implementing sanctions in World War I and through the League of Nations, and do the best that they can to try to cut off the Axis from accessing financing and critical raw materials?

Paid subscribers get access to the second half of our conversation. We discuss:

  • How the blockade and sanctions regimes of WWI differed from WWII;

  • The history behind the construction of the United Nations, and how it was tied to calibrating sanctions;

  • The effect of the nuclear age on the relative morals of sanctions and conventional war;

  • Parallels between the Cuban Missile Crisis and today’s tech restrictions against China;

  • What lessons pro-decouplers should learn from this history of sanctions.

ChinaTalk is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Read more

从《还有明天》出发,谈谈什么是真正的激进(上)

图片

The master’s tools will never dismantle the master’s house.

——奥德雷·洛德,“黑人、女同性恋者、母亲、战士、诗人”

亲爱的媎妹:

见字如面!

前两天看了《还有明天》后发现网上很多人抱怨结局不解气、指责女主软弱又窝囊。看到这些以后很多话实在不吐不快,所以飞快写完了这篇“热门影视测评02”,主要想从电影出发谈谈下面几个问题:投票VS斩男,哪一个才是正解?什么是真正的激进?为什么说部分互联网“激进”女权是假激进真保守,假解放真极权?

图片

个人评分:⭐️⭐️⭐️⭐️

如果只看前1小时50分钟,我会说《还有明天》是一部四平八稳的女权主义电影,她就像一个门门不偏科的学生,虽然优秀,但总觉得缺少一点惊喜。而最后十分钟可谓把整部电影的格局提升了不止一个档次:女主角没有像我以为的那样离家出走,而是去了投票站,并在台阶上和她的丈夫(男权代言人)静静对峙。电影以女主的丈夫退避作为结束,紧接着字幕浮现:“我们像握紧情书一样握紧手中的选票”,让人回味不已。

图片
图片

看完以后到网上看大家的影评,发现很多人和我一样喜欢这个带来了「制度性变革」(institutional change)的结局,却也有很多人表示女主常年遭到丈夫的暴力对待,只是投票实在不够过瘾,这怎么能解气?如果只是抱怨结局我还可以理解,毕竟如果单纯从情绪出发,我也希望施暴者能得到更具像化的惩罚。可是,有些人竟然在批评结局的同时将矛头对准女主,批判她为什么不逃走/反杀,投了票还要回去,再被打也是活该。

这种观点错误地将家暴的发生归因于男性个体,认为可以通过逃跑/反杀解决一切,却终究无法回答“娜拉走后怎样”这一最关键的问题。家暴只是结构性压迫的一种延伸,其根本原因还是女人在政治经济文化等方面都严重失权。其实导演已经非常直白地指明了问题的根源:当女主问老板为什么新来的男员工第一天上班就比自己工资还高,结果对方只是轻飘飘一句“因为他是男的啊”。

所以,在女儿问妈妈为什么不逃走时,妈妈才会说“我能去哪儿呢?”——是啊,如果不改变由结构性失权导致的经济弱势地位,就算女人成功“润”了,等待她的就会是更好的明天吗?

图片
女主工龄远超男同事,工资却比他低不少,谁让人家是“男士”呢...

女儿和未婚夫的故事线也在说明这个道理:看到妈妈被父亲家暴,女儿要怎样改变自己的命运?“相爱”能解决问题吗?“找个好男人”能避免挨打吗?“嫁给富人”能安稳生活吗?都不能。女儿只会重走妈妈的老路,因为社会结构早已为女人和男人设置了既定的轨道和结局。而电影结尾的投票恰恰是改变这种既定轨道的第一步。

其实这种针对女主的批判毫不陌生,如今我们几乎可以在任何一个女性受害者的评论区里看到和受害者情感隔离(例如尊重祝福锁死/不管,别死我家门口就行),或是指责对方不够强/不够女权(因此活该被压迫)的声音,且其中不乏“驴”、“哭丧”等非常难听的字眼。这种对女性的任意审判是对失权者的二次伤害,只会使女权主义运动淹没在网络骂战之中,失去推动实际制度性变革的力量。

而我认为当下这种混乱局面就源于很多人对结构性问题的忽视,以及对“Radical”(彻底的、激进的、颠覆性的)一词或有意或无意的错误理解。

判断一个人是否radical,关键要看她对现有结构的态度。若想达成彻底的解放,我们就不可能不去识别、批判并试图推翻那些正在压迫我们的社会系统,而这主要包括:

  • 父权制

  • 异性恋霸权

  • 资本主义

  • 性别二元论:认为只存在两种性别

  • 顺性别霸权/性别本质主义/生物决定论:根据染色体/性器官等生物因素(sex)强制给个人指派某性别(gender),禁止/歧视跨性别。

  • 对性和情欲的污名(Erotophobia):导致普遍的性压抑和荡妇羞辱,且主要针对不以「生育」为中心的性行为,例如同性恋性行为(恐同)/自慰/老年人性生活等。

  • 城乡二元制度:歧视农村女性/女性劳工。

  • 种族主义

  • 殖民主义/帝国主义

  • 健全主义:歧视身体/心理生病的人,且尤其是女人,想想“疯女人”指控。

  • 主流审美霸权:导致身体/外貌羞辱,包括肥胖恐惧(Fatphobia)

  • 人类中心主义:导致气候/环境危机,使弱势群体、尤其是女性的生存环境快速恶化。

……

需要注意的是,以上这些压迫系统之间有着千丝万缕的联系,没有任何一个系统能脱离其她系统独立存在(参见《为了99%的女性而战|一份女权宣言》)。因此,若想真正消除性别不平等,我们就必须要关注以上所有结构问题。


而与之相反,如果一个人自称Radical却忽视结构压迫、将不平等归因于「女人不够女权/反抗力度不够」,或是忽略现实条件、一味逼迫女性彻底自我解放,那么她很可能会:

  • 苛责甚至辱骂受害者

  • 拥护现有结构:

    慕强恐弱、恐同恐跨是家常便饭,我愿称之为“激进男权”😁最近掀起的对房思琪的批判就是绝佳的例子,我不止一次看到有人以女权的名义指责已故作者宣传“弱女叙事”、损害了“大女人”的光辉形象。其底层逻辑是女性只有变强才配获得自由,可是自由本应是「天赋人权」。

    (不过在此地,“获得自由”甚至可能需要被替换为“生存”...)

    这种荒谬的批斗行为亦和健全主义(歧视生病的人)息息相关。然而,正如Audre Lorde所言,“For the master’s tools will never dismantle the master’s house”——主人的工具永远不可能摧毁主人的房子。换言之,你不能用父权的逻辑反父权。

图片
辱骂已经去世的作者和所谓弱者:“弱女叙事”、“四处搭灵棚的哭丧货”。认为有心理疾病的人不是“正常人”,用“它”指代对方剥夺其人性。
图片
槽点多到写不下😃 1. 性别本质主义/顺性别霸权,把生育和女性身份强绑定,认为“指出男权社会下女性生育遭受的痛苦” = 厌恶女身。潜台词不想生育就是厌女/就不是女人,和专盯女人子宫的老男人有异曲同工之妙。2. 拒绝面对现实,无视资本主义父权对生育劳动的剥削,强行否认这是女性失权的重要原因之一。3. 本末倒置,指责“弱女”是性别压迫的根源,把女性被压迫的结果当成原因。4. 优生主义(弱女不配生育)、剥削弱者的劳动力(弱女就该辅助强女带娃),请问这和男人要求女人无偿家务劳动有何区别?爹味简直快要溢出屏幕🤮
  • 抑止女性创作者创作:

    只要导演、作家、博主等创作的内容有一点不合心意就疯狂网暴,这种审判式的攻击不以沟通为目的,通常表现为毫无逻辑的情绪输出,或是大字报式的批斗。通过毫无事实根据的扣帽子(如厌女/自由人/精神男人/ygbt/学术驴...)占据舆论制高点,仿佛只要给别人贴上“不够进步”的标签,自己辱骂对方的行为就自动具备了合法性。这种暴力批斗和「理性讨论/建设性的批评」有本质区别,后者是为了鼓励或帮助创作者更好地创作,而前者的目的只是捂嘴。

  • 精神胜利法:

    无限拔高女人的主观能动性,认为仅靠个人意志就能消除压迫,仿佛只要默念一百遍“我是强女”就真能变强、抄写一百遍“女人是第一性”现实就真是如此了。既然女人已经如此之强,那么何必推动系统性变革?而哪个女性要是做不到反杀自然只能将她开除女籍、骂她“你弱你活该”,进而剥夺其获得哪怕一丝一毫共情与帮助的机会。

图片
1. 迷惑发言:“自然女”不需要养老,完全无视客观现实,且相当于直接否定了无数女性对自己的亲人/丈夫的亲人付出的无偿care labor的价值(图左)2. 苛责遭老师性侵的女性,无视结构性压迫,说她学点“激女思想”就不会去世(图右)

由于字数限制,本次来信只展示文章的前半部分内容。剩余内容将在下一封newsletter中发布~

就此搁笔,期待下一次和大家见面!

陌生女人1号

二〇二五年四月三日

Before yesterdayReading

Is China Racing to AGI?

2 April 2025 at 18:03

An anon asks the question: Is China racing to develop AGI?

U.S. leaders increasingly frame artificial intelligence as a future-defining competition. Secretary of the Interior Doug Burgum warns of an “AI arms race,” while calls for a “Manhattan Project for AI” grow louder. Corporate giants echo this urgency: OpenAI claims China is “determined to overtake us by 2030,” while Anthropic’s Dario Amodei dreams of a US-led unipolar world — only if America speeds up and beats China to AGI.

But this entire “race” narrative hinges on a crucial assumption: that China is sprinting toward AGI as well.1 Is it?

This piece, inspired by Dwarkesh’s AI scaling essay, stages a Platonic debate between two voices:

  • The Believer, who insists China is racing to beat the US to AGI

  • The Skeptic, who doubts China’s focus and urgency

They clash over several key questions: Are Chinese policymakers truly committed to AGI? Do China’s top AI labs see it as feasible and imminent? Are investors pouring serious money into AGI projects? And does DeepSeek mark an inflection point?

Are policymakers AGI-pilled?

Believer:

China’s senior leadership has been all-in on AI for at least a decade. They launched their “big funds” for semiconductors in 2014. In 2017, the State Council set a clear goal: making China the world’s primary AI innovation centre by 2030. A year later, Xi Jinping himself called AI a game-changer, with a "profound impact on economic development, social progress, and the international political and economic landscape." That’s five years before ChatGPT dropped.

Skeptic:

Sure, they’re all in on AI, but that’s not the same as racing toward AGI. Back then, Beijing prioritized facial recognition, surveillance tech, autonomous driving, and industrial automation — narrow, specialized applications. They didn’t exactly throw their chips in the “AGI” basket, spending only $300m USD on their hyped up “AI megaproject”, which ended up just being a mechanism for funding dozens of small research projects, not some giant cluster.

Believer:

Alright, so you’re saying China is into AI but not AGI. But after ChatGPT in spring 2023, the Politburo explicitly stressed "development of artificial general intelligence" and the importance of building an "innovation ecosystem" around it. That’s a shift toward serious long-term AGI ambitions.

Skeptic:

Yes, that was the first time the Chinese top leaders used the term “AGI”, and that should make us think. But the Chinese term here, “通用人工智能,” doesn’t necessarily translate to “AGI” in the sense of “superintelligence” — it can also mean something more mundane like “general-purpose AI.” This was also a few months after ChatGPT exploded onto the scene, so they may have mostly been just thinking of LLMs at the time.

Believer:

But AGI was just about all anyone was talking about then! Even if the term has two meanings, certainly they would have discussed the idea of AI with human-level intelligence if they were bringing up this kind of term in such a major meeting.

Skeptic:

But here’s where it gets interesting. In China’s political system, a big policy announcement like this one gets followed up by clarifications in party-state media. And guess what? A couple of months later, a People’s Daily op-ed directly referenced the Politburo meeting and explained the difference between "specialized AI" and "general-purpose AI" (通用人工智能). This wasn’t about AI that could outthink humans, but more about systems that can do multiple tasks. The article even said general-purpose AI is still in its "early stages". Nobody in Zhongnanhai is feeling the AGI.

Believer:

That was almost two years ago! In February 2025, Gao Wen — head of Pengcheng Lab and the guy who once briefed Xi Jinping on AI — wrote in People’s Daily that we are in a “transitional phase from weak AI to strong AI”, adding that AI is the “strategic commanding height in great power competition”. If that’s not AGI race rhetoric, what is?

Skeptic:

We can also look at how the wider bureaucracy interpreted the Politburo message on AGI. The Ministry of Industry and Information Technology (MIIT) included “AGI” as a focus area in a new funding round. The projects they describe are aiming to set up data centers with just 1,024 GPUs by 2025. That is a rounding error next to U.S. clusters. The MIIT lumped “AGI” with the Metaverse, humanoid robots, and brain-computer interfaces. Yes, the Metaverse.

Believer:

But local governments are jumping in too — Beijing, Anhui, and Guangdong have all announced AGI initiatives, pooling compute and data resources for large model research. They’re discussing roadmaps that go beyond LLMs, exploring brain-inspired AI and causal reasoning models. Doesn’t that show a shift toward AGI?

Skeptic:

A shift? Maybe, in so far as these are the first policies that specifically focus on general-purpose AI. But the fine print still emphasizes relatively boring applications like healthcare, government services, urban management, and autonomous driving — not superintelligence.

Also, just because they want some fancy new approaches to AI to succeed doesn’t mean they will — the field of AI is littered with the bones of research agendas that absolutely would have created AGI, if only they worked.

Policymakers waking up after DeepSeek?

Believer:

You dismissed the 2017 AI plan and the 2023 Politburo meeting. Fine. But DeepSeek has changed the game. Within weeks, CEO Liang Wenfeng met with Premier Li Qiang and even Xi Jinping. That’s serious state recognition.

Skeptic:

Of course these meetings mean something. DeepSeek’s sudden rise injected fresh confidence into China’s economy, and the state would be foolish not to capitalize on that. Getting Liang in the room and state media then saying basically that the Party supports the private sector, and the private sector supports our national goals is a clear signal: the tech crackdown is over.

But that does not necessarily mean Beijing wants to bankroll DeepSeek’s path to AGI. Li Qiang and Xi meet with plenty of tech companies every year–and by the way, the firms who got the most prestigious seats in the meeting were hardware companies like Huawei, Xiaomi, and BYD, not software firms. These are carefully curated PR moments, designed to showcase government priorities and entrepreneurial success stories. DeepSeek certainly checks those boxes. But not every company that gets a handshake from Xi receives a blank check.

Catching Xi’s eye is like drawing Sauron’s gaze: impressive, but rarely ends well. Remember when Premier Li Keqiang cozied up to Jack Ma in 2013? He invited Ma to exclusive symposia, and lauded Alibaba’s contribution to job creation. We all know how that ended — Alibaba got caught in the regulatory crosshairs during the tech crackdown.

Coming back to Liang Wenfeng’s sit-down with Li Qiang. They supposedly discussed the 2025 Government Work Report. And what did Liang walk away with? A mention of the AI+ initiative — something that was already in last year’s report — and a vague nod to "the extensive application of large AI models," plus "AI-enabled phones and computers.” All of this is very application-focused. “AI+ initiative” basically translates into “AI + literally anything but AGI.” Soon we’ll have AI + rice cookers, AI + karaoke machines, and AI + slightly smarter traffic jams.

If Liang went in hoping to convince Li Qiang to push for AGI, it sure looks like he came out empty-handed.

Additionally, the Chinese government remains cautious. Vice Premier Ding Xuexiang, who’s leading China’s new Central Science and Technology Commission, recently emphasized that China “will not blindly follow trends or engage in unrestrained international competition”. That’s not a race mentality.

Believer:

That’s just diplomatic rhetoric, they don’t want to freak out American policymakers even more by saying they’re racing to AGI. Oh, and by the way, Huawei got prime position in that Xi meeting and they’re the one designing China’s AI chips.

Skeptic:

But their regulations speak to this. Remember that Baidu’s ERNIE bot had its first demo version available in March 2023, but had to wait months for the Cyberspace Administration to finalize its genAI regulations and grant Baidu a license before they made it available to the general public. Going through China’s genAI licensing process takes several months for some companies. This suggests a willingness to slow AI deployment for control and social stability. China’s not racing — they’re speed-walking with a leash.

Believer:

Being cautious also doesn’t equal not believing in AGI. Chinese thought leaders have expressed concerns about existential risks from AGI. If you don’t believe in truly transformative AI, there is also no reason to be concerned about x-risk.

Leading AI Labs and their Investors

Believer:

And the AI labs are charging ahead — policymakers or not!

DeepSeek CEO Liang Wenfeng says flat-out, “Our destination is AGI.” Zhipu.AI 智谱AI founder Tang Jie 唐杰 says they are “conducting AGI-related research around superintelligence and superalignment,” and need to “plan for AGI based on large models” with an “aim to lead the world”. Moonshot AI’s CEO Yang Zhilin called AGI “the only meaningful thing to do in the next 10 years”, while Minimax CEO Yan Junjie compared the road to AGI to the Long March.

Skeptic:

Yes, some start-ups are certainly ambitious. But not everyone shares their enthusiasm. Baidu CEO Robin Li, for instance, claims that “today’s most powerful AI is far from AGI” and that we “don’t know how to achieve that level of intelligence yet”. And Li Kaifu’s 01.AI has also backed away from foundational models.

But ambition alone doesn’t sway investors. China’s VC is struggling. DeepSeek’s CEO complained that they want a quick buck for their money, and are hesitant to support those who make true innovation, suggesting he may also struggle receiving the kind of money he wants.

China’s big tech is not helping the most dynamic independent research labs like Microsoft, Amazon, and Google have. Zhipu’s $341 million from Alibaba and Tencent and Moonshot’s $1 billion from Alibaba are peanuts — Microsoft dropped $14 billion on OpenAI. None of the major Chinese start-ups is valued above $3.3 billion. DeepSeek is a bit of a special case because it is entirely funded by its parent High-Flyer and has no external investments, so its value is anyone’s guess. Forbes estimates it at somewhere between $1 billion and $10 billion.

Even if we assume $10 billion — compare that to OpenAI at $300 billion and Anthropic’s $61.5 billion. It’s not a race — it’s a rout.2

Believer:

DeepSeek has proven that China can leapfrog, not just follow. And investors are taking notice.

Just a year ago, Allen Zhu Xiaohu (朱啸虎), a prominent Chinese VC, dismissed AGI advocates as “delusional” and saw “no point” in engaging with China’s genAI startups. Fast forward to February 2025, and he’s changed his tune: "DeepSeek is almost making me believe in AGI."

And it’s not just private VC money pouring in. In January 2025, the Bank of China launched a massive AI support project — 1 trillion RMB ($140 billion) over five years — funding AI through equity, loans, bonds, insurance, and leasing. The launch event was high-profile, co-chaired by major players:

  • Ge Haijiao (葛海蛟), Chairman of the Bank of China

  • Yang Jie (杨杰), Chairman of China Mobile

  • Zhang Peng (张鹏), CEO of Zhipu AI

  • Zhou Bowen (周伯文), Director of the Shanghai AI Lab

  • Li Meng (李萌), former Vice Minister of Science and Technology

Every major ministry was present: the National Development and Reform Commission, Ministry of Science and Technology, Ministry of Industry and Information Technology, and the State-owned Assets Supervision and Administration Commission. All the big tech firms — Huawei, Tencent, Baidu, Ant Group, iFlytek, China Mobile — were there too.

At the announcement, Ge Haijiao didn’t hold back: "Large AI models are profoundly shaping the global political and economic order” and present “a strategic pillar for ensuring national security." He lines up the country’s AI heavyweights, delivers that message, and then drops the bombshell: “And here is 1 trillion RMB.” If that’s not commitment, what is?

Skeptic:

Let’s break down the Bank of China’s announcement a bit further.

First, the total scale: 1 trillion RMB ($140 billion) over five years translates to $28 billion per year. That’s a serious chunk of money. But compare it to Stargate’s intended $100 billion per year, and it starts looking less impressive.

Second, where’s the money actually going? The new funds will support the entire AI ecosystem — not just AGI research. It covers chips, data, and AI algorithms, but also AI + robotics, AI + the low-altitude economy, AI + biomanufacturing, and AI + new materials. These are all important, potentially transformative fields, but they dilute the focus. Meanwhile, Stargate is laser-focused on super-scaling compute for AGI. It’s a concerted effort with resource centralization. This Bank of China project, in contrast, looks more dispersed, just like that earlier megaproject.

Subscribe now

Let’s be realistic about who’s managing this money: Chinese state banks. These institutions are notoriously risk-averse. They favor safe, mature technologies over speculative bets on moonshot research. No way they’ll push in all their chips to fund AGI breakthroughs, which require Masayoshi-levels of risk tolerance.

Believer:

Yes, there’s uncertainty about how exactly the money will be spent. But the Bank of China head specifically named computing centers in China’s eight designated AI “hubs” as a priority. If anything, that suggests a significant chunk of the funds will go toward centralized compute infrastructure.

Chips and Compute

Skeptic:

Let’s talk compute. No one in China is building GPU clusters at the scale of 100,000 chips like in the US. And despite DeepSeek’s efficiency breakthroughs, developing and deploying superintelligence will demand massive scale — far beyond anything China has today.

Believer:

Again, DeepSeek has changed the game.

Alibaba just committed over 380 billion RMB (~US$52.4 billion) in the next three years to build cloud and AI hardware infrastructure. That’s more than their total capital expenditure from 2015 to 2024 combined. In other words, in just three years, they’ll spend more than the last decade. It’s the largest-ever investment by a Chinese private company in AI infrastructure. And Ali CEO Eddie Wu 吴泳铭 made their ambitions crystal clear: “Alibaba's ultimate goal is to achieve AGI. AGI will be able to perform more than 80% of human capabilities. Since 50% of the global GDP consists of human wages, achieving AGI would create the world's largest industry.”

Policymakers are also signaling increased interest in compute. On a recent visit to China’s top mobile operators — who play a major role in compute infrastructure — Premier Li Qiang emphasized that AI is “bringing profound changes to the world” and urged them to optimize compute resource use.

Skeptic:

You have to do the math: $52.4 billion over three years is $17.5 billion per year. In 2024 alone, Amazon spent $77.8 bn, Microsoft 75.6 bn, and Google 52.6 bn.

Chinese compute investment may be growing, but it still trails far behind what Western investors are pouring into their 100,000-GPU clusters. There’s nowhere near the AGI fever in China that we see in the U.S.

Believer:

If China isn’t building 100,000-GPU-scale clusters yet, it’s not because they don’t want to — it’s because they don’t have the chips. U.S. export controls have slowed their access to cutting-edge semiconductors. If they had the chips, they’d build the clusters.

Skeptic:

China does have chips — over 1 million new AI chips were estimated to be added in 2024 alone. Some estimates suggest even more. In theory, that’s enough to build multiple 100,000-GPU clusters. The problem isn’t just chip supply; it’s that these chips are spread across smaller, fragmented clusters, many of which are underutilized. Their chips are smeared like peanut butter on a bagel — thin, uneven, and oddly unsatisfying

The real bottleneck? Not chips — belief in AGI. No investor is stepping up to bet big on AGI-scale clusters. If there were real conviction that AGI was imminent, we’d see someone pulling these chips into centralized compute hubs. But so far, that hasn’t happened.

Believer:

You really think they’re content with their current chips? Look at the $47 billion third round of the “Big Fund” — they’re doubling down on domestic semiconductor manufacturing. They aren’t just waiting around; they’re going all in.

Skeptic:

Sure, but how do you know AGI is the end goal? $47 billion for chips could also be justified by more narrow economic or military concerns, not a full-blown AGI sprint. And that brings us full circle: China is heavily investing in AI — but that doesn’t mean it’s racing toward the singularity.

Conclusions

So, who wins this debate?

Overall, the Skeptic makes the stronger case — especially when it comes to China’s government policy. There’s no clear evidence that senior policymakers believe in short AGI timelines. The government certainly treats AI as a major priority, but it is one among many technologies they focus on. When they speak about AI, they also more often than not speak about things like industrial automation as opposed to how Dario would define AGI. There’s no moonshot AGI project, no centralized push. And the funding gaps between leading Chinese AI labs and their American counterparts remain enormous.

The Believer’s strongest argument is that the rise of DeepSeek has changed the conversation. We’ve seen more policy signals, high-level meetings, and new investment commitments. These suggest that momentum is building. But it remains unclear how long this momentum can be maintained–and whether it will really translate into AGI moonshots. While Xi talks about “two bombs one satellite”-style mobilzation in the abstract, he hasn’t channeled that idea into any concerted AGI push and there are no signs on any “whole nation” 举国 effort to centralize resources. Rather, the DeepSeek frenzy again is translating into application-focused development, with every product from WeChat to air conditioning now offering DeepSeek integrations.

This debate also exposes a flaw in the question itself: “Is China racing to AGI?” assumes a monolith where none exists. China’s ecosystem is a patchwork — startup founders like Liang Wenfeng and Yang Zhilin dream of AGI while policymakers prioritize practical wins. Investors, meanwhile, waver between skepticism and cautious optimism. The U.S. has its own fractures on how soon AGI is achievable (Altman vs. LeCun), but its private sector’s sheer financial and computational muscle gives the race narrative more bite. In China, the pieces don’t yet align.

This essay deliberately avoids comparing U.S. and Chinese progress or assessing whether Beijing could achieve AGI — whether through innovation, espionage, or other means. Our focus is strictly on intent, not capability: what China seeks to pursue, not what it might accomplish.

A final point of caution: this staged debate covers a snapshot of arguments from early 2025. Things are changing fast in AI, and many of the arguments presented here could change as well. While this piece highlighted that there is currently little evidence that China is racing towards AGI, this may of course change at some point in the future. If the government or major investors started radically centralizing resources while changing their tone in public statements, this could signal change. ChinaTalk will keep an eye out!


ChinaTalk is open to sponsorships. Email jordan@chinatalk.media to start a conversation.


ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.

1

Definitions of artificial general intelligence (AGI) are both numerous and contested. For the purposes of this essay, being “AGI-pilled” refers to the belief that highly transformative, general-purpose AI systems — capable of surpassing human performance across essentially all relevant cognitive domains — are not only in principle possible, but likely to emerge within years rather than decades. This belief justifies large-scale investment in the development of such systems. In this sense, leading Western labs like OpenAI and Anthropic can be considered “AGI-pilled.”

2

Zhipu.AI has received additional investments from multiple state-led funds in March 2025. These may somewhat raise its valuation, but are still in the same order of magnitude as previous investments.

《善意的竞争》:只是霸道女总裁爱上我吗?

图片

我们应该能够充分、平等地体验爱,没有羞耻、也没有妥协。

—— Elliot Page,加拿大演员

亲爱的媎妹:

见字如面!

最近我们看了好多不错的影视剧,所以决定出一个“近期热门影视测评”系列。第一期想来聊聊最近大火的《善意的竞争》:剧中的女同爱情是对“霸道总裁爱上我”的简单复刻吗?酷儿关系为何天然具有解放性?什么才是好的女同、乃至女性角色塑造?

图片

个人评分:⭐️⭐️⭐️⭐️

好看好看好看,大家可以放心食用!女同性恋故事➕丰富有趣的女性角色➕新颖的主题➕颇具张力的戏剧冲突➕出人意料的情节发展➕极易代入的父权竞争社会 = 超适合“东亚女权姐妹”体质的上乘佳作。《善竞》最让我惊喜的地方在于把女性议题融入情节之中,让一切自然而然发生,而没有用猎奇的镜头或态度把「非传统女性/不容于父权体系的女性」当作异类来刻画。

其中最典型的例子当属刘在伊和禹瑟琪的同性关系。《善竞》没有沉迷于出柜叙事(并不是说我们不需要这样的叙事),没有把聚光灯疯狂对准双女主的情感关系、暗示女同和“正常人”有壁,更没有任何角色对此大惊小怪——这一切如此平常,平常到根本没有人在意,就好像女人天生就该爱女人一样。

和《善竞》形成鲜明对比的是《好东西》对女同关系的处理。小叶的男朋友误会她和铁梅是女同,之后俩人又顺水推舟装情侣,一路狂抖包袱。虽然这样的片段某种程度上增加了性少数群体的可见度,但它却并不是我想要的。在整个过程中,导演更多地站在异性恋的角度审视乃至凝视性少数群体,利用(关于较“男性化”女同的)刻板印象制造笑点,把女同性恋者定位成一个引人发笑的「符号」、一处供人猎奇的「景观」。在这样的拍摄手法下,真实而具体的「人」消失了,留下的只有阵阵干瘪而扁平的笑声。

幽默是权力的体现,上位者天然具有取笑下位者的权力,这就好像酒桌上男人拿女人开玩笑,后者还要陪笑一样。我在看这个片段的时候真的非常不适,看到弹幕里一溜的“哈哈哈哈”更是如芒在背。我偏不笑,因为真的不好笑!🙂️

而《善竞》正是对这种叙事的反叛。镜头几乎从不站在所谓主流/异性恋的角度审视二人的关系,而是用平视的视角真诚记录:那些校园日常、暧昧心思、拉扯试探、关心依赖、甚至怀疑和恨意,点点滴滴终于织就了细腻坚实的情感之网。

图片

不过,这也不代表我认为《善竞》对女同的刻画是完美的。事实上,感情线的前半段还是有些比较明显的异性恋范式。霸道富家千金爱上落魄但坚韧的底层女孩,让人想起一些古早的异性恋偶像剧,所以我私心不算特别喜欢。可是随着剧情展开,情感开始流动,权力关系并非如想象中那般坚不可摧。我们发现富家千金原来从小就受制于父亲的权威。因此,在伊和瑟琪的爱其实是伴随着她们对「父亲」的反抗生根发芽的。

而这恰恰是《善竞》和经典“霸道男总裁爱上我”叙事的根本区别。在男总裁和女性的互动中,总裁就是「父」本身。他拥有绝对的财富和权力,能英雄救美、一掷千金,而这些也恰恰是他吸引力的来源。所以,女主与之“相爱”的过程,就是心甘情愿地接受“父亲”控制、进入父权体系的过程。

但《善竞》讲述的女同故事却刚好相反:从想要反抗父亲的权威开始,到对“父权”的成功颠覆为止。女主们相爱的过程就是和“父亲”决裂的过程。两个饱受父权压迫的女人,一段不容于世俗的关系,要想自由呼吸,她们怎可能不逃离父亲的控制?

所以,虽然剧集初始在伊因为家境优渥而占据高位,但若想充分地去「爱」,她就必然要与原生家庭(也就是自身的阶级地位和财富)决裂。所以,感情线从中期开始渐入佳境,并在最后一刻达到高潮、幻化成一曲优美动人的交响乐。最后几分钟,瑟琪滑着滑板冲向面朝大海的在伊,那是全片我印象最深的镜头——广阔的海是自由的象征,而正是从她们的反抗、斗争以及胜利中,我再一次深刻地体会到酷儿关系所蕴含的那如日方升的解放性力量。

除了两个女主角以外,剧中其她女性角色也可圈可点。崔京频繁自慰(包括在刘在伊座位上自慰)的剧情探索了女性的性欲以及女性间的嫉妒、情欲、好胜心等种种复杂情感。而对朱艺莉这个符合父权审美、沉迷奢侈品的女孩,剧集也只是刻画了她在面对各种情况时的自主选择,镜头和叙事并无评判的意味。而且,编剧最终抱着善意给了她一个很好的结局,让人看了感到充满希望。(不过联系到韩国女演员的生存现状,以及最近金秀贤的丑闻,我不禁在想这对于艺莉来说会不会只是换了一种更加“体面“的方式被父权压榨。当然这可能有点过于悲观了!)

图片

不把女性当成“异类”刻画,就是拒绝和主流秩序站在一起、高高在上地对女性进行各种审判,进而也就避免了将女性客体/她者化。换言之,《善竞》编剧笔下的女人们是作为主体存在的——而这正是长久以来男性才有的待遇。当编剧把女性当作大写的「人」而非「女人」进行描绘,当同性接吻的镜头展示的是单纯的「爱欲/情欲」而非「变态/奇闻/笑料/炮灰/茶余饭后的谈资/成全异性恋的背景板/那两个奇怪的同性恋」,人、女人、以及女同性恋自然会在剧情之中变得鲜活、立体、惹人喜爱。

在东亚社会这个极度逼仄的空间里,个性十足的女人们勇敢地摔跟头、朝气蓬勃地成长、放肆地爱也潇洒地恨。就这样,世界好像终于变成了一个所有女人都可以安居的地方——就好像这里「本该如此、也从来都是如此」一样。

P.s. 唯一不足的可能是最后几集剧情稍稍弱了一点,以及瑟琪的演技只能说是差强人意,当然这些不影响总体质量!另外强烈推荐《善意的竞争》漫画主题曲Don't Touch Mine,B站就能找到,ChoA演唱,可谓把女同情欲演绎到淋漓尽致...

就此搁笔,期待下一次和大家见面!

陌生女人1号 兔姐

二零二五年 四月二日

Image
Image

53 在诗意匮乏的时代,我们一起来读读诗

诗会出现在你的日常吗?

在1969年的纽约,一位艺术家觉得诗歌不仅应该被读,还应该被听到。所以它邀请它的各路朋友和听友,打电话热线进来,读一首自己喜欢的诗。这个活动叫做Dial a poem,它发生在1969年,我去年在纽约的MOMA(当代艺术博物馆)看到了这个展览,并听到了当年被朗读的诗歌。

如今半个世纪过去了,当社交媒体充斥在我们的日常时,连读书都成为一件需要文化体力的事情,读诗简直有些过于遥远了。

所以我叛逆地想在这个诗意匮乏的时代,邀请大家一起来读读诗。更关键的是,当房间里全是大象,我们只能用诗的抽象与之对抗。

这是诗歌最糟糕的时代,因为无人读诗。这也是诗歌最好的时代,因为被迫沉默你恰好还可以写诗。

(播客封面:莫不谷用Canva制作)

【Timeline】

03:14 第1个问题:为什么诗歌在这个时代极其重要且非常必要?

16:30 每一个生活艰难挣扎的人,自动会成为诗歌的主角

20:24 霸王花和金钟罩听到选题的反应:为什么我们对诗歌不再问津,逃避甚至拒绝?

35:27 莫不谷:诗歌对我来说有一种像武功心诀一样的感受

42:54 霸王花:令人郁闷的心情变得凉爽的诗歌分享

51:00 莫不谷:大学时读到令我非常非常震撼的现实主义诗歌

56:35 倘若大家正在经历同样的绝望和失望,不妨把诗歌作为武器

59:49 金钟罩:对我决定离职和Gap起到很强安慰作用的诗歌分享

71:00 莫不谷:我最喜欢的诗,这首诗完全可以刻在我的墓碑上

80:00 脱口秀姐妹CC:今天来交流一些低俗的高雅

96:56 莫不谷写诗《放火》:我不需要别人给什么,但我一定要创造点什么

102:05 霸王花:我感到诗意涌动的时刻和创作的现实风格诗歌

118:00 莫不谷:我常去墓园跑步、学习,所以写了首关于墓园的诗歌

120:00 莫不谷:一首写给每个个体的堪称残酷的诗:《月亮是有毒的》

123:00 金钟罩:我在出国前和辞职去世界游荡的路上写出的两首诗歌

128:00 主播三人:希望以后自己能写出什么样的诗?

140:45 《逍遥游》最后祝大家在这个时代能够拥有理性,靠近诗性,活得逍遥!

【播客&文章&书籍&影视】

播客:《52 美妙人生的关键呀,让我们一起扭一扭它》《49 梦里啥都有,我们隐而未发的欲望,奇谲瑰丽的想象

诗歌:大S《白日梦》;小S《被驯服的象》;许立志《一颗螺丝钉掉在地上》《我咽下一枚铁做的月亮;阿特武德《被吊到半死的玛丽》;海子《面朝大海,春暖花开》;佩索阿《不!我只要自由》;夏宇《我和我的独角兽》《甜蜜的复仇》;《古诗十九首》

影视:莫路狂花出道曲!《东亚忍者之歌 Song of East Asian Ninja》游荡者平台主题曲正式上线!《东亚女游荡之歌 Songs of East Asian Female Wanderers》《I wrote a song for the Spanish oranges西班牙的橙子之歌来了!》;电影《相助》;访谈《Prime li的视频》;纪录片《如此打工三十年》;电视剧《甄嬛传》;视频小红书“游荡者的日常”《快看!爱沙尼亚的日出美到惊人!!!

文章:莫不谷游荡者/Newsletter/公众号文章《这夜晚的星空闪烁,春夏正在到来》;《这个世界可爱的人不多,你是其中重要的一个》

【为全球华人游荡者提供解决方案的平台】

游荡者(www.youdangzhe.com),注册完成后可免费阅读由莫不谷和霸王花撰写的3篇文章(Run的800种可能、语言攻略和全球签证攻略),目前游荡者已上线文章分区功能(游荡区、学习区、欢愉区和闲聊搭子区),欢迎大家注册后开启内容创作并在游荡者游荡愉快!找到同类!交易自由!手机用户可把新网址添加桌面,便于日常使用。使用期间如有任何注册、支付、退款等需求,欢迎给客服邮箱wanderservice2024@outlook.com发送邮件。

放学以后Newsletter《新的一年会好吗?答案在这期播客和这些祝福中

放学以后微信公众号《新的一年会好吗?答案在这期播客和这些祝福中

【延伸信息】

永不失联Newsletter订阅链接:https://afterschool2021.substack.com/(需科 学/上 网)

为全球华人游荡者提供解决方案的平台:游荡者(www.youdangzhe.com)

联系邮箱:afterschool2021@126.com (投稿来信及合作洽谈)

同名YouTube:https://www.youtube.com/@afterschool2021

同名微信公众号:放学以后after school

小红书:游荡者的日常

欢迎并感谢大家在爱发电平台为我们的创作发电:https://afdian.com/a/afterschool

片头曲:《寄生兽》Bliss

片尾曲:《流金岁月》bgm

播客封面:莫不谷用Canva制作

放学以后表情包:微信表情包搜索“放学以后”,感谢萝卜特创作。

播客收听平台:

【国内】爱发电、网易云、苹果播客(请科学/上网)、喜马拉雅、汽水儿、荔枝、小宇宙、QQ音乐;

【海外】Spotify、Apple podcast、Google podcast、Snipd、Overcast、Castbox、Amazon Music、Pocket Casts、Stitcher、Radio Public、Wordpress

💾

The Soviet Cold War Machine

1 April 2025 at 18:15

Welcome to part two of our Cold War history series with Sergey Radchenko. Here’s part one.

In today’s epic interview, we discuss…

  • Khrushchev’s removal from power and the transition to the Brezhnev era,

  • How the USSR and China managed their relationships with Vietnam,

  • Sino-Soviet border conflicts, Brezhnev’s negative feelings toward China, and Nixon’s rapprochement,

  • Watergate and the inability of China or the USSR to understand American politics,

  • Why the Soviets decided to invade Afghanistan,

  • Reagan’s approach to negotiations and his relationship with Gorbachev,

  • How to manage the containment paradox and unknown adversary motives when competing with China and Russia today.

Co-hosting today is Jon Sine of the Cogitations substack.

Listen now on iTunes, Spotify, YouTube, or your favorite podcast app.

The Fight For Vietnam

Jon Sine: At the end of part one, we were just talking about the Cuban Missile Crisis. Then, there was the transition in 1964, when Khrushchev was unceremoniously deposed. My question is, in your story of status, how much does prestige carry over with each new leader? Brezhnev’s era was described as “Khrushchevism without Khrushchev,” after all.

Sergey Radchenko: The phrase “Khrushchevism without Khrushchev” was coined by the Chinese. They were unhappy with how the Soviets continued to pressure China to adopt policies that the Chinese deemed unacceptable. They believed that although Khrushchev pursued anti-Chinese policies and was removed, the policies remained in place. Thus, they invented the term “Khrushchevism without Khrushchev” (没有赫鲁晓夫的赫鲁晓夫政策).

Regarding what happens to the standing of new leaders when leadership changes — that’s actually a very interesting question. Khrushchev was removed in October 1964, at which point he basically was Soviet foreign policy. He stood head and shoulders above everybody else in the Soviet Politburo. He could make decisions single-handedly without consulting anybody, like sending missiles to Cuba. There were some people like Anastas Mikoyan who spoke up against this, but even he was very careful in his opposition. Khrushchev was unassailable in many ways during his last years in power.

Once Khrushchev was overthrown — or, “retired” as they put it — you have a new set of Soviet leaders, including Leonid Brezhnev as the Party First Secretary, who later becomes the General Secretary. You also have Alexei Kosygin, who is effectively the prime minister, and Nikolai Podgorny. They form a triplet, a “Troika” (Тройка) of leaders that ran the Soviet Union for a period of time.

Starting from 1964, Brezhnev, who is nominally in charge as the First Secretary, feels out of his depth. Brezhnev came to power not entirely confident of himself, particularly in foreign policy. He was consulting with others in a way that Khrushchev never did.

This explains why, in 1965, Alexei Kosygin went to China in a bid to repair relations with Mao Zedong. Brezhnev didn’t go himself, but he could not stop Kosygin from doing so. Kosygin wanted to repair relations with China, and Brezhnev agreed to let him try. Of course, nothing came of it because Mao Zedong told Kosygin that their struggle with the Soviet Union would “last yet another 10 thousand years – less is impossible.”

Prime Minister Alexei Kosygin (right) fails to make inroads with Chairman Mao. February, 1965. Source.

Meanwhile, Brezhnev, in order to improve his standing and legitimacy as the leader of a communist superpower, extended aid to Vietnam. When we talk about legitimacy and its practical effects, Vietnam is a key case study. Khrushchev didn’t really care about Vietnam, and the Soviet Union was not heavily involved there until 1964. Beginning in 1964, there was escalation with the Gulf of Tonkin incident and increasing American presence, but there also was renewed Soviet commitment to Vietnam.

Why? Because Vietnam was a communist ally in need. Helping Vietnam bolstered Brezhnev’s personal legitimacy and standing as a leader, especially with the Chinese watching. The Chinese were accusing the Soviets of trying to sell out communist movements around the world. Brezhnev, as the leader of this communist superpower, had to help Vietnam then.

That’s why we see the beginning of massive Soviet involvement in the Vietnam War, including advisors on the ground and military equipment. This would not have happened under Khrushchev, but for Brezhnev, there was a deficit of legitimacy, and Vietnam filled this gap.

Jon Sine: Can you explain why you see this change in attitudes towards Vietnam? In your book, you write that the Chinese position under Mao favors conflict, struggle, and violent revolution as something to be promoted, whereas Khrushchev prefers peaceful coexistence. What changes with Brezhnev?

Sergey Radchenko: I struggled with this question in the book because I could not understand why Khrushchev was so committed to supporting Fidel Castro while not caring about the Vietnamese. I have no answer except for perhaps some personal factors. Ho Chi Minh visited Moscow quite a few times in the late ’50s and early ’60s, trying to repair the relationship between the Soviet Union and China. During these visits, Ho Chi Minh contradicted Khrushchev softly and tried to teach him in a way that I think annoyed him.

It’s difficult to read into these dynamics, but somehow Khrushchev just didn’t like Ho Chi Minh and didn’t want to bother with this faraway place that he did not understand. He preferred to focus on another faraway place he didn’t understand — Cuba — for reasons unknown, getting deeply involved there but not in Vietnam.

Was Brezhnev very different? In terms of his knowledge about Vietnam, no. He was on the same page as Khrushchev and knew nothing about Vietnam. But for Brezhnev, it was a matter of demonstrating his commitment to the communist cause of struggle against American imperialism.

Part of the explanation for Brezhnev’s increased commitment is that it coincided with American escalation in Vietnam. If at this point he had done nothing — saying, “I don’t care about Vietnam, let the Chinese handle them” — he would have looked weak and appeared to have abandoned a communist ally. These factors contributed to the increasing Soviet involvement.

A 1958 poster promoting friendship between the USSR and Vietnam. The text reads, “Great is the distance, but close are our hearts!” Source.

Jon Sine: Let’s stay on this discussion of American involvement. The Gulf of Tonkin incident in August 1964 arguably changed a lot in the course of this sort of ménage à trois between the US, the USSR, and China. On the Chinese side, you read into their documents and show that Mao’s Third Front movement (三线建设) was basically inaugurated by the American escalation in Vietnam. That is to say, Mao was able to build the biggest economic change in China since the Great Leap Forward as a result.

In writing this book, what new insights about this did you discover during the research process?

Sergey Radchenko: The most interesting thing for me regarding Vietnam — and there’s a whole chapter on the Vietnam War in the book — is that we in the West typically think about Vietnam as America’s war. We focus on American boots on the ground, how the Tet Offensive and impacted the United States, the bombing campaigns, and so on. We look at the situation through American eyes.

What I was trying to do in this chapter was to understand how Vietnam mattered for the Soviets. What I discovered was that it wasn’t even American involvement in Vietnam that drove the Soviets crazy. The Soviets were really worried about how Vietnam impacted their competition with China.

As you mentioned, the Chinese in the mid-1960s became increasingly radicalized and saw Vietnam as a case study where they could showcase their influence and vision for world revolution. Mao was at the forefront, advising Vietnamese leaders and saying, “The Americans are invading? Well, that’s not a big deal. That’s fine.” At one point he said, quoting this Chinese proverb, “You have green hills in Vietnam — you can go into the green hills and there’s always firewood there (留得青山在,不怕没柴烧).” Reading this, you might wonder what he was talking about, but his attitude was that revolution was on their side, the future was on their side, so they should fight against the Americans and not listen to the Soviets.

The Chinese were really upset about the Vietnamese taking Soviet aid. Meanwhile, the Vietnamese were trying to balance between both sides, telling the Chinese about how much they admired Chairman Mao, how wise he was, what good advice he gave. Meanwhile, they also felt that they needed Soviet aid and weapons to fight against the Americans. They were balancing both sides.

From the Soviet perspective, this was basically a struggle for influence. Who would win in establishing influence in Vietnam, the Soviets or the Chinese? That was their Vietnam War.

The Russian Army welcomes Ho Chi Minh to Moscow, July 12, 1955. Source.

Jordan Schneider: You also bring up the aspect of logistics — how the Soviets were getting aid to the Vietnamese. Partly, they had to go through China. The US innovated during that time too — this is when containerized shipping first came into play.

What can you say about how the Chinese were interfering with Soviet aid to Vietnam?

Sergey Radchenko: It was awkward because the Chinese could not tell the Vietnamese, “We hate the Soviets, so therefore we’ll be blocking the aid they try to send you once it crosses our borders.” But in reality, the Soviets did try to slow it down. There was a pileup of train shipments at the border.

The weapons had to go across China, starting from the Northern border and going all the way to Southern China and then onto Vietnam — sometimes those train consignments were looted by the Red Guards. There was something akin to a civil war going on in China in the 1960s. It was chaos.

The Soviets would always raise this issue with the Vietnamese. They would say, “Look at the Chinese. You say they are your friends, but look what they’re doing. They’re not transporting our weapons, which you desperately need.” The Vietnamese response was, “Please, just let us handle it. We have to be very careful, and we have to understand the Chinese are having a difficult time.” This was a real problem for the Vietnamese and certainly a propaganda point for the Soviets in the late 1960s.

Jordan Schneider: Were the Vietnamese actually that unruffled by the Chinese interfering with the shipments?

Sergey Radchenko: They hated it. There are so many things that the Vietnamese hated about what the Chinese were doing in the 1960s. We have to understand, of course, that the Chinese were extending considerable aid themselves to Vietnam, and their aid was also important in terms of light weapons and railroad workers — about 300,000 people at some point over the entire period. There was actually considerable Chinese involvement on the ground, and it was important to the North Vietnamese.

But what they did not like was the Chinese interfering with the Soviet weapon shipments, and their pressure on Vietnam to stop taking Soviet aid. They did not like this at all.

One thing they really hated was when Chinese domestic politics became radicalized during the Cultural Revolution — from mid-1966 onwards — and the Chinese tried to promote a Cultural Revolution in Vietnam. If you imagine being the Vietnamese at this point, fighting this war against the Americans, and the Chinese come with their crazy radical ideas and Red Guards and whatnot, that is not something that’s going to sell well with the Vietnamese Communist Party leadership.

You see already at that point a cooling in the Vietnamese approach towards China. They don’t like what the Chinese are doing, and they’re more willing to listen to Soviet advice. The Soviet advice, by the way, is basically to talk to the Americans to end the war, or at least engage in peace talks, whereas the Chinese told them to keep fighting.

Jordan Schneider: I would recommend folks check out The Dragon in the Jungle: The Chinese Army in the Vietnam War, which is a fun dive into Chinese archives regarding this question in particular.

Greatness over Grain and Unlikely Partnerships 狐假虎威

Jordan Schneider: Anyway, let’s take it to Nixon. Brezhnev all of a sudden develops this incredible love affair with one of America’s arguably least lovable presidents. What changed about Brezhnev’s approach to Vietnam, the US, and China once Nixon came onto the scene?

Sergey Radchenko: Jordan, as a way of introduction, we have to understand where Brezhnev finds himself in the late 1960s. First of all, there was effectively a war going on with China. They fought a border war in March 1969 over this little island, which is actually closer to the Chinese side of the Ussuri River. The Chinese have a reasonable claim, but they fought a war over this island. Then, the Soviets made noises about a potential preemptive nuclear strike on China. This was a nasty situation, and the Soviets were worried about a Chinese invasion.

The Chinese, by the way, were afraid of a Soviet invasion. “Afraid” does not cover it — they were paranoid. They thought the Soviets were ready to do a 1968 Czechoslovakia-style special operation all over again. By 1969, they were really preparing for a Soviet invasion.

The Soviets, by contrast, thought that the Chinese were going to invade Siberia. There’s even a Soviet-era joke about the war between the Soviet Union and China that lasted for only two days. On the first day, war is declared. On the second day, the Chinese surrender, and 100 million Chinese cross over the border as prisoners of war. Then the Soviet Union declares unconditional capitulation.

There’s this sense in the Soviet Union that China is an existential threat on their border in the Far East, and that drove Brezhnev nuts. He really hated the Chinese so deeply. You can see that in his various commentary about China and the Chinese — how they’re unreliable, how you can never trust them. There are a lot of orientalist tropes there.

One of the things that I was able to do in my book was track where Brezhnev got his ideas about China. It turns out that he got them from 19th-century Russian orientalist literature.

Jordan Schneider: Let’s stay on that weird racist arc for a second. Do you think these ideas precipitated the Sino-Soviet split, or do you think the breakdown in relations emerged from other factors and then Brezhnev just used orientalism as a justification?

Sergey Radchenko: It’s both. You also see these tendencies under Khrushchev, even when the relationship with China was pretty close. Khrushchev never trusted Mao and thought that Mao was this sketchy character, dictatorial in his ways, and so on.

Once the relationship really started to deteriorate in the 1960s, it became so much more pronounced, both for Khrushchev and especially for Brezhnev, who basically just went full-blown racist on China, saying really nasty things about the Chinese and contrasting them very negatively with the Europeans and the Americans whom he thought he could have a good relationship with.

We have to, of course, put this in context. What was happening in China in 1966-67 was crazy. The Soviet Embassy was literally under siege by Red Guards who erected a scaffold, a platform to hang the ambassador. This was canceled at the last moment because Zhou Enlai came out and talked to what I think was a 14-year-old girl who was in charge of this Red Guard contingent trying to storm the Soviet Embassy.

Zhou Enlai said that hanging the ambassador would interfere with diplomatic relations, so they canceled the operation. But from the point of view of Soviet diplomats, they were about to be lynched, and this was all relayed to Moscow in the form of reports. The Soviets were reading this saying, “What is going on? These people are crazy.”

This contributed to this quite racist thinking about the Chinese and complete failure to understand what was going on there. The Soviets were not alone — nobody could understand what was happening in China. The Cultural Revolution was madness in so many ways, but it contributed to Brezhnev’s thinking that Europe was the better place to build bridges.

Already in the late 1960s, even in the mid-’60s, he began his engagement with de Gaulle, then Pompidou, and later with Willy Brandt after the German Social Democrats came to power in September 1969. Brezhnev felt that Willy Brandt was the way to go, and Brandt, of course, had his Ostpolitik and the promise of economic cooperation with the Soviets.

Brezhnev started developing this European détente before he even turned to the United States. At this point, he was saying nasty things about Nixon. Nixon, of course, was known to Soviet leaders — he was in Moscow in 1959 having the famous, or rather infamous, Kitchen Debate with Khrushchev. They knew who Nixon was, and they didn’t like him. Brezhnev also said things about Nixon that were not very complimentary, but then things changed.

undefined
Vice President Nixon and Soviet Premier Khrushchev debate the merits of capitalism in a model of an American house on exhibit in Moscow, 1959. Source.

What changed was, first and foremost, Henry Kissinger visited Beijing in the summer of 1971, and Nixon’s visit to Beijing was announced. Brezhnev saw that as an “Oh my God” moment. I don’t know if he believed in God, but he was thinking, “The Chinese are our enemies. The Americans are going there — this is really bad — so we have to get Nixon to come to Moscow.”

He really invested himself into this summit, which ultimately happened in May 1972, and he developed a fairly friendly relationship with the American president. After that, I think it really built from there. Somehow, he thought that he had Nixon’s trust, and he then went to Washington to visit Nixon and went all the way to Nixon’s Western White House, which is not far from LA.

Jordan Schneider: Brezhnev was pushing for détente with Nixon. You emphasize that when Khrushchev was envisioning himself on the world stage, he thought he could spread his wings and play some big nuclear weapons-backed games. But Brezhnev, as you say, had a different vision of the US and Soviet relationship — he wanted to run the world together. How did he come to want such a world, and how did that manifest in his policies?

Sergey Radchenko: Jordan, to explain that, let’s go back to Khrushchev. Khrushchev was a really optimistic character. He thought that the Soviet Union was surging ahead and doing so well economically, and in terms of science and technology. The launching of Sputnik made him really optimistic — and for good reason. That’s why he proclaimed that in 20 years’ time they would “establish communism."

As the Soviet joke goes, they decided to hold Olympic Games instead because it was never going to be realized. But in 1960, it seemed like this was possible. Then after that, things went downhill pretty quickly.

The Soviet economic situation wasn’t turning out so well. In 1963, the Soviet Union was already importing grain and spending gold reserves. How can you build communism if you cannot even feed your own people? That was a major problem.

By the late 1960s, it was totally obvious that this promise of communism was not being realized. The Soviet Union was not coming closer to that goal of building abundance and joy for everyone that was part and parcel of that initial promise Khrushchev made.

In 1968, there is this memorandum from Andropov to Brezhnev which basically said, “Look, we’re losing the Cold War. We are losing to the Americans because we’re not investing enough in R&D, in education. Our labor productivity is low,” and so forth.

I was able to access the Soviet Politburo Discussion from 1966, which includes hundreds of pages of discussions about the state of the Soviet economy, and they all knew that things were not going well.

They tried to implement reforms, the Kosygin reforms, introducing incentives, basically making the country more capitalist. But it wasn’t working. They tried to tinker with it here and there, but the whole system was just garbage. It was just not delivering.

Because of that, the ideological underpinnings of the Soviet system started to fall apart. People were not buying anymore because the material abundance was not there. Brezhnev understood that. Brezhnev was looking for a new idea, and that is where this external aspect of greatness comes into play. America would recognize the Soviet Union as an equal superpower, and that could be sold to the Soviet people as Brezhnev’s achievement.

Together with this came the idea of peace, and Brezhnev saw himself as a peacemaker leader of the USSR. In his conversations with Nixon, he would often refer to the fact that the Soviet Union and the United States together had enough nuclear weapons to destroy the world six, seven, eight times over, and so they had a responsibility to fix the world — to resolve the nuclear problem, to stabilize China.

The title To Run the World came from the moment in the spring of 1973 when Henry Kissinger went to Moscow, and Brezhnev took him hunting wild boars. It was outside of Moscow in this dacha, and they were in this hunting tower, just Brezhnev, Kissinger, and the interpreter, waiting for the wild pigs to come and feed so they could shoot them.

Kissinger recounted that moment later in a conversation with Nixon, he said, “Look, Brezhnev told me, ‘Don’t take any notes, but I’ll tell you this. What we want to do is we want to run the world together with the United States.’” That struck me as a very interesting proposition from a leader of a communist superpower, working together with the United States. How do you even explain this from a Marxist-Leninist perspective? That doesn’t make any sense.

True, it doesn’t make any sense! But it does make sense from the perspective of selling Soviet greatness to the Soviet people.

Jon Sine: Nowadays, there are strategists talking about how the US could peel Russia away from China. It’s interesting that, at various points during the Cold War, both the Soviets and the Chinese were seeking to align with the US against the other.

What historical lessons from that dynamic do people forget today?

Sergey Radchenko: This is my favorite example of how ideology can be just cast aside. We can talk about China being this revolutionary power that complained endlessly about the Soviet Union betraying the global revolution, but then we get to the late 1960s, and they basically come around and embrace the United States.

An interesting moment discussed in the book was, I think in the fall of 1970, when the American Mao biographer Edgar Snow turned up in China, as he did on occasion. The Chinese leaders thought Snow was a CIA spy, but he was actually just a leftist journalist.

Anyway, Edgar Snow turned up in China and had a conversation with Mao Zedong in which Mao said, “We think that Nixon is a good fellow.” Snow essentially said, “Well, how can you say that? You don’t mean that, right?” Mao Zedong repeats himself, “We think he’s the best fellow in the world.”

Mao, of course, hoped that Edgar Snow would carry this to the Nixon administration — remember, in his mind, Snow was a CIA agent. Later, the summary of this conversation was circulated to the party committees around China to introduce the Chinese people to the idea that China was changing course and turning toward the United States.

Do you know how party committees would have kind of fake debates? They had debates and questions were collected, and those questions were reported back to the center. Questions included things like, “If Nixon is the number one best fellow in the world, why are we having a quarrel with the Soviet Union? Can’t we repair relations with them as well?” The Chinese party officials were confused about this, but Mao would not have any of that. He felt that those people didn’t understand strategy, and that China had to turn to the United States because the Soviet Union was the enemy.

You see, it doesn’t matter that turning to the United States entails turning to “imperialism.” That’s not what matters.

Going back to your question — at that point, both the Soviet Union and China were willing and able to set ideology aside and turn to the United States and try to improve relations with them on the basis of geopolitical great power competition with each other.

Jon Sine: That’s exactly what I was thinking, because normally the story is you have brilliant strategists and Kissinger making that flight to Pakistan and then secretly flying off to China for this engagement. But some have argued, and I think your findings support this, that Mao in some ways, insofar as anyone deserves credit, might be the one making this move.

Reading back through Mao’s various writings, when he talks about the Japanese, for instance, he will sometimes half-jokingly, half-seriously say, “The Japanese are the people that we have to thank the most, because without them, we would never have come into power.” Are you sure there’s no element of speaking tongue in cheek here when he’s saying these things about Nixon being the number one good fellow?

Sergey Radchenko: You always have to be careful with his pronouncements. Mao does say that about the Japanese consistently. What he basically means is that because of the Japanese invasion, the nationalists were weakened — the Kuomintang Party was weakened — which provided the space for the Chinese communists to establish their power.

In a sense, it does make sense what he says about the Japanese. Of course, he’s being sarcastic to a certain extent.

With regard to Nixon, when Nixon and Mao met in February 1972, Nixon tried to engage him in conversation, and Mao kind of brushed it all aside and said, “Oh, you can talk to Zhou Enlai about practical matters. I don’t care about practical matters, I want to talk about philosophical matters.” He says something like, “We like rightists, and we voted for you in the 1968 election."

I don’t think Nixon quite gets that. Nixon is like, “Okay, well, listen...” But Mao is trying to say that he finds the Republicans more trustworthy because although they’re reactionary, as far as Mao is concerned, at least you always know where they stand. Whereas with the Soviets, you don’t. The Soviets would say one thing, and they would cover their actions with leftist phraseology, but in reality they do something completely different.

The same goes for Social Democrats, and he’s full of disdain for Social Democrats like the Europeans — Willy Brandt and all those people in Europe who are basically trying to engage with the Soviet Union, which would allow the Soviet Union to deal with China. He’s accusing the Western Europeans of trying to orchestrate another Munich, where they would sell out China.

Those are the kind of issues that Mao brings up, and I think he’s quite honest about it. It’s not a sarcastic comment when he says that he likes Nixon.

Jon Sine: The Soviets also really liked Nixon. This is probably my favorite part of your book — certainly the funniest. I was actually laughing. They find out about the Watergate scandal and see Nixon about to be removed from power, and Brezhnev is flipping out, thinking that the whole political system in the US is specifically trying to undermine détente. Then he sends a message, I believe it was to Andropov, essentially saying, “You’ve got to help Nixon. You’ve got to find some dirt on his opponents to help him.”

Share

Sergey Radchenko: It is funny. It shows what kind of great material you can find in the Russian archives. In this particular episode, Andropov of the KGB is there, and you have Brezhnev’s aide asking, “Do we have some compromising material on Nixon’s opponents that we can use to help Nixon?”

It is hilarious, but the Soviets never could understand Watergate. Mao also did not understand Watergate. They could not comprehend how you could remove such a wonderful president who was creating détente and bringing the Cold War to an end and who had just won a big election. How? Clearly it’s some kind of conspiracy. At one point they say, “You see? They killed Kennedy, and now the same people are bringing Nixon down.” That was Brezhnev’s take on this. They never get it; they never understand what this is about.

Jon Sine: That was such a crazy time in American politics, from that decade on — the president literally having his brains blown out on TV, to then having a president actually impeached and removed from office. We forget today how things have been very crazy in the past.

The last analogy I was thinking of is when Nixon was about to become president. Some people say he engaged in what might technically be called a treasonous act by being in contact with the Vietnamese and trying to discourage them from agreeing to a peace deal. His rhetoric might sound familiar to people today who pay attention to news about Ukraine. He came in saying that the war in Vietnam was a complete disaster and we needed to get out of it immediately. But what ended up happening was an escalation beyond anything previously seen. I don’t know how much of a warning that would be for today, but I did think it was interesting.

To be fair, I heard Stephen Kotkin bring this up, so I thought it was an interesting analogy, though obviously there are many disanalogies — things that don’t map equally.

Sergey Radchenko: Niall Ferguson also raised this in a recent article comparing Trump to Nixon. The difference between Ukraine and Vietnam was that in the late 1960s, American troops were in Vietnam, which had a direct impact on American society and politics. There were anti-war protests going on in the United States. Today, American troops are not in Ukraine, so you don’t have the same kind of impact — no protests.

Both Nixon and Trump see these theaters as peripheral to America’s core interests. There is a parallel there.

Nixon’s way to get out of Vietnam was to try to coerce Vietnam, including by intensifying the bombing, dropping threats of a nuclear attack on Vietnam (which did not really work), and also working with the Soviets. From his perspective, that was a big part of the whole engagement with the USSR — finding an exit for the United States, “peace with honor,” getting out of Vietnam, and getting the Soviets to facilitate this.

The problem was that the Soviets were not facilitating any of this. From the Soviet perspective, they loved the fact that the Americans were engaging with them. They loved the honor of being a co-equal superpower. But if you’re a co-equal superpower, aren’t you supposed to have clients? You’re not supposed to betray your clients or force them to surrender to the United States. For them, this linkage never worked. They thought they could both help Vietnam and have a good relationship with the United States. That did not really work for the Americans.

Jordan Schneider: I love this line in your conclusion. This is talking about Khrushchev, but it applies to Brezhnev as well. The core question is, “Would he be willing to moderate Soviet foreign policy in return for being accepted as America’s equal? The proposition never worked because of the very fact that being accepted as America’s equal meant rejecting external constraints on foreign policy behavior.” What sort of equality would you talk about if you couldn’t have proxy wars or missiles in Cuba?

“Soviet engagement in the Third World, and indeed American acceptance of this engagement, were part and parcel of what it meant to be a superpower.”

This gets at the other core question as we come to the end of Brezhnev’s effectiveness — if he had stayed healthy and if Nixon wasn’t impeached, they could have kept the good thing going. But once his health deteriorated and Nixon departed, you argue that bureaucratic interests took over, and everything started ramping up again.

What are your thoughts on the transition?

Sergey Radchenko: It’s hugely important. We’re talking about health and power. It’s a big question for the late USSR, and we’re also familiar with this question in the West.

Brezhnev was very charismatic, very active, and very engaged until about 1974, and then he declined rapidly. He developed all kinds of ailments and basically became a figurehead. By the late 1970s, he didn’t really decide anything.

If you look at his summit with Carter in Vienna, he was just reading from pieces of paper. He wasn’t even thinking about what he was reading. When Carter responded, he would turn to his aides and ask, “What should I say now?” They would give him another piece of paper and he would read from that.

The Soviet deep state took hold in the late 1970s in a major way. The bureaucratic interests took over, and the Ministry of Defense became really important. For them, promoting various geopolitical schemes in the Third World was a key issue. They also resisted nuclear disarmament. They thought it was a bad idea. They wanted more investments from the state, so increasingly the Soviet economy became more militarized.

Jimmy Carter and Leonid Brezhnev in Vienna for the signing of the SALT II treaty, 1979. Source.

You have the interests of the International Department of the Soviet Communist Party, whose role is to promote revolutions in the Third World. It’s in their job description — promote revolutions. Previously, Brezhnev would have pushed them away because it’s beyond their pay grade to define policy. But by the mid to late 1970s, they come to influence policy in a major way, and we see that in the increasing Soviet involvement in Africa.

Then you had the Foreign Ministry and the KGB, and they all had their own distinct interests. Those bureaucratic interests increasingly came to dominate Soviet policymaking, and it drifted as a result. Policymaking became more conservative overall because there was no single figure who could break the ice and take charge the way Brezhnev did with the Nixon Summit in 1972. The bureaucracy was against it, but Brezhnev did it anyway.

Well, by the late 1970s, he couldn’t do that because he was no longer mentally alert. It is part of the story of Soviet decline, and it also shows how having an active leader at the top could actually have good results. Not always, because sometimes you could have an active dictator who will do terrible things precisely because he’s not constrained by the bureaucracy. But in the Soviet case in the late 1970s, power at the top was missing and the bureaucracy took over.

Boredom and the Graveyard of Empires

“When everything is calm, measured, stable, we are bored… we want some action.”

~ Vladimir Putin on the invasion of Ukraine, December 2024

Jon Sine: The last important thing that happens under Brezhnev, though he’s not really conscious of it, is the Soviet invasion of Afghanistan. To your point, you have Gromyko, Andropov, and Ustinov — the KGB, defense, the Troika. They go into Afghanistan, and on the US side, my understanding is Carter and maybe Brzezinski assumed this was a play to get all the way to the Persian Gulf.

What was the motivation to invade? If you were to identify some continuity of motivations between regimes, what did you find that was new about what was driving them?

Sergey Radchenko: I found some interesting things. We’ve had many people write about the Soviet invasion of Afghanistan, from Arne Westad to Rodric Braithwaite and many others. It was difficult to find anything new, but I did manage to discover some interesting materials from the fall of 1979.

At that time, there was already a conflict between Nur Muhammad Taraki and Hafizullah Amin in Kabul, but Amin had not yet killed Taraki. Taraki came to Moscow, and Brezhnev had a conversation with him — though “conversation” is in quotation marks because Brezhnev was in no position to have any productive dialogue. He does offer a warning to Taraki, “Look, when you go back, be careful because there’s a problem in your senior party ranks.” Taraki says, “Leonid Illyich, don’t worry about it. Everything is fine."

He goes back, and of course, he’s arrested by Amin’s men and ultimately put in prison and murdered on Amin’s orders.

What we see then is Soviet leadership thinking about how to respond. Their first instinct is actually to work with Amin despite the fact that they consider Taraki’s death an absolute slap in the face, a betrayal of the USSR. They feel they could work with Amin because he’s surrounded by “pro-Soviet people,” people who had studied in the USSR. Brezhnev writes about it in some exchange of memoranda at the senior leadership level, which I discuss in the book. This suggests they might continue working with Amin.

But by December 1979, they decided that they had to remove him. I ask in the book why this happened. There are different possible explanations for Soviet involvement — perhaps the Soviets were just adrift with nobody making foreign policy anymore, but I don’t think that fully explains it.

A more interesting explanation, which I highlight in the book, is that they worried Amin would “do a Sadat” on them. They had lost President Sadat in Egypt, who was supposed to be a pro-Soviet client, but they dumped him because he decided to align with Kissinger and Nixon. The Soviets got outplayed in Egypt.

After losing Egypt, the Soviets viewed almost every country in the Middle East as potentially another Egypt, and Afghanistan fell into that category. They grew suspicious of Amin, thinking he had ties with Americans. They received information that he was in contact with Americans and concluded he was potentially pro-American and could sell them out. They feared Americans would then establish a presence in Afghanistan, creating a strategic problem for the Soviets. They decided to act primarily because they didn’t want Amin to become another Sadat.

Jon Sine: Let’s transition to Gorbachev — and start with the Gorbachev-Trump analogy.

DOGE is American perestroika, at least in the charged anti-bureaucratic approach. Under Gorbachev, the Soviets decreased employment in the ministries and party personnel by something like 30 to 40%, which the current administration would certainly aspire to with the civil service in the US.

You have Gorbachev willingly giving up the Warsaw Pact, which some might analogize to Trump’s stance on NATO. Ultimately, as Axios reported, you have Trump’s desire for a Nobel Prize and the prestige and legitimation of doing something great for a foreign audience. People criticized Gorbachev for similar reasons when he wrote his 1987 book, Perestroika and New Thinking. He had it translated immediately into English, and it was written for a US publisher.

Sergey Radchenko: The book was also published in the USSR. Historical analogies are limited to some extent, but I see some interesting parallels.

One counterpoint I would offer is that the Soviet economy was basically going to hell in the 1980s, and they knew it. They knew they were losing the Cold War and that the promise of deliverance for the Soviet people was just a fake promise. The American economy, if you look at productivity, investment in technology, R&D, and so forth, is far beyond anything else anybody in the world can offer.

The question is one of necessity. I understand some people will say, “America has to carry out reforms because of the national debt” or other issues. In the Soviet case, perestroika was absolutely necessary, and they knew they had to do it. They knew it in the ’60s, but they didn’t act then because they craved stability under Brezhnev and were trying to avoid political upheavals.

They were also, to a certain extent, bailed out by the price of oil and the discovery of oil in Western Siberia, which they could sell for hard cash. This helped feed the Soviet people because they could import grain and some technologies. But they knew the system wasn’t working. When Gorbachev came to power, he had to begin doing something because doing nothing was not an option.

Subscribe now

Jordan Schneider: You write:

Gorbachev had everything in 1985 — an empire (however decrepit), an ideology (however stale), and above all, an office with truly awesome power. What he did not have was greatness, as he chose to understand it, greatness before history. He pursued that fleeting dream for himself and his country all the way to the famous Pizza Hut ad.

When you compare what Gorbachev and these other historical figures define as greatness, and what Trump defines as greatness, you see that the Trump definition is smaller, more personal — “People respect me, and speak to me nicely in the Oval Office.” All these other leaders had a vision of being grand historical figures who achieved something monumental.

There are parts of Trumpism that claim to revive America, defeat wokeness, and bring back ideals. Every once in a while, he’ll mention something about manufacturing, but “Make America Great Again” is much more about Trump personally than it is about a national vision of prestige and greatness.

Sergey Radchenko: It’s related to a certain extent. Soviet leaders desired greatness for themselves and for their country. This book is applicable to almost any other would-be great power or superpower. To a certain extent, it can be read as an allegory of the United States. This is how John Lewis Gaddis saw it in his review in Foreign Affairs — he thought that between the lines, the United States was clearly lurking there, really telling the American story, not just the Soviet story.

The question of legacy is super interesting and important because, as you say, Jordan, Gorbachev is at the helm of a superpower. He has all the power concentrated in his hands, and yet he wants something more. It’s similar to Mao Zedong in 1965. He had all the power, he could do anything. Even after he got rid of people he thought were conspiring against him by 1966, he continued the Cultural Revolution. For what? For legacy, for greatness before history.

If you think about Putin and his invasion of Ukraine, he has all the power, and yet he’s invading Ukraine. For what?

For greatness before history, the way he understands greatness.

If we return to Gorbachev once again and ask, “Fundamentally, why did he do it?” The answer is because he wanted something greater than what he had. He wanted to be remembered as a person who would bring the Soviet idea to life for the first time because he thought his predecessors never made it work. He would make it work, make it globally applicable, end the Cold War — this was his mission. That, I believe, is very important.

Jordan Schneider: We talked in the first episode about national ambition — is it a gas or is it a solid? Will it expand until it hits another obstacle that contains it, or does it have some natural limits? It’s interesting to analogize that not at a national level, but at a personal level.

When you’re in power for 15 years, you get bored. You’re probably already a bit of a gambler if you were able to make it all the way up the system. You just want more. This is why term limits are important — leaders tend to go a little cuckoo after too long. Either they go senile like Brezhnev — who thankfully didn’t start World War III because he was worried about his bladder — or they do something like Putin in Ukraine. It’s a scary thing to contemplate.

Sergey Radchenko: I agree. It’s fascinating to consider how these leaders, perhaps out of boredom, perhaps because they have nothing else to do, take dramatic actions. That might sound ridiculous, but speaking of boredom, Putin was once asked about Ukraine in one of his recent press conferences, and he essentially said, “We were kind of bored, and we decided to do it.” I’m paraphrasing, but he literally mentioned the word “boredom” in his comments on why he invaded Ukraine.

That sounds crazy, but if you’re a leader with all the power in your hands, you need something else. Schopenhauer in the early 19th century reflected that one of the major problems human beings face is generally the problem of boredom. They don’t know what to do with themselves. Now multiply that with absolute power, and you get people like Gorbachev who think, “Why not do something so great that everybody will remember me?” And well, he certainly succeeded.

Jon Sine: There’s a hedonic treadmill effect when it comes to status. You’ve achieved so much and yet, when you look back once you’ve achieved it, you realize you could achieve more, and the last thing you accomplished no longer seems as good. Now you’re wondering, “How can I really leave a legacy that my children and my children’s children will remember?”

What better than changing your borders, acquiring Greenland, getting back Panama?

Sergey Radchenko: Exactly. Or launching Perestroika. That’s why Gorbachev’s book which you referred to, the one published in the West, was actually subtitled “Perestroika For Our Country and For The World.” He was trying to restructure the entire world, not just his own country. That’s the extent of his ambition.

Humiliation and Containment

Jordan Schneider: Let’s talk about the dance that the Europeans and the Americans did with Gorbachev. The goal was to get the Soviets to understand that their relative power was decreasing, and that America and its allies had won the economic, science, and technology competitions. But you also want to allow a leader like Gorbachev to pursue his vision without risking something like a Stalinist revanchism, which wasn’t totally out of the cards.

Could you talk about how the West managed Gorbachev’s moment of defining national greatness as Perestroika rather than, say, conquering Germany?

Sergey Radchenko: First of all, the Americans were quite worried about Gorbachev to begin with, and that’s a well-known story. They were concerned because they thought he could actually succeed in what he claimed to be doing — reinventing the Soviet idea and making the Soviet Union a much more serious strategic competitor to the United States.

Some thought Gorbachev wasn’t for real and his reforms would ultimately be undone. Then, a moment came when American leaders thought they could reach out to Gorbachev in the name of international peace and reach some agreements.

Reagan was the key person here because he also believed in the importance of avoiding nuclear war and felt the great responsibility that was on his shoulders. Many people in the United States like to talk about Reagan “winning the Cold War” and how tough he was on the USSR. For me, the real Reagan was the one who believed that nuclear war had to be avoided and that we needed to talk to the Soviets.

Ronald Reagan and Mikhail Gorbachev laugh in Washington
Ronald Reagan and Mikhail Gorbachev laughing together in Washington, December 8, 1987. Source.

As Reagan said at one point, Soviet general secretaries kept dying on him, but then he finally had a partner in Gorbachev with whom he could talk. They first met in Geneva in 1985, then in Reykjavik. There came a remarkable moment in Reykjavik where they discussed abolishing nuclear weapons altogether. It was “almost decided” to the horror of Reagan’s advisors. Here was an interesting moment when an American leader thought, “What if we actually try to play along and see how far we can get? Maybe we can actually change the course of history.” Reykjavik represented that possibility.

Jordan Schneider: Going all the way back to Stalin and Truman, you have this quote from Henry Kissinger who wrote in a 1957 book, “The powers which represent legitimacy and the status quo cannot know that their antagonist is not amenable to reason until he has demonstrated it. And he will not have demonstrated it until the international system is already overturned.”

We’ve got our answer with Putin, but we don’t quite have our answer yet with Xi’s China. This presents a very difficult choice — on the one hand, if Stalin had turned out to be reasonable, we could have had a different timeline. But if not, the price for running that experiment would have potentially been the Soviet conquest of Europe or even the world.

How are we supposed to think about that dilemma sitting here in 2025, imagining the future of US-China relations?

Sergey Radchenko: Jordan, that’s a great question. The fundamental problem is we don’t know what the other side is really thinking. The other side may not even know what they ultimately want. This is the case with Stalin. Even with the passage of all these years, we don’t know what Stalin’s ultimate ambitions were. Was he limited in his appetites? Would he stop, or unless he met with counterforce, would he keep pushing? Would he keep going until he conquered Europe, and from Europe jump over to conquer Asia and the world?

It sounds fantastic, but “appetite comes during eating” as the French say (“l'appétit vient en mangeant”). If you’re an American policymaker faced with this situation, you’re thinking, “How do we stop that? What is the most reasonable policy we can adopt?” It seems the most reasonable approach is the policy ultimately adopted — containment. You push back on expansionist desires, and the other side has to take this into account. They are deterred and limit their ambitions, but as a result, a struggle unfolds that becomes a prolonged confrontation like the Cold War.

It’s a sad situation, but maybe inevitable and unavoidable. It’s like choosing between two evils. On one hand, containment and counter-containment lead to a Cold War and could potentially lead to a hot war. On the other hand, doing nothing could expose you to a situation where you’ll basically have to accept that the other side dominates everything.

It’s the same with China. The fundamental problem we face today, even with Russia as well, is that we simply don’t know what the other side wants. Does Xi Jinping want to overturn the existing order, or is he just trying to change China’s position within this global order? That question has been debated for 20 years, even before Xi Jinping.

We don’t know the answer. Some people say, “We know Xi Jinping is trying to overturn the world order because here’s the evidence.” I would say, “I’m looking at this evidence, and I’m not 100% convinced.” We simply don’t know whether Xi Jinping himself knows what he wants to do.

Under those circumstances, what’s the best policy to pursue? The policy is probably a combination of containment, firmness, plus clear signaling — “We see what you’re doing. This will be our response.” No jumping around and doing impulsive things, which American foreign policy sometimes tends to do. We’ve seen that with Trump, but also with previous administrations.

With China, consider Nancy Pelosi’s ill-advised visit to Taiwan. What did it achieve? It was very provocative and completely useless in many ways. As a result, we had a breakdown of communication at senior levels between Chinese and American militaries. Did it benefit anyone? No. What was the point? There was no point.

We have to be firm, maintain dialogue, and signal to the other side that we see what they’re doing and are preparing certain countermeasures. But if they back off at the right moment, we will also not proceed with those countermeasures. I think that’s the way to mitigate great power confrontation in this potentially new Cold War that is unfolding before our eyes even today.

Jordan Schneider: I want to apply the lessons of prestige to the US-China relationship, and Putin as well. Are leaders satisfied by the rate of change of prestige or by the absolute value of prestige? Are there ways to give prestige that don’t fundamentally compromise core national interests and power? It seems that prestige is often in the eye of the beholder, and America can give prestige in many different ways — some costly, some much less so.

Sergey Radchenko: Core interests for prestige certainly exist. In my book, I discuss Poland and how it was central to Stalin’s security concerns in Europe. No matter how much prestige he would have received from the allies, he was determined to do what he wanted in Poland because he considered it a fundamental red line core to his security interests. You might say this also applies to Putin or Xi Jinping today.

On the other hand, we sometimes underestimate the importance of recognition, respect, and acceptance. We also tend to underestimate how our actions can humiliate the other side and provoke adverse reactions we would rather have avoided.

Consider President Obama’s rhetoric about Putin being “the kid in the back of the classroom,” which reportedly outraged him. You might ask whether Putin would have still invaded Ukraine if Obama hadn’t made those remarks. Perhaps he would have because Ukraine is central to his neo-imperialist vision of Russia, or perhaps not. It’s a counterfactual — we simply don’t know.

The question is whether he feels he has standing and respectability. I believe it ultimately does matter. The same likely applies to the Chinese leadership.

It’s important to note that, despite our strategic rivalry with China and difficult relationship with Russia, these countries are ruled by autocrats who sometimes make decisions impulsively, for no particular reason, and occasionally in ways that clearly contradict their country’s national interests. They do this because they feel insulted or humiliated. We should not underestimate those sentiments. In dealing with these countries, we should be careful — respectful but firm.

Jordan Schneider: Perhaps even more famously, Barack Obama at the 2011 White House Correspondents’ Dinner made fun of Donald Trump for the birther conspiracy, and some trace Trump’s desire to run for president back to being humiliated on national television.

Sergey Radchenko: That’s a great counterexample.

Jordan Schneider: Maybe this is the 4D chess with Trump — we’re going to stop calling Putin a dictator and say he’s the best thing since sliced bread. Will that make him pull out of Ukraine?

Sergey Radchenko: Maybe not, but it’s a cumulative effect. Over time, these things matter. It’s not as if we suddenly say, “We love you, come back to the G7, let’s have a good relationship,” and he’ll respond, “Yes, I’ll pull out from Ukraine.” We’ve already turned that corner, and it’s very difficult to go back. But on the road to confrontation, these factors do matter, in my view — perhaps not decisively and not always, but they do matter.

ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.

新书讯04:满洲秘密

By: wuyagege
1 April 2025 at 17:28

亲爱的读者周二好,四月的新书讯聊一下满洲(中国东北)相关的历史新书。

20 世纪有两大火药桶,一个位于东南欧,名叫巴尔干;一个位于东北亚,曾经叫东三省,又名满洲,后来叫中国东北。第一次世界大战,第二次世界大战和冷战的爆发,都跟这两个火药桶脱不开关系。

巴尔干火药桶已经得到历史学家充分的检视,而满洲火药桶仍是一团迷雾,深埋着多个国族出于各自原因不愿意直面的秘密。唯有超越狭隘的一族之见,才有可能拨云见日。这一任务留待具备跨国史观念 (transnational turn) 的新一代历史学家去达成。

Stalin, Japan, and the Struggle for Supremacy over China

​Kuromiya, Hiroaki. Stalin, Japan, and the Struggle for Supremacy over China, 1894–1945. London: Routledge, 2022.​

黑宫广昭退休后的集大成之作,还是他最拿手的俄国视角。

1920年代之后,英国元气大伤,美国缩回孤立主义,东亚的命运取决于日本和苏联之间的角力。斯大林的苏联帝国横跨东欧到东北亚,最大的战略危险是同时面对德国和日本两面夹击。日俄战争和日本干预苏联内战的经历仍历历在目,斯大林如何面对日本,如何破除被两面夹击的战略困境?

斯大林在东北亚最核心的战略就是以退为进,将中国“泥潭化”,以中国拖住日本,以中国的多个势力为棋子来替苏联打代理人战争。黑宫重述 1928 张作霖身亡到 1945 日本战败的整个过程,以满洲为核心展示斯大林如何引诱日本一步步深入中国泥潭。以此视角观察,满洲事变(九一八)和卢沟桥事变(七七)背后影影绰绰的苏联布局变得更为清晰。

1945 苏联红军挺进满洲,中华民国精疲力尽,美国三心二意,谁是东亚真正的棋手和胜者已然不言自明。

滿洲帝國的遺產

​姜尚中 and 玄武岩. 大日本.滿洲帝國的遺產:強人政治與統制經濟如何影響近代日韓. Translated by 李雨青. Taipei: 八旗文化, 2018.​

存续时间短暂的满洲国,却是非常重要的政治经济学试验田。这本书以岸信介和朴正熙的满洲经验为中心,讲述满洲统制经济模式对战后日本和韩国的深远影响。

国家资本主义模型,计划 + 市场的融合在东亚的第一次大型试验就在这里。来自东京帝国大学的改革派官僚们在这里第一次展开手脚进行实验,他们所获得的经验并没有随战败而消散,他们随后成为日本和韩国战后国政主导者。

通产省奇迹和汉江奇迹都有着微妙的满洲起源,统制经济和强人政治手牵手,余脉深远:岸信介有个外孙叫安倍晋三,朴正熙的女儿叫朴槿惠。

Making Mao's Steelworks

Hirata, Koji. Making Mao's Steelworks: Industrial Manchuria and the Transnational Origins of Chinese Socialism. Cambridge: Cambridge University Press, 2024.

满洲的另一份遗产在红色中国。

平田康治以鞍山钢铁(满铁-昭和制钢)为主要案例,讲述满洲工业中心如何变成“共和国长子”。

在滿洲的臺灣人

許雪姬,《離散與回歸:在滿洲的臺灣人 1905-1948》,新北:左岸文化,2023。

满洲土地上生活着各种移民人群, 白俄,犹太人,朝鲜(韩国)人,关内来的中国移民,当然还有从台湾来的官僚和工程师们。

满洲国 Grand Hotel

الصورة

最后来本轻松的。

平山周吉 2022 年的物语 《満洲国グランドホテル》荣获了司马辽太郎赏,懂日语的朋友可以一试。

写在最后

新书讯写了四期,还蛮有趣,除了想主题会偶尔头疼。你们想看什么主题的新书,也可以留言~

除每月新书讯外,这月开始更新东德简史,五章正文+两篇番外,我慢慢写,可能一个月只更新一章,敬请期待。

东德简史

1,生于不义 1945-1949

2,反抗的国度 1950-1961

3,顺民的国度 1962-1980

4,江河日下 1981-1988

5,大厦崩塌 1989

番外一:默克尔在东德

番外二:普京在东德

感谢阅读,如果喜欢《不如读书》请推荐给亲友订阅:

Thanks for reading 不如读书! Subscribe for free to receive new posts and support my work.

My Media Diet

31 March 2025 at 17:57

Radiooooo: A Cosmopolitan Dreamland

More and more of my listening this year has shifted away from Spotify. As big a fan I am of AI, I’m sick of Spotify degrading artists and albums in favor of mixes and playlists by ghost musicians. In its place for me has come the magical website and app Radiooooo.

Radiooooo takes human-submitted tracks, categorizes them by country and decade, and lets you explore. On their whimsical global map, you can either let their algorithm take you on a world tour, or you can decide that today’s the day to spend an hour with the tunes of pre-revolutionary Iran.

Most of the world maps I look at reflect geopolitical competition, but Radiooooo’s map flips that on its head, using music across time to show how cross-pollinated we are. The selection is biased towards selections that show global influence, and it is so cool seeing that global influence! Japan had excellent do-wop in the 1970s! A Swedish singer in 1959 made a song about Puerto Rico! A calypso song from Hong Kong in 1960! Soviet crooners from the early Brezhnev era! Radiooooo is a life-affirming reminder that joy and creativity persist even in decades of dictators and war.

Beethoven provided my other favorite listening experience of the year. Norman Lebrecht’s Why Beethoven: A Phenomenon in One Hundred Pieces is a brilliantly written collection of short essays, which introduce, in some creative oblique way, a Beethoven composition in just a few pages.

Chinese TV

再见爱人 See You Lover — China’s “Road Rules for Divorce” show — is a reality show where three celebrity couples all married for at least ten years and on the verge of divorces take an 18-day road trip together. It is some of the most gripping content I’ve ever seen. Discussed here and with on the podcast episode below. It’s the year’s only must-watch piece of Chinese content.

Over the course of the show, you get to watch modern Chinese women wake each other up to the fact that they can demand more from their neglectful and abusive husbands through an interplay of independent earning power and global feminist ideas (filtered primarily through Japanese writers). Score: 9.8/10

欢乐家长群 Growing Together — this light series is about four couples in Beijing raising young children, and the drama that takes place through a parent’s WeChat group. Contemporary marital drama is much more engaging in 再见爱人, but this show is a sanitized window into how rich parents try to raise their kids in a post-cram school era where prioritizing self-actualization holds some weight relative to academic achievement. Score: 7.5/10

I got four episodes into 大明王朝1566 before getting overwhelmed. It’s quality content that requires real attention, which paternity leave did not allow for. Will try again this year!

山花烂漫时 She and Her Girls [they should just let me name these shows, this is awful] is a dramatized true story of a teacher who founded an all-girls high school in a poor rural village.

It’s a high quality drama and illustrates why top flight Chinese dramas have a hard time finding footing internationally. In one episode, the teachers start jumping ship because of bad working conditions and low pay. The only ones who stick around are the Party members — who use the power of red songs to get their kids to study. It’s a lot to stomach. Chinese audiences understand that elements like these are non-negotiable if you want to make a show showcasing the reality of modern poverty, but it causes global audiences to tune out.

“In the war against the Japanese, as long as there is one Party member, there’s no way that group would lose ground. Today I learned we have six teachers who are Party members. Could it be that we lose ground for the Party at this women’s high school? The long march was so long, I know this semester will be hard, but not that hard! This semester let’s all review again the Party Membership Oath. Before we were colleagues, now we are comrades. 之前我们是同事,现在我们是同志!”

Ep 5 is the standout episode for me where the teachers go out into the countryside to recruit students only to encounter desperate poverty preventing these girls from studying. It’s a refreshing departure from the normal poverty porn you see in food shows like Bite of China.

Other themes include:

  • A morally complex protagonist! Usually in a propaganda-boosted storyline like this, with someone sacrificing their personal life for poor people, the main character gets turned into Mother Teresa. In this show the teacher is proud — and resentful of the students who still rebel when given the opportunity.

  • The real importance of the gaokao — ep 19 opens with a dream sequence of everyone getting super high test scores (which doubles as satire of traditional gaokao shows), and closes with a speech by the lead teacher about why the gaokao isn’t absolutely everything. The show illustrates how deeply that exam pressure has traumatized these girls, and by the next episode the head teacher is back to stressing everyone out about test scores. College is the only social mobility path for these rural youth, but if you don’t get accepted anywhere, at least you didn’t have a child at 14!

  • Alcohol — being raised by alcoholic parents contributes to a lot of the girls’ misery, but the founder of the school still uses booze to bond with teachers and schmooze donors and officials.

  • Bureaucracy — we run into unhelpful lower level officials, only to of course find an understanding higher-up who swoops in to save the day once the protagonist remonstrates and proves their worthiness.

Other TV

Culinary Class Wars — Koreans with nice “Great British Bake Off” vibes cooking food that’s novel to me. Funny how the “Chinese-style” Korean chefs all have a very flat interpretation of Chinese cuisine. Most remarkable is the star judge, Korea’s first three star chef. He has a crazy backstory:

Anh was born in South Korea but moved to the United States in 1993 at the age of 12. He grew up in California where his parents ran a Chinese restaurant. He enlisted in the US Army after high school, and after the September 11 attacks and the start of the Iraq War, asked to serve in Iraq. Anh served as a mechanic, where he helped fuel combat vehicles like helicopters and tanks.After his Army service, Anh was set to attend mechanic school to become a Porsche mechanic. After seeing a group of culinary school students walk by in chef coats, however, Anh changed course and enrolled in culinary school.

Score: 8.3/10

Industry Season 3 — this show is fun, but it’s no Succession. Score: 7.8/10

Sports

Most of the ‘make you gasp’ moments I have following professional sports are surprises about how awful things can get. On a weekly basis for years now, Josh Allen, QB of the Buffalo Bills, has delivered these moments. Watching him is balletic — his invincibility lets you forget all the crushed bodies left in the NFL’s wake.

Dipping into independent fan podcasts after big games or trades is one of my great pleasures. The best sports drama arc of 2024 were the Mets gloating at the Soto signing.

In an increasing number of sports, women’s competition offers the superior viewing experience. This was especially apparent watching the Olympics this year. For men’s indoor volleyball, the jump serves result in too many out serves and dead time, but the women’s game has more action and dramatic points. The same goes for women’s tennis, where Caitlin Clark is a highlight. However, the WNBA is still a little too slow for my taste.

I started playing soccer in adult leagues in NYC this year! It would be fun to get a ChinaTalk team going if there’s any interest…

Movies

We’ve since learned that ChinaTalk doesn’t only get people promoted, it also gets them cast in movies! Phil spent this fall as one of Adam Sandler’s four sons in Happy Gilmore 2. He’s also the lead in an excellent play currently on in Atlanta from March 29 to May 4th.

Billy Madison — the teacher deciding to go out with Billy Madison after he gropes her on the bus is a little bit of a stretch…

Happy Gilmore — I had more fun with this one than Billy Madison. Do they not make short dumb well-made movies anymore, or do I just not watch the ones that come out?

抓娃娃 ‘The Successor’ (2024) — this is one of China’s top grossing movies of the year. The Successor is a ‘Truman Show’ where rich parents decide to raise their child in poverty after spoiling their first kid while training him to take over the family’s business empire. It was interesting to see the backlash to tiger parenting across this movie and the parents WeChat group show. You can stream this movie on Iq.com with English subtitles! Score: 8.8/10

The Divorcee, 1930 — I went on a 1920s kick, which led me to Ursula Parrot’s epic 1929 novel Ex-Wife and its 7.1/10 movie adaptation. The best discovery was this labor of love website, pre-code.com, that I’m excited to dive deeper into.

Anora — I fell asleep in second act. Score: 7.6/10

Video Games

Elden Ring (Levels 60-85)

Parenting an infant considerably shrinks your physical universe — before I became a parent, I was traveling twice a month. Returning to Elden Ring helped me feel less claustrophobic as I made the transition to life within a 5-block radius.

This game is the perfect experience for the pre-VR/AI world, and a marvel of human creativity.

One day I went to the Met Museum after a few hours of playtime, and thought, “Elden Ring’s world is more beautiful and striking than half of these landscapes.” There are few experiences in content consumption quite as satisfying as beating an Elden Ring boss on the 7th try. There were a few times in the past year, in real-life contexts, that I opted to beat “a boss” creatively instead of beating my head against a wall trying the same thing over and over again. Perhaps you could credit my “Elden Ring brain” with helping me find lateral solutions to these problems.

Fromsoft’s triumph is a testament to the magic of maintaining a singular focus on honing a craft for decades. Sure, hardware could not handle something as ambitious as Elden Ring in the past, but the gameplay and design heights of Elden Ring would not be possible had the developer not already made six games in the ‘Soulslike’ genre.

Even if you’re not a big video game person, I’d really encourage you to try to experience the game for yourself. If there’s no chance you will, consider watching this video:

Magic the Gathering: Bloomburrow Quick Drafts

It’s dark magic. Doing drafts gives you slot machine randomness with just enough brain-engagement to trick you into thinking it’s an intellectual activity. Deck-building is genuinely interesting, but then you’re back to casino-like brainstem stimulation during actual gameplay. Generally, there are only one or two moments in each game where you’re rewarded for thinking. The beauty of playing on chess.com is that the apps have a direct feedback loop to teach you how to improve after each mistake you make, which doesn’t exist in MTG Arena.

Von Neumann has a great quote comparing chess and poker: “Chess is not a game. Chess is a well-defined form of computation. You may not be able to work out the answers, but in theory there must be a solution, a right procedure in any position. Now, real games are not like that at all. Real life is not like that. Real life consists of bluffing, of little tactics of deception, of asking yourself what the other man is going to think I mean to do. And that is what games are about in my theory.”

Maybe in-person magic has that in-person bluffing aspect coupled with the feedback of learning by chatting after the games — but playing online has left me hollow. After spending down $20 in gems, I blocked the game.

Black Myth: Wukong.

If you’re looking to appreciate the artistry of the game without having to git gud, check out this 4k bilibili playthrough.

Asian food in nyc superlatives

Read more

AGI and the Future of Warfare

28 March 2025 at 18:39

Part 2 of our interview with Shashank Joshi, defense editor at the Economist, and Mike Horowitz, professor at Penn who served as Biden’s US DAS of Defense for Force Development and Emerging Capabilities. Here’s part 1.

In this installment, we discuss…

  • AI as a general-purpose technology with both direct and indirect impacts on national power,

  • How AGI might drive breakthroughs in military innovation,

  • The military applications of AI already unfolding in Ukraine, including drone capabilies and “precise mass” more broadly,

  • Whether AGI development increases the probability of a preemptive strike on the US.

Listen to the podcast on iTunes, Spotify, YouTube, or your favorite podcast app.

I’m hoping to expand on this show with an interview series exploring AI’s impact on national security. Too often today, debates center on “superweapons” lazily pattern-matched to the nuclear era or go in circles on cyber offense vs defense. The goal instead is to repeat the exercise Dario did for biotech in Machines of Loving Grace: deeply explore the bottlenecks and potential futures across domains like autonomy, decision-support, stealth, electronic warfare, robotics, and missile defense. Guests will be engineers and technologists who can also explore second-order operational and strategic impacts.

But this needs a sponsor in order to happen! If you work at an AI firm, defense tech, VC, university or think tank and want to help facilitate the best conversations about the future of warfare, please reach out to jordan@chinatalk.media.

Military AI and the Drone Revolution

Jordan Schneider: Let’s talk about the future of war. There is this fascinating tension that is playing out in the newly national security-curious community in Silicon Valley where corporate leaders like Dario Amodei and Alex Wang, both esteemed former ChinaTalk guests, talk about AGI as this Manhattan Project-type moment where war will never be the same after one nation achieves it. What’s your take on that, Mike?

Michael Horowitz: There’s a lot of uncertainty about how advances in frontier AI will shape national power in the future of war. I’ve been, historically, extremely bullish on AI from a national defense perspective. I remember when Paul Scharre and I were in a small conference room at CNAS with basically everybody in Washington D.C. who cared about this. Now it’s obviously attracting much more attention.

But I think the notion that AGI will inherently transform power in the future of war makes a couple of questionable assumptions. The first is that AGI is binary and immediately causes a jump in capabilities, which essentially means that you can then solve all sorts of problems you have been trying to solve that you couldn’t solve before.

That might be true. It also might be true that you have continuing growth in AI capabilities that may or may not constitute AGI. In that case, you never have one specific moment, yet you still have ever-increasing frontier AI capabilities that militaries can then potentially adopt.

This other assumption is that technical breakthroughs are the same thing as government adoption, which the history of military innovation suggests is incorrect.

I worry that US companies will lead the world in AI breakthroughs, but the US Government will lag in effective adoption due to legacy bureaucracy, budgeting systems, and the relationship between the executive and Congress.

Maybe the PRC will get there later, but adopt the upside faster.

Jordan Schneider: What is the AI eval that would convince Mike Horowitz that this is the next stealth bomber, or the next nuclear weapon?

Michael Horowitz: I just think that’s the wrong way to think about it. AI is a general purpose technology, not a specific widget. If what we are looking for is the AI nuclear weapon or the AI stealth bomber, I think we’re missing the point.

When general purpose technologies impact national power, they do so directly and indirectly. They do so directly through, “All right, we have electricity, now we can do X,” or, “We have the combustion engine now we can do X.” They do so indirectly through the economic returns that you get, which then fuel your ability to invest in the military and how advanced your economy is. AI is likely to have both of those characteristics.

The thing that’s not entirely clear yet is whether there’s essentially a linear relationship between how advanced your AI is and what the national security returns look like.

Jordan Schneider: Just to stay on defining terms, the direct applications we’re talking about, like the AI for science, are still in the very early phases. There’s a really fun book by Michael O’Hanlon called Technological Change and the Future of Warfare, that includes this cool table of all of the different vectors on which technology can get better over a 20-year horizon.

Show me some crazy material science breakthroughs that you can put in weapon systems, and I will be convinced that this stuff is really real and going to matter on a near-term horizon on the battlefield.

But I think the Dario Amodei framework strikes me as really not grappling with the challenges both on the scientific side as well as on the adoption side. Maybe Shashank, before we do adoption, anything on what this can potentially unlock that we would want to?

Shashank Joshi: We could split the applications of that general-purpose technology up a million different ways. The way I have tended to do it in my head is thinking about insight, autonomy, and decision support.

Insight is the intelligence application. Can you churn your way through satellite images? Can you use AI to spot all the Russian tanks?

Autonomy is, can you navigate from A to B? Can this platform do something itself with less or no human supervision or intervention? The paradigmatic case today, which is highly impactful, is terminal guidance using AI object recognition to circumvent electronic warfare in Ukraine.

The third interesting thing is decision support. This includes things that nobody really understands in the normal world, like command and control. It’s the ability of AI to organize, coordinate, and synchronize the business of warfare, whether that’s a kind of sensor-shooter network at the tactical level for a company or a battalion, or whether it’s a full theater-scale system of the kind that European Command, 18th Corps, and EUCOM has been assisting Ukraine with for the last three years.

This involves looking across the battlescape, fusing Russian phone records, overhead, radio frequency, satellite, IM satellite returns, synthetic aperture radar images, and all kinds of other things into a coherent picture that’s then used to guide commanders to act more quickly and effectively than the other side. That’s difficult to define. But if we’re talking about transformative applications, that is really where we need to be looking carefully.

Michael Horowitz: I agree. General Donahue is a visionary when it comes to what AI application can look like for the military today, and in trying to at least experimentally integrate more cutting-edge capabilities.

To the Dario point though, it’s all a question of timeframe and use cases. If you imagine AGI as instantly having access to 10,000 Einsteins who don’t get tired, then that’s going to lead to lots of breakthroughs that will generate specific use cases.

These could lead to new material science breakthroughs that decrease radar cross-sections, new advances in batteries that finally mean the dream of directed energy becomes more of a reality, or advances in sensing in the oceans that create new ways of countering submarines. It could lead to all sorts of different kinds of things. The challenge is it’s difficult in some ways, ex-ante, to know exactly which you’ll get when.

Jordan Schneider: Perhaps Dario would say that your framework, Shashank, is weak sauce and he’s talking about an entirely new paradigm.

Which of these applications are currently being developed in Ukraine?

Shashank Joshi: The autonomy piece is super interesting because of the pace of change. To sum this up, when I was talking to Ukrainian first-person view (FPV) strike operators, they were saying that if you are a member of an elite unit with loads of training, you can get a hit rate of up to 70-75%. But if you’re an average strike pilot, this is not easy. Sticking those goggles on, navigating this thing — you don’t know when the jamming is going to kick in and cut the signal. You have to get it just right, and you’re getting like 15-20% hit rates maybe.

What I am seeing with the companies and entities building these AI-guided systems for the final 100 meters, 200 meters, 300 meters, and increasingly up to 2 kilometers in some cases, is that the engagement range is going up. You can hone in on the target beyond the range of any plausible local jamming device. That’s a huge deal. More importantly, the hit rate you’re getting is 80% plus. That’s phenomenal. That changes the economics, the cost per kill — that changes the economics of this from an attrition basis.

There are all these interesting ripple effects. You can achieve this with like 30 minutes of training. Think about what that unlocks for a force, particularly sitting here in Europe where we have these shrunken armies with no reserves, with the manpower requirements as well as the training times to bring new people in when you have attrition in a war in the first round.

Subscribe now

This little tactical innovation — terminal guidance, AI-enabled — looks very narrow, but it has all these super interesting and consequential ripple effects on the economics of attrition, the cost per kill, lethality levels, the effectiveness of jamming, and on manpower and labor requirements. That’s why it’s so important to get into the weeds and look at these changes.

Michael Horowitz: We’re seeing at scale something that we kind of thought might happen, but it just always been theoretical rather than something real. The argument for why you would need autonomy to overcome electronic warfare has been obvious for decades. When they were questioning if the technology is there or if we want to do it this way, there were different kinds of approaches.

What we’ve seen is that when you are fighting conventional war at scale, if you want to increase your hit rate and overcome jamming when facing electronic warfare, you can update software to try to counter the jamming. You can try to harden against jamming, and although it increases costs, you can use different concepts of operation to try to get around it, to sort of fool the local jammers.

But to the extent that autonomy becomes a hack that lets you train and operate systems more effectively in much less time — that’s a game changer, and it’s not one that we should expect to be confined there. Imagine all of the Shaheds that Russia fires at Ukraine with similar autonomous terminal guidance out to a couple of kilometers. Imagine all sorts of weapon systems with those kinds of capabilities. We’re seeing this at scale in Ukraine in a way we just had imagined.

To be clear, it’s not just in the air. Let me now just give my 15-second rant on the term “drone.” We are currently using the term “drone” to refer to a combination of cruise missiles, loitering munitions, ISR platforms, and uncrewed aerial systems that themselves launch munitions. All of those are getting called drones right now, even though we actually have correct names for them. It might be helpful if we used those names since we’re talking about different capabilities, essentially. But yeah, plus one to everything.

Shashank Joshi: Mike, do you want me to start using terms like UAS, UUVs? I’m going to get sacked if I start using all those acronyms.

I’d like to ask you about something since you’ve raised the issue of different domains. One of the questions I often get asked by readers is: where are all the drone swarms? Where are the swarms that we were promised? Maybe this is a kind of ungrateful thing because we bank a bit of technology and we are desensitized to it and then we forget everything else.

What interests me is, when writing about the undersea domain a while ago and submarine hunting, I was struck by how difficult it is — this is obvious physics — to communicate and send radio frequency underwater. Radio waves don’t penetrate water very much, if at all. Acoustic modems and things like that are very clunky. So the technologies we have relied on for things like control signals, navigational signals, oversight in the air domain operate very differently in places where signals don’t travel as much — the curvature of the earth or in the water. Do you think that uncrewed technology and autonomy operates in some kind of fundamentally different way in those domains or will it be less capable in those places?

Michael Horowitz: That is a great question. Let me give you a broad answer and then the answer to the specific undersea question.

We’ve essentially entered the era of precise mass in war, where advances in AI and autonomy and advances in manufacturing and the diffusion of the basics of precision guidance mean that everybody essentially can now do precision and do it at lower cost. This applies in every domain — it applies in space, in air for surveillance, in air for strike, to ground vehicles, and can apply underwater and on the surface now. The specific way that it plays out will depend on the specifics of the domain and on what is most militarily useful.

If the question is “Where are the swarms we were promised?” and what we end up with is a world where one person is overseeing maybe 50 strike weapons that are autonomously piloting the last two kilometers toward a target, there may be actually military reasons why we don’t want them to communicate with each other. If they communicated with each other, that would be a signal that could be hacked or jammed, which then gets you back into the EW issue that you’re trying to avoid. There’s an interaction in some ways between the “swarms we were promised” and some of the ways that you might want to use autonomy to try to get around the electronic warfare challenge.

This points to the huge importance of cybersecurity in delivering essentially any of this. Part of the issue is that swarms potentially are vulnerable to some of the same issues that face FPV drones and other kinds of systems in different kinds of strike situations. It wouldn’t be surprising to me if you then see a move more towards precise mass in the context of autonomy without swarming in a world where you think that you’re getting jammed.

Now underwater, you absolutely have the physics issue that you’ve pointed out — communication underwater is just more difficult. To the extent that something like swarming requires real-time coordination, that becomes even more difficult the further away things are from each other. It wouldn’t be surprising then that the underwater domain would be challenging here.

To take it back to the AGI conversation we were having, two notes are relevant here. One, the “where were the swarms we were promised” question reminds me that how we define artificial general intelligence often is a moving target. We’re constantly shifting the goalposts because once AI can do things, we call it programming. Artificial general intelligence is always the thing that’s over the horizon.

Going back to my belief that there might not be one AGI moment, it reflects that way that the definition of AI has tended to be a moving target. But specifically, if you’re in the “AI will transform everything” paradigm, one of the things you would probably try to use AGI to do that would have transformative impact would be to solve some of these communication issues in the undersea domain that can potentially limit the utility of uncrewed systems in mass undersea. That’s an example essentially of a science problem that then maybe major advances in AI help you address when you have paradigm-shifting AI.

Debating China War Scenarios

Jordan Schneider: I want to talk about two odd theories for why a US-China-Taiwan war could kick off. One is China’s dependence on TSMC, and the other is this idea that if one side is close to AGI, then the other would do a preemptive strike to stop their adversary. What do you think about these scenarios?

Michael Horowitz: Those are great questions. A lot of what we know from the theory and reality of the history of international relations and military conflicts suggests that war in either case would be pretty unlikely.

Let me start with the Taiwan scenario. I am extremely nervous, to be clear, about the prospect of a potential conflict between China and Taiwan. There is real risk there. But the notion of China essentially starting a war with Taiwan over TSMC would be kind of without precedent. Put another way, there are lots of paths through which we could end up with a war between China and Taiwan. The one that keeps me up at night is not an attack on TSMC.

Jordan Schneider: Ben Thompson just wrote a piece that defines China’s reliance on Taiwanese fabs as an important independent variable in Beijing’s calculation of whether or not to invade. How do you think about that line of argument in a broader historical context?

Michael Horowitz: The best version of the argument, if you wanted to make it, is probably that if China views itself as economically dependent on Taiwan, it would then seek to figure out ways to get access to the technology that it needs from Taiwan. You could imagine that effort happening in a couple of different ways.

One is to mimic what TSMC does, which obviously they’re already attempting to do. Another would be attempting to coerce Taiwan to get better access to TSMC. But starting a war with Taiwan where tens of thousands, if not hundreds of thousands of people are likely to die, and which could trigger a general war with the United States and other countries in the Indo-Pacific over the fabs — I think that’s relatively unlikely since China would have lots of other ways to try to potentially get access to chips that they need.

Jordan Schneider: You keep coming back to the straw man of going into TSMC to take the chips. But there’s another line of argument that as long as China still needs TSMC and is able to buy NVIDIA 5-nanometer or 3-nanometer NVIDIA chips and needs the output from those TSMC fabs to run its economy in a normal, modern way, then that would drive down the likelihood of China wanting to start a conflict.

Michael Horowitz: There’s certainly an argument in that direction, which is to say that economic dependence could generate incentives to not start a conflict as well. My belief tends to be that the probability of war between China and Taiwan will be driven by broader sociopolitical factors.

Jordan Schneider: I would also agree. Putting on my analyst hat for a second, I can think of several factors that are orders of magnitude more important to China than TSMC in deciding whether to invade — domestic Chinese political dynamics, domestic Taiwanese politics, the perception of America’s willingness or Japan’s willingness to fight for Taiwan.

How about the preemptive strike over AGI?

Michael Horowitz: The argument is insufficiently specified, and I will say my views on this could change. For a situation in which, say, China would attack the United States because it feared the United States was about to reach AGI, that presumes three things.

  1. It presumes that AGI essentially is a finish line in a race — that it’s binary and once you get there, there’s a step change in capabilities.

  2. It presumes that there’s no advantage to being second place, and that step change in capabilities would immediately negate everything that everybody else has.

  3. It assumes that advances are transparent, such that the attacker would know both what to hit and when to hit it to have maximum impact.

There are a lot of reasons to believe that all of those assumptions are potentially incorrect.

It’s not clear, despite enormous advances in AI that are transforming our society and will continue, that there will be one magical moment where we have artificial general intelligence. Frankly, the history of AI suggests that 15 years from now we could still be arguing about it because we tend to move the goalposts of what counts as AI. Since anything we definitively have figured out how to do, we tend to call programming and then say that it’s in support of humans.

It’s also not clear that these advances would be transparent and that countries would have timely intelligence. You need to be not just really confident, but almost absolutely certain that if somebody got to AGI first, you’re just done, that you can’t be a fast follower, and probably that it negates your nuclear deterrent.

If you believe that AGI is binary, that if you get it, it negates everything else, and that there’ll be perfect transparency — in that case, maybe there would be some incentive for a strike. Except that military history suggests that these are super unlikely.

What we’re talking about in this context is a bolt from the blue — not the US and China are in crisis and on the verge of escalation and then there’s some kind of strike against some facilities. We’re talking about literally being in steady state and somebody starts a war. That’s actually pretty unprecedented from a military history perspective.

Leaders tend to want to find other ways around these kinds of situations. If you even doubted a little bit that AGI would completely negate everything you have, then you might want to wait and see if you can catch up rather than start a war — and start a war with a nuclear-armed power with second strike capabilities. It’s so dangerous.

Jordan Schneider: I am sold by arguments one and three. If the story of DeepSeek tells you anything, it’s not even fast following like three years with the hydrogen bomb in the Cold War; it’s fast following like three months with a model you can distill.

If it’s not a zero-to-one thing, then maybe the more relevant data point is Iran and Israel in the 2000s and 2010s. You don’t literally have missiles being fired and airstrikes, but you have this increasingly nasty world of targeted assassinations and Stuxnet-like hacking of facilities.

What is now a happy-go-lucky world in San Francisco could become a lot more dark and messy. Mike, what could trigger that potential timeline?

Michael Horowitz: Let me be more dire and ask, what’s the difference between that and the status quo?

Obviously there’s a difference if we’re talking about assassination attempts and those kinds of things. But every AI company around the world, including PRC AI companies, is probably under cyber siege on a daily basis from varieties of malicious actors, some of them potentially backed by states trying to steal their various secrets.

To me, this falls into a couple of categories. One is cyber attacks to steal things — hacking essentially for the purposes of theft. A second would be cyber attacks for the purposes of sabotage, like a Stuxnet-like situation. A third would be external to a network, but physical actions short of war — espionage-ish activities to disrupt a development community.

On the cyber attack aspect, there is a tendency sometimes to overestimate the extent to which there are magic cyber weapons that let you instantly intrude on whatever network you want. Are there zero days? Yes. Are cyber capabilities real? Yes. Many governments, including the United States government, have talked about that, but I’m not sure it’s as easy to say “break into a network” that is, to be clear, pretty hardened against attack, and just flip a switch like, “oh, today we’re going to launch our cyber attacks."

There are effectiveness questions about some of those things. But also those networks are constantly being tested.

Stuxnet was really hard to achieve. Stuxnet is probably the most successful cyber-to-kinetic cyber attack arguably in known history. It’s this enormous operational success for Israel against Iran.

Jordan Schneider: But the difference with Stuxnet versus what we’re talking about, is that a data center in Virginia or Austin is much more connected to the world. They hire janitors. It’s not like in a bunker somewhere.

Michael Horowitz: Those are more accessible, but there are also more data centers. Targeting any one data center in particular is not likely to grind all AI efforts to a halt.

Frankly, if there is one AI data center that is widely regarded as doing work that will be decisive for the future of global power, that’s going to be locked down. The company will have incentives to lock it down, just like defense primes have incentives to lock their systems down, even if we’re not talking about defense companies. Companies like Microsoft and Google have incentives to lock down non-AI capabilities as well.

My point isn’t that there won’t be attempts and even that some of those attempts won’t succeed, but there’s sometimes a tendency to exaggerate the ease of attack and its structural impact. In a world where we’re talking about hitting a very accessible facility in Virginia, that means there’s probably similar accessible facilities in other places that also can potentially do the job.

Now the toughest scenario is the espionage one where you’re talking about essentially covert operations targeting companies. It wouldn’t surprise me if some of those companies are intelligence targets for foreign governments. The challenge analytically is that these arguments quickly enter the realm of non-falsifiability.

If I tell you that I think this kind of espionage or that kind of espionage wouldn’t be that likely, you could say, “Well what about this?” We’re not going to be able to resolve it with facts. Non-falsifiable threat arguments make me nervous analytically. Maybe this is the academic in me that makes me want to push back a little bit because I feel like if an argument is legitimate, we should be able to specify it in a clearly falsifiable way.

Jordan Schneider: Like what? Give me the good straw man of that.

Michael Horowitz: The best straw man argument would be if you could basically demonstrate that the PRC is not really trying to target for collection varieties of AI companies and that it would be relatively easy for them to do so. That would raise the question of why they’re not doing that now. Then we get back to the question about what’s the point at which they would start those kinds of activities.

They would need to have enough information that they believe some company or set of companies is getting close to AGI, but not enough that they would have done something previously — assuming they’ve got good capabilities on the shelf that they would pull off if they have to.

The non-falsifiable part is about to what extent they could ramp up attacks, to what extent there would be defenses against those attacks, and to what extent those non-kinetic strikes would actually meaningfully delay the development of a technology. Another way of saying this — my prior is that there’s lots of espionage happening all the time. I want to see more specificity in this argument about what exactly folks mean when they talk about escalation.

Jordan Schneider: One of the things that has been remarkable about China, at least how it deals with foreigners, is that you haven’t seen what Russia did with all these targeted assassinations. The sharpest we’ve gotten, at least with dealing with white people, has been the handful of Canadians who were grabbed and ultimately let go after a few years in captivity following the Meng Wanzhou arrest.

People are very focused on China starting World War III out of the blue. But there are also world states in which China becomes much more unpleasant while not necessarily kicking off World War III.

Michael Horowitz: I’m a definitive skeptic on the “China starts World War III over AGI” point. I buy the notion that China could become more unpleasant as we approach some sort of AGI scenario — including non-kinetic activities, espionage, etc. I tend to maybe not view those as decisive as some others do potentially.

You’re right that they certainly could become a lot more unpleasant. If the question is why they haven’t already, the answer is probably twofold. One, there’s an attribution question — suppose Chinese espionage involved doing physical harm to AI researchers or something similar. If they were caught doing that, they’ve now potentially started a war with the United States, and they’re back to the reason why they wouldn’t launch a military strike in the first place.

If it’s non-attributable, then the question is exactly how much are they going to be able to do? I wonder whether there is something about their broader economic ties with the United States that maybe makes some of the worst kinds of these activities less likely in a way that is less troubling to Russia.

Jordan Schneider: This is a decent transition to precise mass in the China-Taiwan context. What can and can’t we infer about military technological innovation in Ukraine to what a war would look like over the next few years?

Michael Horowitz: It’s not necessarily the specific technologies, but it is the vibes. By that, what I mean is the advances in AI and autonomy, advances in manufacturing, the push for mass on the battlefield that we see already in publicly available documents and reporting on how the PRC is thinking about Taiwan. We see that already in the US in the context of Admiral Paparo and Indo-Pacom and the Hellscape concept or something like the Replicator initiative in the Biden DoD — and full disclosure, I helped drive that, so I certainly have my biases.

We see that if you look at some of the systems that Taiwan has been acquiring over the last couple of years. You essentially have a growing recognition that more autonomous mass, or what I’d call precise mass, will be helpful in the Indo-Pacific. It’s unlikely to be the exact same systems that are on the battlefield in Ukraine, but variants of those scoped to the vastness of the Indo-Pacific.

Shashank Joshi: I have a few thoughts on this. One way to think about what Mike is saying is for any given capability, you can have more intelligence that is defined however you like, whether that’s in terms of autonomy or capacity to do the task on the edge at a lower price point.

That capability could be a short 15-kilometer range small warhead strike system in the anti-infantry role. It could be a 100-kilometer system to take out armor with bigger warheads, or it could be significantly longer range systems that have to be able to defeat complex defensive threats. Obviously, the third of those things is always going to be more expensive than the first.

What that revolution in precise mass, if it is a revolution (we can debate if that’s what it is), does is push you down. The capability per dollar is going up and up. That is the essential point.

Michael Horowitz: Just to be clear, that’s the reverse of what we saw for 40 years, where in the context of the precision revolution, you were paying more and more for each capability, whereas now we are seeing the inverse of that in the era of precision mass.

Shashank Joshi: Is the transformative effect comparable across each level of sophistication or capability or range? Are there specific things to FPV-type systems because they, for instance, rely on consumer electronics, consumer airframes, and quadcopters — they can draw upon a defense industrial base or an industrial base that has existed for commercial drones? Is it easier to have that capability revolution for intelligent precision mass at one end of the spectrum relative to building a jet-powered system that has to travel significantly further, has to defeat defense mechanisms, may have to have IR thermal imaging, etc? Is a revolution comparable at each end?

Michael Horowitz: I was with you up until the end about defeating all of those systems, because the thing that’s so challenging about this for a military like the United States is it’s a different way of thinking about fighting. You’re talking about firing salvos and firing at mass as opposed to “we’re going to fire one thing and it’s going to evade all the adversary air defenses and hit the target."

Look at Iran’s Shahed 136. That’s a system that can go, depending on the variant, a thousand-plus kilometers. It can carry a reasonably-sized warhead that, in theory, could have greater or lesser levels of autonomy depending on the brain essentially that you plug into it. That’s not going to be as sophisticated a system as an advanced American cruise missile that costs $3 million or something. But it doesn’t have to be because the idea is that these are complements where you’re firing en masse to attrit enemy air defenses. Your more sophisticated weapon then has an easier path to get through. It’s just a different way of thinking about operating, and that creates all sorts of challenges beyond just developing the system or buying the system.

A Shahed-129 drone on display at an IRGC aerospace fair in Tehran, June 2021. Source.

Shashank Joshi: That’s a really interesting point. This gets us to a phrase, Mike, that is very popular in our world and you and I have talked about this, which is the mix of forces that you have and specifically the concept of a high-low mix. It’s not just that small drones will replace everything. You have a high-low mix where you will have some, albeit fewer, very expensive, high-end capabilities that can perform extremely exquisite, difficult tasks or operate at exceptionally high ranges. Then you will have a lot more in quantity terms of lower-end systems that are cheaper, more numerous, and that will not be as capable — they can’t do things that a Storm Shadow cruise missile or an ATACMS missile can do, but they can do it at a scale the Storm Shadow can’t do or the ATACMS can’t do.

The most difficult concept when I’m writing about this for ordinary people who are maybe not into defense is explaining the mix, the interaction of those two ends of the spectrum. Here’s the difficult bit — pinning down what is the right mix. Do we know it yet? Will we know it? How will we know? Does it differ for countries? That’s where I’m struggling to understand all of this.

Michael Horowitz: It probably differs for countries and even within countries differs by the contingency. For example, if you are fighting, if you’re back in a forever war kind of situation and you’re the United States, then you might want a different mix of forces than if you’re very focused on the Indo-Pacific and on China in particular. What comprises your high-low mix probably changes.

The way that I’ve tended to think about this is you have essentially trucks, which are the things that get you there. Then you have brains, which is the software that we’re plugging in. And then you have either the sensor or the weapon or the payload piece. In some ways, what I think we’re learning from Ukraine that is applicable in the Taiwan context is that sometimes it matters a lot less what the truck is than what the brain is.

Shashank Joshi: The other thing that I struggle to get across is the relationship between the “precise” bit and the “mass” bit, particularly the role of legacy capabilities in this. The great conflict I see today is in the artillery domain. Do strike drones replace or supplant artillery? Strike drones now inflict the majority of casualties in Ukraine, not artillery, as was the case at the early phase of the conflict.

There is this really interesting line in the British Army Review published in 2023 — “There is a danger that the enemy will be able to generate more combat echelons than we have sensors or high-end long-range weaponry to service.”

You can have these remarkable AI kill chains that can spot soldiers moving and feed that data back to your weapons — but if you don’t have the firepower to prosecute those targets and then keep prosecuting them week after week in a protracted conflict, you don’t have deterrence.

That is what we are waking up to now.

Michael Horowitz: The important thing here is the notion that you didn’t need a high-low mix, that you could just go high, presumes short wars where you can just use your high-end assets and sort of shock and awe the adversary into submission. Whether we’re talking about a forever war situation or in the Indo-Pacific, if you’re fighting in a world of protraction, then you need much deeper magazines in all ways, including in your platforms frankly.

Maybe in that world AI is actually helping you with what you’re manufacturing and how you’re manufacturing and can deliver a bunch of other benefits on the battlefield. A big challenge here is that I don’t think these capabilities necessarily mean there’s no role for traditional artillery. Although if they can do the same job better at lower cost, then they will eventually displace those capabilities, or militaries that don’t adopt them will fall behind. We’re more at a complement than a substitute stage right now for those capabilities. But things could change.

The challenge right now for a military like the US is you have all these legacy capabilities, and maybe you wish to invest less in them to be able to invest more in precise mass capabilities, which is something I advocate. But then the question becomes what are you doing with those legacy capabilities and when across what timeframes?

This does tell you some things that are really important from a force planning perspective. For example, the “one ring to rule them all” approach to air combat that led to the F-35, where you’re just going to have one fighter that will operate forever and presumably be useful for every single military contingency. Turns out that means it’s optimized for none of them, and that generates a bunch of risk. One of the things that the new administration is probably going to be doing is figuring out how to address those risks. I don’t want to hate on the F-35 and its stubby little wings. Sorry, I’ll stop now.

Shashank Joshi: I’m really glad you raised F-35 because this gets to another one of my points of thinking. You said, “If it can do the same job,” right? What are the jobs it can and can’t do?

When you look at a simple mission, let’s take an anti-tank guided missile — it’s looking pretty clear to me that a kind of mid-range, low-cost strike system, one-way attack strike system is looking like it can do the job of an ATGM very effectively at significantly lower cost. Therefore, unless the ATGMs also get transformed, you’re going to see a supplanting of roles.

However, I’m also acutely conscious that we are thinking about some battlefields that are in some ways uncluttered. We’re looking at a stretch of the Donbas in which anything that moves is going to be the target that you want to hit — it’s going to be a Russian military vehicle. You’re not going to accidentally hit a school bus full of children. In the Taiwan Strait, you are using your object recognition algorithms to target shipping. You’re not accidentally going to hit something in the context of an amphibious invasion of Taiwan.

Michael Horowitz: Yeah, if the balloon goes up, it’s not like there’s lots of commercial fishing just chilling in the Strait.

Shashank Joshi: Well, and if it is, it’s probably MSS operatives. You’re fine. But in the air power situation too, right? You may be doing stuff like this. However, urban warfare is not going away. I’m imagining fights over Tallinn, over Taipei, over thin, cluttered, complex, multi-layered subterranean environments like Gaza, like Beirut, like places like that. I worry a lot more about the timeline over which autonomy will suffice.

To end on one last point before I spin off the RAF, the Royal Air Force believes that an autonomous fighter aircraft will not be viable prior to 2040. Now maybe that’s ultra conservative, but that’s based on some of the assumptions about the tasks they think it will need to do. I know you have your debate over NGAD and long-run capabilities. So are there limits to this process?

Michael Horowitz: The limits are how good the technology is. Frankly, I’ve argued this for years in the context of autonomous weapon systems. The last place that you would ever see autonomous weapon systems is in urban warfare. Not only whatever ethical moral issues that might surround that, but that’s just way harder to do than figuring out whether something is an adversary ship or an adversary plane or an adversary tanker in a relatively uncluttered battlefield.

Shashank Joshi: The second part was to do with the air picture, right? Countries are now having to decide what our air power is going to look like in 2040, and how much can we rely on the technology being good enough by then? You’re right, that is the question — will it be good enough? But you have to make the bet now. You have to make the judgment now because of the timelines of building these things.

Michael Horowitz: There are two questions there. One is, how good is the AI technology? Frankly, even if Dario is overestimating how quickly we get to something universally recognized as artificial general intelligence — and just to be clear, dude’s way smarter than me and I’m not saying he’s necessarily wrong, just that there’s uncertainty here. He’s super smart and thoughtful.

Side note, the fact that CEOs of today’s leading companies post their thoughts on the internet and come on shows like this is super useful. Now that I’ve taken off my Defense Department hat and I’m back in academia, it’s excellent for understanding how a lot of the people designing cutting-edge technology are thinking about it and interacting with government policy and militaries.

If you’re talking about what the future of a next-gen aircraft airframe would look like or what a collaborative combat aircraft could do, my bet is that the militaries are underestimating how quickly AI will advance and the ability to do that. What they might be accurately assessing is, given the way the process of designing new capabilities works today and given today’s manufacturing capabilities, how long it would take to actually design and roll out a new system.

From a parochial American perspective, shortening that timeline is way more ambitious than what we were attempting to do with Replicator. I expect the new team actually will continue to push forward a lot of those things, whether they call it Replicator or not. You could have a scenario in which your AI is at a level that you believe you could have more autonomous operations of a collaborative combat aircraft, and you have some advances in manufacturing that mean you could produce maybe more of them at a slightly lower price point, but still be unable to do so before 2040 for various bureaucratic and budgetary reasons.

Jordan Schneider: Looking back over the past four years, where are the places that the Biden administration made progress in defense innovation?

Michael Horowitz: We accomplished a lot in getting the ball moving, specifically toward greater investments in lots of important next-generation capabilities like collaborative combat aircraft and varieties of precise mass through the Replicator initiative. But there is a lot more to be done.

It was a journey in two ways. One was bringing everybody along on that defense innovation journey to get to the point where folks bought into the importance of some of these emerging capabilities for the future of warfare, specifically in the Indo-Pacific. Then there’s what we could accomplish before the clock ran out.

To start with the first piece, taking everybody on the journey — nothing happens in the Pentagon without getting lots of people on board. It took a little while for there to be consensus on the state of the defense industrial base. If you look at varieties of think tank reports about what DoD should do differently, there are often all these suggestions like, “DoD needs to scale more of this system, scale more of that system, build more of that.” All of that is great, except that even if you fully fund really exquisite munitions like the Long Range Anti-Ship Missile or JASSM or something like that, you’re going to be really capacity constrained because those facilities just have limits. You can change those limits, but you can’t change those limits quickly.

What can you then do to scale capability relevant for the Indo-Pacific in the short term? The answer is precise mass capabilities — more attritable systems, more AI-enabled and autonomous systems, systems that maybe sometimes, but not always, are built by non-traditional companies. There’s a coalition of the willing that pushed through all of those in a bunch of different contexts, including Replicator, which DIU I think did a brilliant job of leading implementation of, and that generated some growing momentum.

But it also highlighted some real limits. Reprogramming 0.05% of the defense budget to fund multiple thousands of attritable autonomous systems for the Indo-Pacific under the first bet of the Replicator initiative required over 40 briefings to Congress, including a ton by the Deputy Secretary of Defense who’s really busy. Congressional oversight is really important, but the degree of effort required to reprogram essentially less than a billion dollars demonstrates a budget system unable to operate at the speed and scale necessary given the rate of technological change and given the threat that the US is under from Chinese military advances in the Indo-Pacific.

We did some good things. We moved the ball forward. I’m proud of some of the things that we did, but there’s a lot more work to be done.

Jordan Schneider: Ray Dalio recently said that “America can’t produce things, any manufactured goods, cost-effectively.” Shashank, if Ukraine can figure out how to build a million and a half drones a year, couldn’t the US do this too if we were more determined?

Shashank Joshi: Let’s unpack this a little bit. First of all, Ukraine is doing this under wartime conditions. It’s ripping up rules — not all of them, but a lot of them. It doesn’t have to worry about pesky things like health and safety standards, like “Can I build this explosive warhead facility next to this town?” You try doing that in the US or the UK or in Europe. The Ukrainians can get around that because they’re at war, it’s fine.

Secondly, what are they building these things with? The Ukrainian supply chain for UAVs and for lots of other things, including electronic warfare systems, is still full of Chinese stuff. They haven’t got it all out. Yes, there are companies beginning to find alternative supply chains and finding stuff from Taiwan and other countries, but the US government can’t do that. There’s the little pesky matter of the NDAA which prohibits you from just sticking Chinese components into all of your drones and building them out. So you have to get supply chains right. That’s problem number two.

Problem number three is to what standard are you building these things? Are you demanding that they can cope with this level of electronic warfare, radiation hardening, cybersecurity standards? Those are all the requirements being placed on UAS in many Western governments. I won’t speak directly to the US exactly, because Mike knows it better than I do, but certainly in my country I’m aware of this. The Ukrainians can build things to a satisfactory standard that would never stand up to the scrutiny of an accounting or auditing official in a Western country to the same way.

The final point on all of this is the level of mobilization for the defense industrial base is quite different. Ukraine has nationalized its IT sector. It had a brilliant IT sector before the war. It had some great technicians, tech-minded people, software-minded people. That is just not the case in most of our countries. We can’t nationalize our tech industry to go work to build software-defined weapons at a low cost very effectively and quickly. Those are some of the reasons I can see for the discrepancy. Mike, do you disagree?

Michael Horowitz: I think that is all correct. I’m probably more optimistic than you about the ability to build at low cost. Although what exactly you define as low cost? If we think about attritable as replaceable kinds of systems, and what’s attritable probably varies depending on how wealthy you are and how big your defense budget is. What’s attritable for the United States might be different than what is attritable for Ukraine.

There’s a lot of possibility to get lower cost systems. One example is last year there was an Air Force DIU solicitation for a low-cost long-range cruise missile that would be about $150,000 to $300,000 a pop. That’s way more expensive than an FPV drone in Ukraine, but that’s a lot more capability than an FPV drone in Ukraine. Given that existing systems, some of those existing systems cost a million dollars or more — if you can deliver that at what is a fraction of the cost, you’re buying back a lot of mass in a useful way.

The thing that I will be really interested to see, shifting actually to Europe for a second — if you look at what Task Force KINDRED has done for Ukraine, my question is, when is the UK going to buy those capabilities for the UK? If these are militarily useful capabilities for fighting Russia, that seems like something that might be useful for the United Kingdom’s military.

Or say the German company Helsing — they produced 6,000 drones that were going to get sold to Ukraine. Obviously that’s in the context of millions being produced, but presumably those 6,000 are pretty good from a quality perspective. Why isn’t Germany buying those drones?

Shashank Joshi: Helsing is a really good example. In Ukraine, distributed in Ukrainian manufacturing facilities, they are building these UAVs where the hardware is not completely standardized. The software has to adapt to the different airframes being built by different producers. But they’re good enough, they’re doing a good job, they’re making a difference, they’re producing Lancet-like capability at considerably lower cost as far as I understand it.

But Helsing is also building drones for NATO countries in southern Germany in its own factories. The advantage is that it controls the supply chains. It can standardize the process, it can build software-defined weapons in ways where the hardware is optimized to take that on — in the way that Tesla once built other people’s cars and put its code in them. But since that initial phase, it’s built its own hardware because it’s easier to build a software-defined car if you do it yourself. But it’s going to be more expensive and it’s not going to be as cheap and quick and easy as the Ukrainian manufacturing. I believe there are some trade-offs that you see even within the same company.

Michael Horowitz: Absolutely. That also is why you need more open architecture all around and why in some ways the government needs to own more of the IP. If you think about what I said before about trucks and brains — if you’re buying a truck that only can have a brain from the same company, then you’re locked into a manufacturing relationship that’s almost necessarily going to generate higher costs over time than if you can swap out the brain over time with something that might be more advanced. Frankly, whether it’s lower cost or not, it’s probably a better idea to be able to swap it out.

This reflects differences in not just the defense industrial base, but in how the US and western militaries have thought about requirements for capabilities over time in ways that now require challenge.

Shashank Joshi: There was a really interesting speech. The Chief of Defence Staff in the UK, the head of our armed forces, in this case Admiral Sir Tony Radakin, gives a speech every year at Christmas at the Royal United Services Institute, the think tank in London.

Michael Horowitz: Is there port?

Shashank Joshi: There is wine after the lecture, but we’re not drinking port during the lecture.

Michael Horowitz: But I’m imagining like brandy and cigars and port and like a wood-paneled room.

Shashank Joshi: You’re not far off on the room itself.

Jordan Schneider: Are there oil portraits? Do we have like Hague and stuff?

Shashank Joshi: There are oil portraits. I must declare I’m on the advisory board at RUSI, so I’m very fond of it. But anyway, I’ve gone down this rabbit hole. The reason I brought it up is because I wanted to quote a line from that speech he gave at Christmas where he said:

[W]e have only been able to demonstrate pockets of innovation rather than the wholesale transformation we need.

Where we have got it right is because we used an entirely different set of permissions which elevated speed and embraced risk so that we could help Ukraine.

But when we try to bring this into the mainstream our system tends to suffocate the opportunities.

He then proposed a “duality of systems… Whereby major projects and core capabilities are still delivered in a way that is ‘fail-safe’ – clearly the case for nuclear; but an increasing proportion of projects are delivered under a different system which is ‘safe to fail’”.

Mike, this is pretty much what you told me when we spoke a few weeks ago, right? A willingness to embrace failure — not for your Ohio-class SSBNs or your bombers, but for your smaller systems where the cost of failure is not terrible and you need to fail to innovate. That is so much of what it’s going to take for us to be able to be more Ukrainian in our own systems.

File:Royal United Services Institute interior 16.jpg
The interior of the Royal United Services Institute. Source.

Jordan Schneider: Let’s contrast that with a quote on the US side of the pond from Bill LaPlante, who is the DoD’s top acquisition executive during the Biden administration:

“The Tech Bros are not helping us much… If somebody gives you a really cool liquored-up story about a DIU or OTA, ask them when it’s going into production, ask them how many numbers, ask them unit costs, all those questions, because that’s what matters… Don’t tell me it’s got AI and Quantum in it. I don’t care.”

Michael Horowitz: Bill LaPlante thinks that it’s still 1995. The thing that was remarkable about that quote is even at the time it was so profoundly incorrect in the way that it described the ability to scale emerging capabilities. Shortly before we left office there was a speech or maybe it was from under a question or something where he said that he had learned from the Houthis and what they’d done in the Red Sea that it was possible to produce low-cost munitions. Unclear why what was happening in the US had not triggered that revelation. But he got there.

That is the mindset that requires challenge. If you view the only things worth using as an extremely small number of exquisite platforms, then that takes you down a road where emerging capabilities, even those you can scale, might seem less useful because they might require using force differently — if you’re operating at mass rather than just operating those exquisite capabilities.

The bigger challenge though, is every single major defense acquisition program in the US military is either behind timeline, over budget, or both. Most are both. What that suggests is that the current system, which is designed to buy down risk and produce these great capabilities — and it does, but just slower than it’s supposed to, with higher prices than it’s supposed to — in a way that suggests that the current system is not succeeding.

Risk is the right way to think about this. The scale and scope of the challenge posed by the Chinese military is unlike anything I have seen in my lifetime from an American perspective. That means that the assumption that undergirded the 90s, frankly, about the inevitability of American conventional military superiority, is just no longer the case. It’s not just that we can’t sit on our laurels, which is something that I think I wrote maybe a decade and a half ago. It’s that we are being actively pushed and challenged across almost all domains.

What that requires is accepting more risk in the capability development process, which I feel comfortable doing not only because I’m generally bullish on the ability of emerging technologies to deliver, but also because the status quo system just isn’t working.

Shashank Joshi: What we can’t ignore then is what is stopping us — or you in the US case — from taking that risk. Often it’s the politics. You talked about how shifting 0.05% of the budget requires this Herculean bureaucratic political effort on the Hill to plead with Senators and Congresspeople, “Please let me move this $15 million here and there.” That’s not sustainable if you’re trying to make a systemic effect.

You have to have appropriators who are willing to say, “Actually I trust you with this money, and I trust you to be able to spend it in a way that’s flexible and won’t lock you into a spending path for the next six months without wasting it. And let’s test you on that in six months,” but not micromanaging everything.

I don’t know how you’re going to be allowed to have the failure that you need to have the innovation you need if Congress doesn’t trust people to innovate at scale, not just in these little pockets of innovation as Radakin called it.

Michael Horowitz: Every single rule that the appropriations committees have exists because of something that happened in the Department of Defense in the past. To be clear, we are in a different era now with a different set of risks and a China challenge that’s unlike anything we have faced before. We need a new bargain in some ways with the appropriations committee to be able to innovate at the speed and scale we need.

Keep in mind the Pentagon’s budgeting process was invented by Robert McNamara during the Vietnam era and has not changed since then. In the best of times, it is a two-year cycle between when one of the military services decides it wants to invest in a technology and when it gets the money to invest in that technology.

We’ve spent several years of the last decade and a half in continuing resolutions, which means Congress can’t pass and appropriate a budget. This means you can’t start new programs, which then delays adopting new capabilities even further. Something has to give there.

Jordan Schneider: I want to recommend the podcast series Programmed to Fail: The Rise of Central Planning in Defense Acquisition by Eric Lofgren, who’s now working in the Senate. He used to run Acquisition Talk and do shows with me about this stuff. You know, it wasn’t just McNamara — it was McNamara trying to learn from the Soviets.

Subscribe now

Michael Horowitz: The assumption was we’ve got the technology we need. We think that our basic tech development system works. What we need to be able to do is produce this stuff and produce good stuff, and then we’ll beat the Soviets. It worked, but we’re in a different period now.

Shashank Joshi: There’s an interesting book I wrote a review essay on for Foreign Affairs about a year ago by Edward Luttwak and Eitan Shamir, head of an Israeli think tank, called “How Israel Fights.” I don’t find all of it persuasive, but it raises the question of how this country in the 1960s — this agrarian society that is poorer than many parts of Southern Europe in GDP per capita — produced these anti-ship missiles that are able to defeat the Soviet weapons being carried by the Egyptians of that era and the Syrians.

What did they do right? What have they done right? There are many things that they’ve done wrong, and there are many cases in which tech innovation did not help them strategically or even contributed to complacency. But there’s something about that innovation, including innovation under conditions of peacetime or semi-peacetime, that I think we should be thinking about.

Jordan Schneider: I have a five-hour Ed Luttwak episode in the tank that I’ve been dreading editing. But we did get into it, and there is something about this topic. It comes back to some of the Ukraine stuff — Israel is a semi-mobilized society. It’s playing at a smaller scale.

There’s this great anecdote where someone walks into an office and says, “You should arrange the tank this way instead of that way.” Then they do it because somebody thought it was a good idea. You take all his stories with a grain of salt, but still conceptually, the fact that this is all among friends in this small network of, by the way, the best minds in the country.

Michael Horowitz: Whereas in our system, fourteen different people can say no and stop a capability, but no one person can say yes and move it forward.

Jordan Schneider: One of the many shames of the Trump imperial presidency is that despite having enough control of Congress to do this well, getting Pete Hegseth to be the one to lead it is just one of these unfortunate timelines we’re in because the President couldn’t give two shits about this stuff. Maybe there are enough tech people around him though.

Michael Horowitz: Let me muster a point of optimism here, frankly, on this. In the brilliant article that Shashank wrote in The Economist on some of these questions, I sounded a similar note. If you look at Hegseth’s testimony, his discussion of defense innovation is very coherent. He makes points that are not structurally dissimilar to the ones that we have been making for whatever the last period of time has been.

If you look at Stephen Feinberg’s testimony yesterday to be Deputy Secretary of Defense, he actually makes some very similar points, and you hear some of those echoed by various tech sector folks that look to be entering either the White House or the Defense Department.

There is a potential opportunity here for the Trump administration to push harder and faster on precise mass capabilities, on AI integration, and frankly, on acquisition reform in the defense sector, because the president right now seems to have a strong hand with regard to Congress. Whether the president’s willing to use political capital for those purposes is not clear. How the politics of that will play out is unclear. But if the Trump administration does all the things that it says it wants to do from a defense innovation perspective, that may not be a bad thing. There are a lot of things they want to do that are very consistent with things that many of us have advocated for over the years.

Shashank Joshi: I really admire your dispassionate assessment of that and the willingness to think about it apart from the politics. My concern is that you have people who are good at radicalizing and disrupting many businesses and sectors and fields of life. But the skills required to do that are different from the skills needed in a bureaucracy like this.

Just because you were able to navigate the car sector and the rocket sector doesn’t mean you know how to cajole, persuade, and massage the ego of a know-nothing congressman from — I’m not going to name a state because that’ll end up being rude — who knows nothing about this and who simply cares that you build the attritable mass in his state, however stupid an idea that is, and who wants you to sign off on the $20 million.

I worry that they will either break everything — and what I’m seeing Doge do right now with a level of recklessness and abandon is worrying to me as an ally of the United States from a country that is an ally — but also that they will just not have the political mouse to navigate these things to make it happen. Just because Trump controls Congress and has sway over Congress doesn’t mean that the pork barrel politics of this at the granular level fundamentally change. You need operatives, Congressional political operatives, and a tech pro may have many virtues and skills, but that isn’t necessarily one of them.

Michael Horowitz: No argument. There’s a huge gap between being willing to lean further forward on defense innovation and transformation and the ability to bureaucratize, essentially, as a friend of mine is fond of saying, and be able to get the job done delivering. The Pentagon is the world’s largest bureaucracy and it will continue to be the world’s largest bureaucracy even with whatever is happening. That requires a lot of bureaucratic political acumen to be able to deliver results. It is a very open question whether this administration will be able to deliver on that. Frankly, there are early signs that are concerning. But again, it’s still early days.

Jordan Schneider: I want to reflect a little bit about the role of inside knowledge and outside knowledge when it comes to understanding what’s happening in Ukraine as well as the future of war. What does stuff like Shashank’s reporting get? What can it not get? How does all of the open source analysis that’s happening today about the war in Ukraine filter into discussions about budgets in Congress and R&D?

Michael Horowitz: Systematically drawing insights from open source material, both Shashank’s and others, in ways that inform what we do in the Defense Department is important. In the context of Ukraine lessons learned, there are actually a number of different efforts, both classified and unclassified, that try to dive into those things. In the case of Ukraine specifically, there’s heroic effort both inside and outside the context of just the Ukraine desk at the Pentagon or on the Joint Staff to do that.

The challenge sometimes is making some of those insights more visible and then connecting them to the change that you wish to see. Part of the issue is that folks are really busy in ways that are sometimes even difficult to comprehend on the outside. Your read time, even to read things that you really want to read, is just extremely limited.

The role of an influential columnist like Shashank that lots of people trust is invaluable because for a lot of senior folks, that might be the only outside thing they read about defense in a given week. This points to the importance of networks. I think about this a lot from the perspective of how academics should bridge the gap between academic research and policy — networks play a huge role there.

In the case of Ukraine, I think there actually has been really good pickup inside, at least in the US, on what lessons learned look like because of a lot of the great reporting out there. Sometimes people would say, “Why haven’t you purchased this drone that Ukraine uses?” It’s a great question, but it’s really challenging to consume all of the information out there that you should consume. A lot of it ends up getting mediated through staffs.

Jordan Schneider: Here’s a deep cut for you guys. Ian Hamilton, who ran the Gallipoli campaign, was a journalist who covered the war in Manchuria between Imperial Russia and the Japanese and saw the future of war and wrote about it really clearly. But even though this guy was there and then was on the battlefield running it in Turkey, he was not able to instantiate the lessons that he saw firsthand into the way he ended up killing a few hundred thousand people who probably didn’t need to die if he had made smarter decisions. None of this shit is easy.

Michael Horowitz: As part of research for a future book, I was at the National Archives in College Park last week looking for information on US military procurement decisions in the early 20th century surrounding General Purpose Technologies. I found some really interesting back and forth between the War Department and the Wright brothers about the airplane that sounded a lot like modern debates. The Wright brothers are saying, “Well, send us the cash and we’ll send you the airplane.” And the War Department’s responding, “Prove it works and meets these metrics, then we’ll pay you and then you deliver the airplane.” It was like, “Oh dear God, maybe nothing has changed” in some ways in some of these debates.

Jordan Schneider: Shashank, any reflections on the role of popular writing in all this?

Shashank Joshi: I’m amazed by the cut-through we can sometimes get. People will say, “I can’t take my classified system on a plane to read, but I can take a copy of the Economist.” So you suddenly have this responsibility. I’ve had deep experts on something like armored warfare and tanks say, “I’ve been screaming this message into the ether for years, but it was only when you quoted me that this general read me."

We’re sometimes in the strange position of being — I don’t want to say conduits because we would never wish to be uncritical conduits for anything — ways that can short-circuit these networks and cut across them in strange and amusing ways. I have the grave responsibility of not only telling my readers about the big stuff, like what’s Trump going to do next or where’s Ukraine headed, but — if this is not too condescending — feed them their vegetables, make them think about Mike’s essay on precise mass.

Maybe I have to bury it in a piece that’s about the future of drones, but I can make them think about budgeting. Mike tells me you have to understand budgets to understand innovation. Then I think, “Okay, now there’s a challenge. My editors may not like me talking about budgets for a page, but this is my job to get it across and to make people read it and listen to it."

I’m fortunate that I have access to expertise like that of Mike and others to be able to translate that. Fundamentally, Jordan, I see my job as not giving people the answers. It’s just giving them a sense of the debates that the knowledgeable people are having. That’s not to say one person’s right, one person’s wrong, but to say, “Here’s the lay of the land. Here are the arguments on each side. Here are the debates,” and give them a flavor. Let them peer through that window into the world of the conversation that Mike may be having with his colleague on what they disagree on.

Jordan Schneider: What you do, what Mike Kaufman does, what Mick Ryan does, what Mike Lee can do — which I imagine would be harder if you are sitting on the Ukraine desk and your job is to cover what’s happening in electronic warfare — is to think about this all synthetically and across the different domains. In your case, Shashank, even across regions.

This is what I feel like I do in some sense with ChinaTalk as well. Shashank, you have an editor who it seems like you can just bowl through at this point, and I’m very glad you can basically write budget articles. Picking what you want your readers to read and think about is the game. Doing that in a really thoughtful way when there is so much new happening, so many battles occurring in every moment, and so many tactical innovations and counter-responses is essential. This is particularly important for the public at large and also for senior leaders who only have about 3% of their time to really sit down and absorb this stuff — or God forbid, the president.

Shashank Joshi: Or the vice president who has lots of it.

Jordan Schneider: Mike, continuing on, why outside writing matters.

Michael Horowitz: Now that I’ve left the government again, when I think about how, as an outsider, as an academic, to try to influence policy, one way to think about this is: if you want the US Government and the national security arena to be doing something and they’re not doing it, there’s usually one of two reasons.

First, you might be wrong. There could be classified information or some other information you don’t have access to that shows you’re wrong.

Second, you’re right, but your bureaucratic allies are losing. It is hubris, given the size of the national security agencies, including the Pentagon, to think that you have some idea that literally nobody in the entirety of the Pentagon, the intelligence community, and the State Department has thought of. Generally, somebody wants the same thing that you want, but they’re losing. The question is how you give them ammunition to help make the case to move that policy forward.

In the kind of writing that involves advocating for policies or making arguments for policies, I think it makes sense to think about this in terms of how you’re providing support to your bureaucratic allies, even if you don’t know who they are and even if they don’t know you until they see something you write show up in the Early Bird in the morning at the Defense Department or through some sort of press clippings. That’s how I think about the role of outside writing and how you can try to influence policy.

Jordan Schneider: It’s always weird for me when I write something in ChinaTalk and get an email from someone I’ve never met saying, “Thanks for this. This was helpful.” Just putting your arguments with some good, thoughtful analysis out into the ether sometimes works in mysterious ways.

Michael Horowitz: There’s this fiction that you write an op-ed and it somehow ends up on the desk of the president, and then all of a sudden US foreign policy changes. That’s just not how the real world generally works, especially if what you’re trying to influence is policy within a bureaucracy.

I’m encouraged to hear people within the government should be listening to ChinaTalk and getting insights from it. That’s terrific. That’s evidence for this idea that you have allies and you’re trying to help give them ammunition to make the case in whatever fora they’re engaged in.

Jordan Schneider: I got one last question for you guys, if you don’t mind. All of this, compared to reading and writing and doing podcasts about the PLA itself, is just so much more interesting because I’ve read a lot of PLA books and I’ve really thought about doing more shows about it. It’s so difficult to talk in hypotheticals when you’re reading doctrine stuff and doing the OSINT and whatever. It’s just so hard to actually learn things and talk about them in an interesting way. I’m curious if you guys have any advice for me about what better PLA coverage in outlets like ChinaTalk or just more broadly could look like to get people thinking more seriously about all this stuff.

Michael Horowitz: This is super hard. We think about this a lot actually in the context of PhD students, junior faculty, and what the academic China-watching community is doing in this space. Especially given the way that Xi has consolidated control in China, there’s still a lot available, but there’s so much less available frankly than there was 20 years ago or even 10 years ago. That creates analytical challenges because what you can get raises questions like, “Well, why could I get this?” There are essentially, to be really nerdy, selection effects that govern what you’re able to access.

The truth is there’s no military in the world where we probably have a greater uncertainty parameter about its potential performance in a conflict than China’s military — the PLA — because it’s just been decades since it fought. We know what weapons they have. We know what their doctrine says. The ability to put all that together, as we know from the Russia-Ukraine context or any war in history, is very different than what it looks like on paper.

I do not envy the task. All you can do in some ways is acknowledge that irreducible uncertainty and do your best to give folks the information that is available.

Shashank Joshi: I would just add to that: let’s learn from our analytical errors in the past and think about how they might apply. I’ve really enjoyed some of the writing done by people like Sam Bresnick at the Center for Security and Emerging Technology at Georgetown on the way the PLA thinks about AI. He doesn’t say, “Oh, they’re miles behind, that’s useless.” But he does give a flavor of Chinese debate, saying they worry about many of the same things that we do. They worry about explainability, about control, command, oversight. They even worry about ethics — that’s not completely absent. They’ve got issues around compute capacity and all these other things.

It’s just helpful to be reminded that what they say on the page about intelligentized warfare and this and that — they’re grappling with some of the same challenges that we all are. I admire the work of people like Sam and others and his colleagues and many others who think about these from a very fieldwork-based or empirical perspective, getting their hands dirty, reading the stuff, talking to people, looking through journals. That’s great work.

Jordan Schneider: Well, as Sam’s peer advisor in high school when I was a senior and he was a sophomore, I am going to take complete credit for all the brilliant work that he’s done both in Beijing and now in Washington. All right, last thing. Each of you give one book for everyone to read.

Michael Horowitz: My junior colleague at the University of Pennsylvania, Fiona Cunningham, is the most talented academic scholar of the Chinese military of her generation. She has a book that just came out with Princeton University Press titled Under the Nuclear Shadow: China’s Information-Age Weapons in International Security. I would highly recommend checking out Fiona’s book. [We recorded a pod already!]

Shashank Joshi: The orthodox choice is someone who came up earlier — Paul Scharre’s book, Army of None, which I still think is fantastic on the issue of autonomous weapons. It’s brilliant on the history and thinks about it historically — fantastic book for anyone still thinking about autonomous weapons.

But the slightly left-field choice I want to put out there is The Billion Dollar Spy by David Hoffman, which is the story of one of the CIA’s most difficult operations in Moscow running Adolf Tolkachev, a Soviet engineer. The reason it’s relevant to this conversation is it’s about the application of technology to operations — in this case, intelligence operations, running an agent in Moscow, communications technology, miniaturization, the way that the emerging plastics industry and transistor industry affects the CIA’s choices in the ’50s, the way that changes with satellites. I love the idea of thinking about this in a completely different field, intelligence, espionage, and the parallels and ideas that may spark for us thinking about the defense world.

ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.

Pharma Access with Chinese Characteristics

26 March 2025 at 18:54

This continues a series examining China’s role in the global biotech and pharmaceutical industry, and focuses on China’s drug markets. Our first piece explored the role of AI in China’s biotech ecosystem. Future articles will cover the Chinese biopharmaceutical research, innovation, and supply chains. If you have tips on any of these topics, please reach out.

As China prospers, so do its citizens’ expectations for high-quality healthcare. That means expensive drugs – but how can it make them affordable?

The government is transforming the landscape of China’s pharmaceutical markets, cutting drug costs, and pushing domestic firms to become more competitive. But even as policy advances, demand is evolving just as quickly.

As incomes rise, so does China’s disease burden. The country is seeing more of the chronic and complex illnesses common in wealthier nations — cancer, cardiovascular disease, and aging-related ailments — that often require expensive, innovative treatments. Forecast to grow to USD $264.5 billion in 2026, China’s pharmaceutical market is the second largest in the world behind the United States. The US is still by far the world’s largest pharma market, at over US$600 billion today.

China’s healthcare ecosystem remains challenged by corruption and other inefficiencies. Though domestic firms are improving, they are never going to fully meet the growing need for novel biotech and other expensive therapies. And though China’s new policy mechanisms have cut the price of many medications, they can’t keep patients, physicians, and companies happy all at once.

That means China continues to rely on imports for some of the most advanced treatments — those with high price tags that even aggressive government policy can’t cut away. For the sake of its citizens and broader social stability, China will have to navigate the difficult balance between affordability, access, and market power in the global pharmaceutical landscape.

The need for reform

In the 1990s, financial pressure to offset the losses from government funding cuts drove public hospitals to inflate prices and over-prescribe, causing widespread problems. The Chinese Food and Drug Administration (now known as the National Medical Products Administration, or NMPA) suffered from incompetence, corruption, and a large backlog of drug approvals. Pharmaceutical companies marked up drug prices and leveraged kickbacks1 to generate sales. In fact, generic drugs2 — which made up 95% of China’s drug approvals at the time — sold at gross profit margins of 80-90%.

At the same time, between 2009 and 2017, China’s total health expenditure on pharmaceuticals more than doubled, from 754.38 billion RMB to 1,820.3 billion RMB (or about $US110.45 billion to $268.76 billion).3

Data from China National Health Development Research Center

These trends were unsustainable for a growing China.

The Healthy China 2030 plan, unveiled in 2016, recognized the need for reform in the “three spheres” 三医 of healthcare, medical insurance, and pharmaceuticals.

To improve the quality of pharmaceuticals, the NMPA implemented Generic Quality Consistency Evaluations4 in 2015 to assess generic drugs relative to their original brand-name counterparts. In 2017, China joined the International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use (ICH) and started adhering to the ICH-GCP (Good Clinical Practice) guidelines. Additional reforms streamlined and strengthened the drug approval regulatory process.

To improve the affordability of pharmaceuticals, China has begun to leverage two systems: the National Reimbursement Drug List (NRDL) and centralized volume-based drug procurement (CVBP).

National Reimbursement Drug List (NRDL)

Established in 2018, the National Healthcare Security Association 国家医保局 (NHSA) determines which patented drugs are included on the National Reimbursement Drug List (NRDL) and therefore covered under public health insurance schemes, which cover over 95% of China’s population to some degree.

How the NRDL works:

  • The NHSA assesses novel drugs based on safety, efficacy, affordability, and clinical value.

  • Drugs that demonstrate both high clinical value and low cost are categorized as Class A on the List and fully reimbursed by public insurance. Class B Drugs are clinically effective but higher cost, and thus only partially reimbursed.

  • Innovative drugs, orphan drugs, and drugs for rare diseases receive priority.

  • Nevertheless, innovative drugs that exceed a certain price threshold – including the most groundbreaking new cancer treatments – remain excluded.

  • The current 2024 NRDL includes 3,159 drugs, including 91 new additions.

Centralized Volume-Based Drug Procurement (CVBP)

The National Reimbursement Drug List improves access to clinically important drugs, especially new innovations.

China takes a more aggressive cost-cutting approach to generic medications. Piloted in 2018 and rolled out nationwide in 2019,5 the national centralized volume-based drug procurement system 国家组织药品集中采 combines the buying power of all public hospitals and leverages it to negotiate the lowest possible prices from suppliers.6

How CVBP works:

  • Every few months, the government7 invites suppliers to engage in a competitive bidding process for a selection of drugs.

  • Drugs submitted for bidding must first pass Generic Quality Consistency Evaluations, and then win by offering the lowest price.

  • Winning suppliers gain the right to sell a certain volume of medications — usually up to 70% of the previous year’s total consumption — to public hospitals. Domestic firms overwhelmingly win procurement rounds due to their ownership of raw-material manufacturing, large production capacities, and overall economics of scale.

  • The National Health Commission monitors hospitals and physicians to ensure that they prescribe bid-winning drugs.

China's First Bulk Procurement Rounds. | Source: https://doi.org/10.3389/fphar.2021.804237 Figure 1

Since the two pilot programs, the government has held 10 centralized volume-based procurement rounds and 8 iterations of the National Drug Reimbursement List and continues to fine tune both programs.8

Pills without the price pain

So, have the government’s programs worked?

If price reductions were the goal, then yes.

The National Drug Reimbursement List’s first round achieved price reductions of over 50% on treatments for chronic hepatitis B and lung cancer. Subsequent rounds achieved similar reductions in the 50-60% range, as the following chart indicates.

Source: Baipharm, 2025.

The pilot round of centralized volume-based procurement had similar results, with an average price cut of 52% for 25 drugs. A hepatitis B drug experienced the highest price cut — a whopping 96%, for a final bidding price of US$0.09 per tablet. That same drug would cost US$10.94 in the United States and US$15.84 in the United Kingdom.

Examples of Price Cuts from China's “4+7” Pilot Procurement Round. | Source: China National Health Development Research Center

The CVBP program has allowed China to access some of the lowest prices for generic drugs in the world.

Between 2018 and 2023, generic drug manufacturers cut prices to win contracts, on average, by over 50%, and sometimes over 90%. Studies of the “4+7” pilot show evidence that after a procurement round, consumption of quality-assured bid-winning drugs increases and overall drug spending decreases. Here are some notable examples:

  • Insulin: median price reduction of around 42% for a 2021 round applied to 42 insulin products — improving affordability9 and saving an estimated total of US$2.85 billion in the first year of contracts with the winning suppliers.

  • Lung-cancer treatment: price reduction by 83%, from 108 RMB to 18 RMB, saving patients an estimated 8,100 RMB per treatment cycle.

  • Stents to treat coronary heart disease: price reduction by 90%, from over 10,000 RMB (US$1,500) to around 1,000 RMB (US$150). As a result, the number of stents supplied to patients grew, and more medical institutions — including 500 additional second-tier hospitals10 — carried out stent implantation surgeries.

What do lower prices mean?

Government statistics estimated that as of 2022, CVBP created national savings of over 260 billion RMB (approximately $36.3 billion USD). The National Healthcare Security Administration (NHSA), which oversees public health insurance, reallocates over half of its savings from generic drug procurement to cover innovative medicines through the National Drug Reimbursement List.

In terms of affordability, one analysis of the pilot program’s impact found that the proportion of affordable drugs increased from 33% to 67%, and the mean affordability improved from 8.2 days’ wages to 2.8 days’ wages. Urban residents benefited more, a reflection of larger urban-rural healthcare disparities. In theory, improved affordability means patients are more likely and better able to adhere to their prescribed treatments.

Still, impacts on patient health outcomes have yet to be comprehensively evaluated. And the depth of the latest cuts has caused concern among both patients and physicians.

Saving cents, losing sense

In December 2024, the largest procurement round thus far resulted in the steepest price cuts yet, with some products discounted by over 90%.11 Not everyone was happy about this.

“You get what you pay for” is a common sentiment for consumers. So when generic aspirin tablets drop to an astonishing 0.03 RMB (US$0.0041) per pill, people start wondering what was sacrificed in pursuit of lower costs. On Chinese social media, patients have vocalized concerns that the generic drugs are less effective or cause more side effects.

One particular phrase by Zheng Minhua 郑民华 — surgeon-director of Shanghai Ruijin Hospital and member of the Chinese People’s Political Consultative Conference — went viral recently: “blood pressure doesn’t drop, anesthetics don’t bring sleep, and laxatives don’t release shit” 血压不降、麻药不睡、泻药不泻.

The phrase is part of a proposal submitted early this year by Zheng and his colleagues outlining concerns about the efficacy, reliability, and flexibility of bulk-procured generic drugs. Another doctor’s article — which has since been taken down — questions the validity of data published from evaluation trials of procured drugs.12 Some industry participants report that companies are able to replace excipients (non-active ingredients) of drugs after passing consistency evaluations without re-doing testing.

The government has since responded. However, these controversies come at a moment of broader frustration with China’s healthcare system, which stringent COVID-19 lockdown policies and sluggish post-pandemic economic recovery has only made worse.

Can you ever cure corruption?

Even with the government’s attempts to keep drug prices down, per capita medical costs have more than doubled in the last decade. While rising costs likely stem from a variety of systemic and economic factors, in China, corruption is a standout issue that the government can’t ignore.

In 2023, the National Health Commission initiated a crackdown on corruption, including bribery, misuse of insurance funds, rent-seeking by administrative officials, and unethical conduct. Over 160 hospital chiefs were detained.

Widely publicized, this anti-corruption campaign – which rewarded reports of corruption and the imposed strict monitoring of doctors – fueled distrust between the public and physicians.

Importantly, a majority of such medical corruption involves commercial bribery, in which pharmaceutical and medical suppliers give kickbacks to healthcare providers, leading to overprescriptions and inflated costs for patients. Legally, drug companies wield similar influence through heavy marketing, with one report showing that Chinese pharma companies had aggregate sales expenses of about 2.6 times that of R&D. Heavy workloads and low pay make doctors more receptive to drug suppliers’ incentives.

That pharmaceutical firms still turn to bribery and marketing to sway doctors suggests that government efforts to shape drug availability based on quality and cost have yet to fully succeed.

Despite the government’s efforts, domestic policy and industry still today falls short of the Chinese public’s desire for innovative drugs.

Imported drugs still matter

Accessing imported brand-name drugs can be a challengeeven if patients are willing to pay out-of-pocket. One reason may be Beijing’s focus on boosting Chinese drug manufacturers, actively cultivating their capabilities and favoring them during CVBP and NRDL selection processes.

This tension played out strikingly through Pfizer’s Paxlovoid, a treatment for severe COVID-19. Paxlovoid received conditional regulatory approval in February 2022 – so it was authorized to sell in China – but despite high demand, the drug never got included on the National Reimbursement Drug List.

So what happened during NRDL negotiations?

Officials cited Pfizer’s high asking price as the reason for no deal. Meanwhile, two COVID-19 treatments did make the list – the traditional Chinese medicine Qingfei Paidu and the domestically-produced antiviral pill Azvudine – reinforcing evidence of China’s preference for cheaper homegrown treatments.

Pfizer defended its price and rejected the notion of a future deal for domestic manufacturers to distribute a generic version of Paxlovoid in China, a common practice for low- and middle- income countries. “They are the second highest economy in the world and I don't think that they should pay less than El Salvador,” said CEO Albert Bourla at the time.

The Chinese public was left to deal with the fallout. Facing extreme shortages, people turned to the black market and to unproven Indian generic versions that sold for as much as 50,000 RMB, over twenty times the original price.

The case of Paxlovoid highlights both the extent of China’s progress and the gaps that still exist.

Yes, the domestic pharmaceutical industry is becoming increasingly formidable. But some important innovative drugs still remain beyond China’s borders, leaving it dependent on imports.

The value of China’s imports of finished drugs has been increasing over time. | Source: UN Comtrade

The Chinese government has made substantial – and in some cases, excessive – strides in ensuring medicine is affordable for its citizens. But novel pharmaceuticals come with high price tags that can’t be easily negotiated.

In 2024, the size of China’s innovative drug market13 reached a milestone of over 100 billion RMB (about USD $13.89 billion). Out-of-pocket payments or commercial health insurance covered over half of that market. It’s clear that China’s national medical insurance programs aren’t alone enough to meet China’s medical needs – especially if the goal is to provide the same level of care to its whole population that is expected in a G7 country.

Industry projections expect China’s combined market for innovative drugs and medical devices to exceed 1 trillion RMB (USD $137 billion) by 2035 – a full 30% of the global pharmaceutical market.

Is this a market opportunity for multinational companies? Or, will China be able to achieve its ideal world: a thriving domestic pharmaceutical ecosystem, managed by rigorous national insurance and bulk procurement programs, so that its citizens get the medicine they need.

ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.

1

Between the 1990s and 2018, provincial governments organized drug procurement. A “two-envelope” process awarded supplier status to pharma companies who met basic quality standards and offered the lowest prices. However, medical institutions retained the power to sign supplier contracts, which governments struggled to regulate. As a result, kickbacks, uneven access, and high drug costs continued.

2

A generic drug has the same active ingredient(s) and is bioequivalent to a brand-name drug, but is often more cost-effective and sold under a different name once the original drug’s patent expires.

3

As measured by China National Health Accounts. In 2017, drug spending made up 34.42% of China’s total health expenditure, meaning over a third of all healthcare costs were for medications.

4

The GCE tests generic drugs along two dimensions relative to their original brand-name counterparts, known as originators or innovators. The first is in vitro pharmaceutical equivalence (does it have the same active ingredient, dosage form, strength, and quality standards?) and in vivo bioequivalence (does it perform the same way in the body?). The US FDA and other regulatory bodies use similar frameworks to approve generic drugs. In 2017, the NMPA began adhering to international standards for pharmaceutical development and registration, which continues to this day.

5

The government’s “4+7” pilot was implemented in four municipalities and seven cities before rolling out nationwide in 2019. The 4 municipalities were Beijing, Shanghai, Tianjin, and Chongqing. The 7 cities were Guangzhou, Shenzhen, Xi’an, Dalian, Chengdu, and Xiamen. In total, they make up around one-third of the national market.

6

China’s use of pooled procurement for medicine pricing, by the way, is not unusual. Many countries, including Indonesia, Canada, and much of Europe, leverage bulk purchasing power and negotiations to reduce drug prices for public medical institutions.

7

The NHSA designs and oversees policy, the National Health Commission 国家卫健委 oversees hospitals, and the Joint Procurement Office (JPO) oversees implementation of centralized volume-based procurement, which is operated day-to-day by the Shanghai Pharmaceutical Centralized Bidding and Purchasing Management Office.

8

To mitigate risks of shortages, centralized procurement selection allows for multiple winners and runner-ups. In cases of medical necessity, hospitals may purchase a small number of drugs outside the program. And to compensate for lost income, hospitals receive government subsidies and charge higher service fees.

9

A study of the round assessed affordability as the number of daily wages needed by the lowest-paid unskilled government worker to obtain a 30-day supply, and found that insulin had changed from about 1.63 days’ wage to 0.68 days’ wage. Relative to costs of insulin in other countries, the reduction accomplished by China shows some improvement.

10

China’s hospitals are organized by a three-tier system based on their care, education, and research capacities.

11

778 products (the highest number of products so far) and 493 companies were involved in this procurement round.

12

The proposal points to evidence such as the fact that patients using bulk-procured anti-blood clot medication had higher incidence of strokes and pulmonary embolisms. It also expresses the need for greater flexibility given to physicians and hospitals if a bid-winning drug doesn’t apply to a specific medical case, or isn’t the right formulation (such as for pediatric use).

13

“Innovative drug market” includes newly developed pharmaceuticals that introduce novel mechanisms of action, significantly improve clinical outcomes, or address unmet medical needs. This market is typically measured by the revenue generated from patented, first-in-class, or breakthrough drugs.

幸运鬼和理性怪?

我有时候觉得自己简直幸运地像个鬼一样。

我只知道有开心鬼,不知道有没有幸运鬼,倘若有,那可能是我。

说像鬼一样,是很多时候我都幸运地有点离奇乃至诡异了。感觉只要出门,就有幸运的事情发生,甚至有时候在家坐着,幸运也能从天而降。我有时候都想说:幸运之神也不必如此捂了嚎风地眷顾我。

今天又发生了一个离谱事件,我正在上的非常intensive(强度巨大)的荷兰语课迎来了又一次考试,这次考试又重要又难,我的同班同学大部分是难民,倘若这次考试通不过,很有可能就会收到官方通知说无法继续下一个难度的荷兰语课程,那就迟迟无法开启在荷兰用荷兰语上学的进程,整个生活进展都会受到巨大的影响。

因为我是自愿报名的,所以对我影响没有特别大,但是我这个人吧,该说不说,热爱学习,(注意:以下开始都不是好词,很容易陷入优绩主义的陷阱)上进心强 ,有完美主义倾向,还有点强迫症,就是不太能接受自己在考试的时候不爽:爽的意思就是要游刃有余,挥洒自如,下笔有神,愉悦流畅。所以考试前我也在疯狂学习,每天学习十来个小时,即使每学几个小时就起来运动一下,但是还是搞得肩颈和背剧痛,眼睛也感觉要瞎了。人真的不能每天坐着学习那么久。

到了考试当天,我早早起床,吃了早饭带了午饭来学校接着复习(吃饭很重要),准备把前些天逐字逐句苦心孤诣复习的几十篇文章全部再跟着音频通读一遍,加强一下短暂记忆。结果我戴着耳机正在读的时候,突然有一个陌生人(应该是跨性别者,男跨女)走向我并问我:你也是在学荷兰语并且一会要考试吗?

我说是的,她立刻开始和我说她刚考完,然后开始和我说考了哪几篇文章,哪些文章完全不用看。

我真的当场震惊,想说上天和这位陌生的朋友也对我太好了吧!

当我俩正在激动讨论哪篇考了,哪篇没考时,又来了一位女性,是她的同班同学,她俩开始一起和我说每一篇文章具体考了什么和答案分别是什么!

当她们帮我回忆完所有文章和答案离开后,我真的坐在座位上进入了贤者时间,感觉一切好不真实。

过了好一会才反应过来,感觉像是被幸运的陨石砸中了。

后来我仔细想想我幸运的原因:这一切并不是平白无故的,它归根结底是因为我早起早去了学校,坐在了考场门口,并在朗读课文(当然没影响其它任何人哈,是一个在路边半封闭的,允许发出声音的空间。其它还有全封闭的地方是不能发出声音的,我没有坐在全封闭的地方)。如上这些条件真的缺一不可,因为我隔壁空间也有个女孩,她也在复习荷兰语,但是她没出声读,所以这两位幸运天使就走向了我,而不是她。

我在这里面无意识的情况下释放了非常积极主动的信号,且选择了正确的时间和地点。没有这些积极主动和正确,这些幸运绝无可能发生。

于是我当场得出结论:幸运是对主动者的奖励

我想完了这些后就陷入了纠结,还剩俩小时,接下来还接着学习吗?朋友劝我吃吃饭玩一会,但是我想了一下还是决定放下手机,按照原定计划把文章全部通读复习一遍,包括那些刚刚两位天使说绝对不会考的文章。

毕竟学好了,学熟了都是自己的,考试只是手段,学会才是目的。我应该把时间花在目的上。

所以我就接着在那个半封闭空间听着听力进行跟读,中间还有我同班同学看我太沉浸式学习跑过来吓我。

等到考试的时候,不仅同学们紧张,老师都特别紧张,我就不太紧张,主要是幸运天使光临过我。

结果考试一开始,我点开题目一看,第一篇出现的,就是两位幸运天使和我说绝不会考的文章!接下来10篇,没有1篇是她们刚考过的!老师换了全部的题目

我就在考场感到庆幸:我才是自己的幸运天使!我的理性决定才是我当下幸运的最大来源

因为我认真复习了,且在考前还在全盘复习,我才能在发现所有题目都和我知道的不一样的时候,还能情绪稳定地把题目做完,把百分之九十以上的考题全做对。

所以幸运到底从何而来呢?

我用我一天坐过山车一样的故事总结一下:从主动和理性中来,缺一不可!

幸运鬼也得是理性怪,幸运鬼必须得是理性怪。

幸运只会频频降临在理性思考的头顶之上

BTW:放学以后的春日回归播客,不出意外,将在4月1号深夜,4月2号(下周三)凌晨准时上线更新。4月,让我们在诗意盎然中相见!

China on that Signal Chat

25 March 2025 at 19:04

Taiwan and China on the Leak

Lily Ottinger and Nicholas Welch report:

The Trump administration leaked its plans to bomb the Houthis by accidentally adding The Atlantic’s editor-in-chief to a Signal group chat. This situation calls for humor, so we curated a mini-roundup of reactions from the Sinosphere. Enjoy!

From the Taiwanese platform PTT:

(Pushed) “I wonder what the American ancestors would think when they see these unworthy descendants.”

(Pushed) “I shouldn’t, but I really want to see the Indo-Pacific war plan leaked…”

(Pushed) “This is too ridiculous. Could it be because the reporter just happened to have the same name as the person they wanted?”

Indeed, such an error would be more difficult to make in Chinese given the huge variety of characters used in Chinese names. From Weibo:

American names are too simple. … 😂

This is truly an epic level of face-losing disgrace. (这是史诗级别的丢人现眼)

The following comments come from another Weibo post about Trump’s response to the leak:

“I don’t know, that didn’t happen” has hoodlum energy (無賴嘴臉)

Shameless people are invincible (人不要脸,天下无敌)

Another commenter in that thread invoked the phrase, “Not listening! Not listening! The turtle is chanting buddhist scriptures!” (不听不听王八念经). This meme is originally from a nursery rhyme satirizing people who shut down when confronted with an opposing viewpoint.

Some other relevant Chinese idioms:

  • 泄露天机 (xiè lòu tiān jī) — “To divulge the will of heaven”

  • 掩耳盗铃 (yǎn ěr dào líng) — “Covering one's ears while stealing a bell”

  • 事后诸葛亮 (shì hòu zhū gě liàng) — “To become Zhuge Liang after the fact” — Zhuge Liang was a military strategist widely regarded as a genius thanks to his portrayal in Romance of the Three Kingdoms. This is a bit like saying, “Everyone’s an Einstein in hindsight.”

Emergent Ventures Grants for Taiwan

Jordan Schneider reports:

Tyler Cowen’s Emergent Ventures fellowship is a special institution. It is a no-overhead fellowship where Tyler reads short essays and gives grants to individuals to support personal projects and career development. Emergent Ventures recently received a donation to start finding talent in Taiwan and is looking for applicants. If you are or know anyone who could use a grant to get a personal project off the ground, do consider applying and select Taiwan on the regional dropdown menu!

For some inspiration, see here for an article on Tyler’s selection critera and a site that has compiled brief descriptions of all the past winners who’ve done things like use AI to decipher ancient text, compose Bach-style fugues, cover local politics, and a thousand other things.

Emergent Ventures helped get ChinaTalk off the ground, giving me a grant in 2018 to buy microphones and Chinese lessons. I’ve since attended three conferences (which I count as some of the most exciting weekends of my life), and met up with winners from around the world. What I’ve taken most from this community is a renewed optimism in what the future can bring and an ambition that I can continue to improve and broaden what I do and really make a dent.


Mark Carney — China Nerd?

Irene Zhang reports:

It’s official: Mark Carney is Canada’s 24th Prime Minister. The prominent economist and political novice, who led the Bank of Canada through the 2008 financial crisis and headed the Bank of England from 2013 to 2020, won the incumbent Liberal Party’s leadership contest on March 9 and was sworn in as Prime Minister on March 14. (In Canada, the leader of the largest party or coalition in Parliament becomes Prime Minister, making Mark Carney an oddity as a PM who doesn’t represent a constituency of his own.)

Britain voted to leave the European Union in 2016, three years after Carney became the first foreign head of the Bank of England. In the ensuing years, Carney worked to open the UK to new trading opportunities — including China. He traveled to Beijing with Britain’s then-finance minister in late 2017 to seek a $1.34 billion trade deal, though a UK-China free trade agreement never materialized. Along the way, he’s had to wade through choppy waters. In August 2019, as Chinese paramilitary troops amassed near the Shenzhen-Hong Kong border and protests in Hong Kong reached a boiling point, Carney awkwardly pulled out of a high-profile dinner in London with China’s ambassador to the UK. He faced criticism from the British political arena for bending over backwards to promote Chinese investment at a time when China was repeatedly violating human rights in Hong Kong.

Subscribe now

Carney’s successful banking career comes with a special level of insight into China’s economic situation. He was bullish on the Chinese economy as of 2012, calling on Canadian firms to adopt China strategies and “reorient” toward China. But at least as early as 2018, he became deeply concerned about China’s domestic financial situation. During a BBC interview, he warned that China was one of the top risks to the global economy because its financial sector “has many of the same assumptions that were made in the run-up to the last financial crisis.” He elaborated on this in February 2019, telling a Financial Times audience that the possibility of a Chinese economic slowdown the second-most-important risk to global growth:

“While China’s economic miracle over the past three decades has been extraordinary, its post-crisis performance has relied increasingly on one of the largest and longest running credit booms ever, with an associated explosion of shadow banking. … The Bank of England estimates that a 3% drop in Chinese GDP would knock one per cent off global activity, including half a per cent off each of UK, US, and euro area GDP, through trade, commodities and financial market channels. A harder landing would have significantly larger effects, as these channels would likely be accompanied by negative spillovers to global confidence.”

Later that year he called the world’s reliance on the USD as the reserve currency “risky”, but said the Chinese RMB was far from ready to step in as a replacement.

Carney arrives in Ottawa with some knowledge of the top man himself — he met Xi Jinping twice while heading the Bank of England: once in 2017 and once in 2019. Last year, when rumors about him running to replace Trudeau first began to swirl, Carney met Xi Jinping again at the China Development Forum in March as head of Bloomberg’s board. These previous run-ins may not count for much in the short term, as Canada-China relations remain on ice. Just one day before Carney was elected to lead the Liberals — and by extension Canada, for now — China slapped 100% tariffs on Canadian rapeseed oil, oil cakes, and pea imports, as well as a 25% tariff on seafood. These are in retaliation to the 100% EV tariffs and 25% steel and aluminum tariffs Canada imposed on China last October, a coordinated move with the US at a time when relations between Washington and Ottawa looked very different.

What do you do when you get into a tariff war with your second-biggest trading partner to help your biggest trading partner, but now your biggest trading partner keeps talking about annexing you while trying to decimate your economy? You elect a central banker, of course.

Carney’s technocratic appeal was already compelling to voters as Canada struggled with productivity stagnation — now, the trade war is only adding to his political edge. A federal election is required to happen in Canada before October 20 this year. Before Trump made anti-Canadianism his newest fixation, the incumbent Liberals were set for massive losses while opposition Conservatives hoped for a majority in Parliament. But Conservative leader Pierre Poilievre’s previous amiability towards the American president has practically doomed him to quisling status only two months later, and now some polls predict Carney’s Liberals could even win a majority if the election were called today. The events have been quite jaw-dropping, and the polls still contain much volatility, but if Carney manages to cement his mandate, he will lead Canada through unprecedented headwinds on both American and Chinese frontiers.

Mood Music — Canadian Disco:

Why China Doesn’t (Really) Frack

Caleb Harding reports:

China’s natural resource endowments have been succinctly summed up as “Rich coal, poor oil, small gas” (“富煤、贫油、少气”). Despite claiming “small gas,” however, China actually has the world’s largest reserves of shale gas, an unconventional type of methane deposit that must be tapped with hydraulic fracturing. China is one of only four countries in the world with commercial fracking operations (the others being the USA, Canada, and Argentina), which began after President Obama shared fracking technology with the PRC in 2009.

President Obama with Hu Jintao in Beijing, November 17th, 2009. Source.

Chinese media acknowledges that the shale gas revolution was a game changer for U.S. energy supply and security, but China has yet to realize its own revolution with its immense reserves — why? The simple answer is geology.

China’s shale reserves are deeper, more scattered, and in more mountainous terrain than those in the United States. Guo Tonglou 郭彤楼 is a chief engineer at Sinopec, one of just two companies involved in commercial scale extraction of shale gas in China. Commenting on China’s situation, he said that, “Although we are a country rich in shale gas resources, shale gas extraction in our country is difficult and costly. This means that it is impossible for us to adopt the same development approach as the United States" (虽然我们是页岩气资源大国,但我国页岩气开采难度大,成本高。这就意味着,我们不可能采取美国那样的开发方式).

Shale wells in Sichuan (home of China’s main shale plays) and Texas. Cherry-picked, yes, anecdotal, yes, but makes the geological difference a bit more visceral.

Other barriers exist, including a lack of natural gas infrastructure and imperfect legal/policy frameworks, and China has been working to address those issues.1 But the fundamental issue remains the economics and geology at the well — if it was cheaper to extract, China could rapidly expand their natural gas infrastructure, just as the U.S. did in the decade after 2005. But that seems unlikely anytime soon.

But that hasn’t stopped Beijing from announcing ambitious targets for fracking. In 2018, the Chinese Ministry of Finance and the State Administration of Taxation introduced a preferential tax policy to reduce the resource tax on shale gas production to 4.2% from 6.0%, which in 2023 was extended through December 2027. Following the release of China’s 14th Five-Year Plan in 2021, policy directives continued to support the development of unconventional natural gas resources, including adjustments made to the pipeline tariffs last year, and a new subsidy policy released just this month that incentivizes higher production.

The Sichuan basin represents most of China’s shale gas production. Source

Rather than hoping for a shale revolution, Guo Tonglou’s aspirations for Chinese shale gas are modest — he’s optimistic that, with more technical breakthroughs, shale gas could account for ⅓ to ½ of Chinese natural gas production. That would be an increase of their current production from shale from 25 bcm to 100-200 bcm annually, or from ~0.6% to ~ 2-4%2 of China’s overall energy. In contrast, the U.S. produced 836 bcm3 of natural gas from shale formations last year, which could provide around 30% of the U.S. overall energy.

But is the juice really worth the squeeze? With a comparative advantage in renewables and coal, why invest in such a difficult domestic energy source? As one CRS report estimates, China’s demand for gas through 2030 can be met under existing contracts, and demand through 2040 can be met without having to increase trade with Western suppliers.

Economically, it doesn’t seem to make much sense. But in the age of great power competition, economics are no longer the primary concern — security is. And as the world’s largest importer of LNG, China has valid concerns. Shale will likely never reach the scale in China that it did in the U.S. But until China feels totally secure in its energy supply, its development will continue.


Chinese is Hard, Even for Natives

Moly from Chinese Doom Scroll reports:

Question: “Why does the 8th line on the Chengdu subway have an announcement reminding people to not bite fellow passengers?” [撕咬 — sīyǎo, to bite.]

A response from Chengdu Subway’s official account: “Uh! Sweetie. The subway is saying, ‘Do not broadcast noise from electronic devices to annoy fellow passengers.’ [滋扰 — zīrǎo, to annoy. In local dialect, it might be pronounced very similarly, I have no idea.] Create a civilised city, be a civilised citizen. The subway invites you to travel in a civilised way.”

“Not only are you not allowed to bite fellow passengers, you also have to be careful of the spy between the train and the platform.” [“Spy” “奸细” jiānxi here sounds like “gap” “间隙” jiànxì]

“I’ve also wondered forever why Chengdu’s East Station bans Cadillacs. [凯迪拉克— Kǎi dí lā kè] Does it actually mean something?”

“Are you sure it’s not, ‘No solicitations?’” [喊客拉客—hǎn kè lā kè]

“Shanghai buses play announcements, ‘Please dance when the door opens.’ It’s very confusing.”

“‘Please be careful as the doors open.’ They should also broadcast the same thing in Shanghaiese.”

“I came to Shenzhen for uni and heard the subway broadcast, ‘Watch out for kasayas,’ [袈裟 — jiā shā] And thought it meant to watch out for scammer monks or something.”

“So what was it actually?”

“To avoid getting pinched? [夹伤 — jiā shāng] Probably?”

“It’s okay, at least it’s better than the Shanghai subway. The announcement there is, ‘No bathing, performing, social, San Francisco.’”

“Huh? So what the hell is it talking about?”

“No begging, performing, selling, or passing out ads.”

“So what does, ‘Please go through the fried chicken passage.’ mean?” [炸鸡 — zhájī]

“Go through the turnstiles.” [闸机 — zhájī]

“There’s a minnan dialect announcement that just goes Lou Cha Lou Cha endlessly…”

“Lou Cha means depart hahahahahaha”

“Our maths teacher: Do not bite the classroom.” [干咬 — gānyǎo. I’m pretty sure it was supposed to be 干扰 — gānrǎo, disrupt the classroom.]

“The Beijing subway told me to not use yellow corpses.”

“Then I figured out it meant, ‘Do not use child leashes.’”

“One time, I was passing by an intersection and heard someone calling out, ‘Crazy~ Bread rolls~ Crazy~ Bread rolls~’ And I was like, ‘Just how crazy can bread rolls get? I gotta check it out!’ So I turned the corner to see the sign on the car, ‘Honey bread rolls.’ Someone was just calling out ‘Honey~ Bread rolls~’ with a dialect.”

“I’m the only one here with a straight answer. Biting is too cruel. We don’t engage in such inhumane methods. We usually prefer to swallow whole.”

ChinaTalk is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

1

For example, in 2011 China’s State Council approved changing the legal status of shale gas from a “natural resource” to an “independent mining resource,” which expanded private companies access.

Bidding opportunities went from being closed to foreign entities, to encouraging joint ventures.

In 2014, China’s National Development and Reform Commission issued a new policy that requires that pipeline operators “provide unused pipeline capacity to new customers on a fair and nondiscriminatory basis,” making it easier for natural gas operations to gain access.

2

Natural gas makes up about 8% of China’s energy supply, at 394.5 bcm. About 59% of that, or 232.4 bcm, is produced domestically. 25 bcm, or 11%, came from shale gas.

Assuming:

  • National energy consumption remains constant, with 8%/394.5 bcm = 0.0203 %/bcm representing the conversion from bcm of natural gas to percentage of China’s overall energy

  • Other sources of natural gas (currently 207 bcm) remain constant

  • Shale gas rises to constitute half of domestic production (thus also reaching 207 bcm)

Shale gas would then be 207 bcm * 0.0203%/bcm = 4.2% of China’s overall energy. If instead it was ⅓ of domestic production, that would be 104 bcm * 0.0203%/bcm = 2.1% of China’s overall energy.

3

In 2023, about 78% of total U.S. dry natural gas production (37.87 trillion cubic feet = 1,072.4 Bcm) was from shale formations. Thus shale gas = 1072 * 0.78 = 836 bcm. The U.S. is a net exporter of natural gas, which makes percent of overall energy from shale gas calculations more challenging. But as a back of the envelope calculation, with 38% of energy from natural gas, and 78% of natural gas from shale, shale provides roughly 30% of the U.S. energy.

Inside the Soviet Cold War Machine

24 March 2025 at 18:49

Sergey Radchenko’s book, To Run the World: The Kremlin’s Bid for Global Power, is a masterwork! In my mind, it’s in pole position for best book of 2025. Sergey takes you into the mind of Soviet and Chinese leaders as they wrestle for global power and recognition, leaving you amused, inspired, and horrified by the small-mindedness of the people who had the power to start World War III.

We get amazing vignettes like Liu Shaoqi making fun of the Americans for eating ice cream in trenches, Khrushchev pinning red stars on Eisenhower’s grandkids, and Brezhnev and Andropov offering to dig up dirt on senators to help save Nixon from Watergate.

Sergey earns your trust in this book, acknowledging what we can and can’t know. He leaves you with a new lens to understand the Cold War and the new US-China rivalry — namely, the overwhelming preoccupation with global prestige by Cold War leaders.

In this interview, we discuss…

  • Why legitimacy matters in international politics,

  • Stalin’s colonial ambitions and Truman’s strategy of containment,

  • Sino-Soviet relations during the Stalin era and beyond,

  • The history of nuclear blackmail, starting with the 1956 Suez crisis,

  • Why Khrushchev’s vision of abundance couldn’t save the Soviet economy.

Co-hosting today is Jon Sine of the Cogitations substack.

Listen now on iTunes, Spotify, YouTube, or your favorite podcast app.

Why Authoritarians Need Legitimacy

Jordan Schneider: Let’s start with prestige. How do you define it, and is it anything bigger than just schoolyard dynamics?

Sergey Radchenko: That’s a deep question, and that’s where the Soviet experience is not actually unique. If you think about the Soviets pursuing prestige, what’s so unique about that? Everybody does that. People strive for status, people strive for recognition, and states do the same thing.

That is why in the introduction to my book I try to zoom out a little bit — I’m trying to say is that the striving for recognition is not something unique to Soviet leaders. In fact, I show that even during the Cold War, the Soviets were not unique. The Chinese also had similar motivations and sometimes also the Americans.

We’re starting to throw around difficult terminology — recognition, prestige — what do those things mean? To add to the complexity, what the Soviets really wanted was legitimacy. Everyone wants legitimacy, but the Soviets had a particular need for legitimacy perhaps because of their lack of internal legitimacy. They had a project on their hands. They wanted to pursue a communist revolution. They were trying to sell it to the Soviet people saying, “Look, soon we will be living under communism. Everything will be so fine. We’ll all be so wealthy.” The abundance, material abundance was supposed to come soon — and it didn’t.

It didn’t arrive because the Soviet project had already started losing steam by the late 1950s. It became increasingly clear that this was happening, creating a deficit of internal legitimacy. Because of that, they wanted external legitimation. They wanted others to say, “You’re great, you’re wonderful, you’re a superpower,” because then they could turn to their own people and say, “Look, we are so strong, we’re a superpower, we’re great because they say so.“

Who are they? Of course, the ones in a position to recognize — the Americans. Others mattered too. If Mongolia or Papua New Guinea said, “This is the great Soviet Union,” that was important. But if the Americans said that, it was all the more important because they were themselves a superpower, so their recognition was absolutely crucial for the Soviets’ sense of greatness. That is where those things connect — prestige, recognition, external recognition, and legitimacy.

Jon Sine: I’m trying to understand what legitimacy from an international perspective gets you. I’m reminded of the phrase, “Foreign policy is domestic policy by other means.” How is a foreign policy of recognition — this external validation — important from a domestic standpoint? Is this what is driving it, or is there some other source driving this need for recognition and legitimacy?

Sergey Radchenko: There may be an acceptable definition — legitimacy is legality and justice put together somehow. It’s a term that reflects legality — a legal position — and justice, meaning you deserve that position.

The Soviets really wanted to be seen to occupy a place they deserved. They tried to sell this to their own people, saying, “Look, we are in this position, we’re great, we deserve to be great, they recognize us as great, we are legitimately great.” This is to be distinguished from raw power.

I show in the beginning of the book how Stalin sometimes actually wanted legitimacy instead of raw power. In some cases where he controlled certain territories, he would back off because he wanted legitimation of his control — legitimate control, not just absolute control by sheer power.

I can give you a couple of examples. One example would be the separatist rebellion in Xinjiang. Stalin supported the rebellion in Xinjiang but then backed out. He basically sold the rebels down the river and engaged in a relationship with the central government in China. Although that meant he lost immediate control of Xinjiang, he acquired a certain sense of legitimacy for his other claims in China.

That is where power and legitimacy were in a state of interplay, and sometimes — not always — but sometimes Stalin would prioritize legitimacy over power. That’s not to say he would always do that. There was a certain bottom line where he would say, “No, I refuse to give up that particular thing because it’s absolutely essential for my security or what I feel is necessary for the USSR.” But sometimes he would say, “I will give this up if you recognize that whatever else I hold here is legitimate,” and that is the essence of the Yalta framework that was constructed in 1945, which Stalin was very fond of and wanted to preserve.

Jon Sine: Staying on the Yalta agreement, one of the things that quite surprised me in the book is how interested Stalin was in a concert of powers and having his view of the USSR’s proper sphere legitimated. Could introduce the percentages agreement?

Sergey Radchenko: The percentages agreement is a well-known episode in the early history of the Cold War. Churchill went to Moscow and met with Stalin in October 1944. Churchill in his memoirs recounts it by saying basically, “I wrote down percentages of influence that the Soviet Union and Great Britain would have in southeastern Europe, and then I gave the paper over to Stalin. Stalin made a tick on it, handed it back to me, and I told Stalin, ‘Maybe we should destroy this piece of paper, because isn’t that very cynical that we basically decided the fate of millions of people in this way?’ And Stalin said, ‘No, you keep it.’”

What’s interesting — this story, of course, is decades old. We’ve all known about it from Churchill’s memoirs. But this is where writing international history becomes really interesting — now we have the British version of the record of conversation, we also have the Russian version of the record of conversation. We have the actual piece of paper in the Churchill archives, the percentages agreement, and crucially, something that has not really been considered before: we have extensive discussions that followed that agreement between the Soviets and the British about specific percentages.

When I was reading those discussions, I thought, “This is so strange, the Soviets really took this stuff seriously.” They argued about an extra 5% in Hungary, this kind of stuff. You think, “Really? What does it even mean?” It sounds bizarre, it sounds absurd, but that shows that Stalin really took that seriously. He felt that he had a legitimate sphere of influence in Eastern Europe. He was very cynical about it, but he basically expected the British and the Americans to defer to his desires in Eastern Europe, and in turn, he would accept the British and American sphere somewhere out there.

He wanted his sphere recognized externally. The problem was that what Stalin considered a legitimate sphere of interests or sphere of influence was not seen as legitimate by the Americans, in particular. That explains a lot about the way the Cold War unfolded.

Jon Sine: Throughout the rest of the book, it’s not clear if any of this is really worth it. But in this particular instance, when you talk about America pushing back on Stalin, you do seem to think that it was very necessary for America to have done that. Basically the Cold War, in a sense, was not inevitable from an ideological perspective, but rather happened because Stalin was going to push until he found resistance. Could you elaborate on your views on that?

Sergey Radchenko: It’s a difficult question. Historians have looked at the early Cold War and have tried to understand whether Stalin was perhaps open to compromise. Some have said that he was looking for great power compromise, and I’m also inclined towards this view. Others would say, “No, we had to stop Stalin’s expansionism, and Stalin could never be trusted,” and so on.

The problem from the US perspective is we’re all wise in retrospect — we can look at the Soviet archives and hopefully understand what Stalin actually was thinking. But the problem at the time is you are facing this difficult person called Joseph Stalin. Nobody knows what he’s thinking. Nobody knows what his ambitions are. Does he want to take over the world or not? Does he want to take over Europe? I show in the book that Soviet post-war planning entailed Soviet control over much of continental Europe, as far as Sweden. They had great appetite.

There’s great uncertainty. You’re in the present moment, trying to deal with the situation, you don’t know what the other guy thinks, so what is the safe policy to pursue? I cannot imagine a safer policy than that which was actually pursued in terms of what George Kennan formulated in The Long Telegram, the policy of containment, the policy of pushing back.

Did that help bring about the growing distrust with the USSR and in some ways precipitate the Cold War? Maybe. Of course, if you’re pushing back on a guy, he’ll think, “Look, they’re aggressive, they’re pushing back on us.” But if you don’t push back on the guy and it turns out he’s actually determined to take over the world, you’ve got a problem.

In the end, what happened — containment — is, I guess, the best worst solution in the situation that we had in the early Cold War. It’s a tragic situation. It’s a situation that results from the lack of trust and from inability to understand or know in real time what the other side is thinking. We can project this onto our own days and ask similar questions, but maybe we can do it later.

Jordan Schneider: It’s fascinating how you see these leaders ending up resorting to the lowest common denominator of understanding the other side. Stalin basically says, “The West only understands force. When you have an army, they talk to you differently. They will recognize and learn to love you because everybody loves force.” Truman basically did the same thing. He was like, “I got a bomb, they gotta listen to me. I can bully him however I want.“

There are moments over the course of the Cold War where leaders start to feel their way to slightly different dynamics, where it’s not just this whole idea of prestige and national power as this gas that inevitably expands until you hit another great power that’s going to push back on you. Perhaps there are different ways in which you can look at those points of contention and try to devise different operating systems.

It’s remarkable and almost speaks to the argument of prestige being this gaseous force that these two countries who had just fought a world war together were not able to figure out how to get their interests to align in a way that wouldn’t lead them down to what the world ended up having to live with for the next 50 years.

Sergey Radchenko: It’s partly that, Jordan, and partly just a different conception of the world. We can get attached to the person of Stalin and say that he was paranoid and crazy, but if you look at Soviet post-war planning, it was based on a kind of 19th century imperial model. The Soviets get to control a lot of territory like the Russians when the Russian empire was expanding. Stalin was thinking in 19th century terms, very unremarkable given that particular historical period.

Churchill was like that, by the way. The Americans were not thinking in those terms. They were coming from a different perspective, and they did not want to allow the Soviets to basically put Eastern Europe into their pocket and walk away. From that, you already see a clash of visions for the post-war order, multiplied by the lack of trust, multiplied by security concerns. You can see how you would have the slide towards confrontation

Churchill and Stalin in Moscow in 1942. Source.

There were also other factors. One point I raise in my book is that early on — 1945, 1946 in the early post-war world — Stalin was hopeful that the communists would come to power in many places unaided. People would just vote for them because they were so popular and wonderful. Then it turned out that in Eastern Europe nobody thought that, so he had to falsify elections in Eastern Europe. This raises a question. It becomes much more difficult to pretend you can have some kind of cooperation or a coalition of governments.

The same thing happens in Western Europe where the Italian Communist Party and the French Communist Party were initially very popular but then ultimately had to effectively leave the government. Stalin realizes that he will not be able to obtain more pro-Soviet governments or Soviet leaning governments through coalition partnerships and resorts to brute force. That’s another factor I wanted to highlight.

Jordan Schneider: If Stalin is expecting 70% of Europe and the West is expecting 70% of Europe, then was there ever really a way to find some happy medium if their desires of what they wanted the world to look like after 1945 were fundamentally incompatible?

You zoom in very directly to 1947, where Stalin wakes up to the fact that unless he starts playing hardball, he might be losing on a lot more fronts than he was hoping to. But in the hypothetical where communism happens to be more popular and win elections, is this something the CIA even stands for in the first place? It’s hard to imagine, if that is Stalin’s base case, that you really could have ended up anywhere else.

Sergey Radchenko: Exactly. That’s why I’m so cautious. I’m not a revisionist historian saying, “If only the Americans behaved themselves better and understood Stalin or accepted Stalin’s sphere of influence, then it would all have turned out fine.” Maybe it would have, but another thing we need to keep in mind is that Stalin measured his appetites in part as a result from the policy of containment.

If the Americans were not pushing back, who knows whether his appetites would have stayed where they are or would have extended even further. That is why I’m saying we don’t know, and he himself may not have known at that time. That’s why containment was the safe policy.

It is a tragic outcome. Yes, it does contribute to the Cold War, but who could come up with a better policy than containment when you don’t know who you’re dealing with on the other side? Frankly, looking at Stalin’s paranoia and his dark view of the world and his obsession with conflict, it would have been a very unwise policy to try to indulge Stalin too much. I’m saying that as somebody who’s read a lot of Stalin. I think containment was a wise strategy.

Jordan Schneider: I love this little riff you go on in your book, which is a historical wrinkle I wasn’t aware of, where Stalin was like, “Let’s see if we can get some colonies in Libya.” Can you explain that?

Sergey Radchenko: It’s remarkable, isn’t it? Absolutely remarkable. Here again, 19th century European imperialism and colonialism — that’s how Stalin was thinking. “Oh, Africa? Of course. Those other guys have colonies in Africa. Why shouldn’t we have a colony in Africa? We can try our hand.” I think Molotov said something along the lines of, “We can try our hand at colonial administration.”

They, by the way, early on felt that the Americans had promised them something along those lines. Then when James Byrnes became Secretary of State, it turned out that he didn’t want to go in that direction, and Stalin felt his expectations were betrayed.

Why did he want this colony in Africa? Strategic reasons maybe. More important are the issues of prestige that we talked about, because if you’re a great power, of course you want to be in Africa. That’s where all the great powers made their imprint. That’s why he wanted to go there.

Jordan Schneider: Let’s turn to Asia. What happened with Hokkaido?

Sergey Radchenko: The Hokkaido story is fascinating. Basically, Stalin got this idea that the Soviets should help in the liberation of Japan and proposed effectively landing on the island of Hokkaido.

Remember, in 1945, the Japanese controlled not just Hokkaido, but also what they called Karafuto, or southern Sakhalin Island — that’s the place where I grew up — and the Kuril Islands as well. The Soviets liberated those islands as part of the Yalta agreement, which was stipulated in the agreement.

Stalin then said, “We can also land on Hokkaido. There’s a wonderful island off Hokkaido. Let’s liberate Hokkaido as well.” He proposed that and sent a telegram to Truman. Truman basically said, “There’s no way this is going to happen,” and remarkably, Stalin backed off.

You have to ask the question, why did he back off? The obvious answer — we’re talking about mid-August 1945 — would be that he was afraid of the atomic bomb. That’s one possible answer. Is it credible? There’s something to it.

Another possible answer is that he was interested in preserving some sort of cooperative relationship with the United States. The Yalta agreement was in place, the Americans respected it, and he wanted that recognition and that legitimacy offered by Yalta.

In fact, he continued for some years, until 1949, almost 1950, to abide by the principles of Yalta because he wanted American recognition of the legitimacy of his gains. When Truman said, “No, sorry, you cannot. That’s just too far. You were never there, do not even come close,” Stalin backed off.

Subscribe now

It’s interesting — it wasn’t just fear of the nuclear bomb, it was also concern about maintaining a cooperative relationship with the United States. It could be both. History is multi-causal.

Socialist International Relations(唇亡齿寒?)

Jordan Schneider: Another example of Stalin not going all in right after 1945 was the way he handled Mao and the nationalists. Can you talk about how Stalin handled China throughout the civil war?

Sergey Radchenko: That is a very interesting aspect of Soviet foreign policy which I think is understudied. If you look at histories of the Cold War, they all focus on Europe, on the German question more or less, and few people actually examine China.

What you see in China’s case is something quite remarkable. Soviet involvement with China predates the Second World War and goes back to the 1920s. The Soviets were helping both the Kuomintang and at one point were crucial to the establishment of the Chinese Communist Party. Then you have this period of prolonged confrontation between the Kuomintang and the Chinese communists.

The Chinese communists, following the Long March, retreated towards Northwestern China and were headquartered in a place called Yan’an. That is where the end of the Second World War finds Mao. He receives a telegram from Stalin in August 1945 saying, “Go negotiate, go to Chongqing,” which is the military capital of China, “Negotiate with Chiang Kai-shek for a coalition government.“

You can imagine how Mao would react to this. “What? Hang on just a second. We are not friends. We have had a very difficult relationship.” Yes, there was a period of united front and during the war against the Japanese they even cooperated in some ways, but Mao was not keen to have this sort of relationship with Chiang Kai-shek or to enter into a conversation with him.

Mao was basically forced by Stalin to do that because Stalin had concluded the Treaty of Alliance with Chiang Kai-shek. That was a result of very painful negotiations in Moscow. Chiang Kai-shek sent T. V. Soong, who was the head of the Executive Yuan, to Moscow for those negotiations.

What Stalin was able to wrestle from Chiang Kai-shek was not only concessions in China — there’s a discussion of the railroad that the Soviets sold to Manchukuo but now wanted back — but also, crucially, he got Mongolia. The Soviets had been controlling Mongolia for a long time, but de jure it remained a part of China. Stalin wanted Chiang Kai-shek’s government to say, “Okay, we give up on Mongolia.” That was arranged in Moscow in 1945. These were massive concessions.

What did Chiang Kai-shek want in return? I mentioned one of those demands: the end of Soviet support for Xinjiang separatists, and Stalin agreed. Stalin claimed, “We don’t know what’s happening in Xinjiang. This is not us.” Although the whole war was being fought with Soviet weapons and instructors. He said, “No, no, no, that’s not us.” Basically, he gave up and betrayed those separatists in Xinjiang, saying, “Okay, we’re done with you.“

Then the question of Mao Zedong came up, and Chiang Kai-shek said, “Look, you’re supporting Chinese communists.” Stalin replied, “Well, Mao is not really a communist, he’s just a nice guy but he’s not a communist.” As a result, Stalin got his gains in China, he got Mongolia, and he basically told Mao to go negotiate with Chiang Kai-shek.

Stalin had to understand what this meant. If Mao Zedong negotiated with Chiang Kai-shek for a coalition government — we’ve been there, we’ve done that already before. Mao did not have a superior force, and Stalin understood this. Stalin thought that the Kuomintang was a much stronger force in China in 1945. This basically meant surrendering to the Kuomintang. I think Mao also understood that.

There’s this very interesting period from 1945 until the Chinese Civil War broke out again where Stalin was uncertain who he would support or what his position in China should be. He was interested in a solid relationship with the Kuomintang because they had guaranteed his control of Mongolia and those parts of his imperial interests, ratified by the Yalta agreement. That was much more important to him than Mao’s cause and communism.

Later he changed his mind, but that was because the Cold War started to unfold in Europe and he thought that maybe he should have stuck with Mao Zedong after all. But that took place over a period of time, which is why Mao later said on numerous occasions, “Stalin blocked our revolution and regarded me as a half-hearted Tito.” Mao had every right to say that, because Stalin did not believe in the victory of the communist revolution. When it happened, it was a great surprise for him.

Jon Sine: America also sent George Marshall (there’s a great book on this called Marshall’s Mission to China) to try to ensure this compromise between the communists and the nationalists. You’re telling us we have Stalin pushing on the communists to make this work. So what goes wrong?

Sergey Radchenko: What doesn’t work is that Mao does not want to have anything like that in place. Stalin has limited leverage over him — he cannot force him to agree to a coalition government. He can only advise him to have a coalition government.

They have long discussions in the fall of 1945 when they meet in Chongqing. There’s extensive discussion there about what future China might look like. The Chinese communists actually propose effectively dividing China along a north south scenario where the north is controlled by the communists and the south is controlled by the Kuomintang.

Chiang Kai-shek cannot accept this because he is a nationalist, a patriot. He does not want a division of China like that, so he turns this down. They never could agree, and then ultimately you’ve got George Marshall’s intervention in the period of quiet.

Later, Chiang Kai-shek blamed George Marshall for interfering too much, claiming that if his hands weren’t tied, he would have been able to wage war much more effectively against Mao Zedong and would have defeated him.

That’s a whole different story as to why nothing works out and why the civil war is won by the communists. Many historians have written about this — it’s not part of my book — but there were serious problems with the Kuomintang government and the state of the Kuomintang army. That’s the bottom line.

Jon Sine: Let’s fast-forward to the late ‘40s, at which point the communists are moving into Manchuria. You have this piece of evidence — I think it’s Stalin writing to a general operating in Manchuria on behalf of the Soviets — and he says something to the effect of, “If Mao and the communists move in and try to take any of the material, open fire or use force to prevent this.” This really speaks to your narrative. Could you explain what that source is and what Stalin was doing there?

Sergey Radchenko: That comes from one of Stalin’s telegrams. I think it was to Rodion Malinovsky, who later became the Soviet Minister of Defense if I recall correctly.

The situation was that Chinese communists were trying to get into Manchuria because that’s a very important base to have, and they’re trying to occupy Manchurian cities. The Kuomintang government was obviously determined to prevent them from doing that.

The Soviets were actually in control at that time in Manchuria before they ultimately withdrew. Stalin instructed his forces on the ground to deny Chinese communists the ability to enter the cities. This is remarkable if you think about it because, aren’t they supposed to help Chinese communists? They’re the Soviet Union, a communist country — aren’t they supposed to help communists?

The communists say, “We want to get into the cities, we want to control Manchuria,” but Stalin sends a telegram saying, “Keep them out. Open fire if they try to do that.” I can’t remember the exact term that Stalin used — it’s like “these people” or something like that. It’s as if he’s saying, “I don’t trust these guys.” He says, “They are trying to get us into conflict with the United States.“

Is he right? Of course he’s right. The Chinese communists would have been very happy at this point if the Soviets and the Americans came to blows. But that’s still too soon right after the end of the Second World War, and Stalin is of the mind that he should: A) avoid a conflict with the United States, especially over China, and B) work with the Kuomintang government.

He is pursuing a policy of great duplicity. He is tolerating the Chinese communists in the countryside, but he’s keeping them out of the cities. The Soviets were helping communists by providing them supplies over the border. I don’t want to say that he completely rejected the Chinese communists — he was basically playing both sides.

It’s very clear that he was in fact trying to keep communists from taking over cities in Manchuria, and I think that is really a new piece of evidence that contributes to our understanding of what his China policy was from 1945 until 1946

Jon Sine: Let’s zoom out and talk about Stalin’s China policy from a macro perspective. I was reading Stephen Kotkin’s book, the first volume of his Stalin biography, and he basically pillories Stalin’s China policy. He looks into what the Soviets were themselves talking about back in the ’20s when Stalin was suggesting that the communists should ally with the nationalists, which ended in an absolute massacre of the communists by the nationalists.

Kotkin’s view is that Stalin’s China policy has been a string of really atrocious decisions about how to operate there. How does this ultimately reverberate with Mao, thinking about his position vis-a-vis Stalin, someone who has really screwed him in a number of ways, and who then has to deal with Mao coming into power in ‘49? Mao has to maintain a sort of subservient position to Stalin as the long-term leader of the communist camp. How does that all play into this?

Sergey Radchenko: Clearly Stalin and Mao did not really like each other, and Mao had very few good things to say about Stalin. However, he understood that Soviet support was really important for the long-term survival of the communist regime in China and for recognition of the communist regime. Here I mean literal recognition because once the Chinese take over, they need recognition from the Soviets.

The Soviets were very keen to offer that right away, although not without some side stories. For example, when the Chinese communists take over Nanjing, the Soviets follow with the retreating Kuomintang to Guangzhou. The Chinese communists are asking, “What are you doing? Even the Americans are staying in Nanjing!” The Soviets respond, “You don’t understand strategy.”

For Mao, it was essential to get Stalin’s support, Stalin’s diplomatic recognition of the Chinese communists as a legitimate regime, but also aid — Mao counted on Stalin’s aid in the reconstruction of China and on Soviet protection.

There are all kinds of good reasons why, despite his not particularly good feelings towards “Comrade Main Master,” as Mao called Stalin, he still went along and respected Stalin. I think I say in the book that it’s almost in the way that, in a large family, perhaps a son would respect his father in a Confucian way, although Mao always liked to emphasize that he was not a Confucian. It’s almost like he’s paying respect without really liking Stalin very much at all.

I don’t know if you saw that in the book, but I dug out this absolutely hilarious document which I quoted from. Mao sent his wife, Jiang Qing, to Moscow, and she was there in 1949 undergoing medical treatment. She wrote letters to her “beloved chairman” and these letters are in the Stalin archive in Moscow, which obviously shows that they were being read and delivered to Stalin.

The funniest thing is, Jiang Qing surely knew that they were being read, so she would start the letter with, “Oh, my dear beloved chairman, I miss you so much, loves and kisses.” And then she would say, “Oh, we have to be very clear in our condemnation of Tito’s revisionist clique” or something like that.

Sergey Radchenko: It was very important for Mao to actually prove to Stalin that he was a loyal supporter, that he was not a Tito who could betray him. That’s what he tried to do. Very pragmatic on his part.

Mao did declare in 1949 that he would “lean to one side,” the Soviet side, which led later to historians debating this moment, asking, “Was there a lost chance? Was there an opportunity for the Americans to stay in China and avoid 20 years of disengagement?“

Back in the ‘90s when these debates were being held, the conclusion by most historians — people like Odd Arne Westad and Chen Jian and others — was that no, there was no lost chance because Mao was so ideologically connected to Stalin. Also, he had in mind this project of Chinese communist revolution that was so important to him. He had to “clean the house before inviting guests,” i.e., kick out the Americans and then carry out these revolutionary transformations.

I am not 100% sure about this because I don’t think anything is definite until it actually happens. My own view is that perhaps if the Americans had a different approach to Communist China at that time, there would’ve been an opportunity to establish relations.

Stalin and Mao coordinated their policy on this question. The expectation on the part of the Chinese communists would be that the Americans would derecognize Taiwan effectively, or derecognize the Kuomintang government and extend recognition to the communist government. That was the key issue for them. Beyond that, it’s not like they were absolutely determined to break diplomatic relations with the United States.

Jordan Schneider: This older brother, younger brother, father-son dynamic between Stalin and Mao plays out wonderfully the way you talk about the face-to-face meetings they have. Mao gets really disrespected the first time he goes to Moscow. Stalin makes him wait for a while and basically says no to everything he asks. Slowly but surely over the course of Stalin and even more dramatically with Khrushchev, the national power balance as well as the revolutionary legitimacy (i.e. who is leader of the communists) shifts over time.

Let’s maybe use that as a way to start talking about the Korean War. How did their relationship and this Chinese-Soviet dynamic play out in 1950, which brought about that horrific conflict?

Sergey Radchenko: The Korean War has been studied by so many people. We have had a lot of documents that have been studied closely by historians since the 1990s on the Chinese side, the Russian side, and even on the North Korean side. This is really a well-studied conflict.

I wondered if I’d have anything new to say about this. In a massive book like this, I had to say something about the Korean War. I was actually quite lucky in the sense that I did find some new evidence. It wasn’t a smoking gun type evidence, but it came close. I’ll explain the nature of the evidence.

Basically, the story goes like this. Kim Il-sung in North Korea wanted to reunify the country and kept asking Stalin for permission, saying, “Comrade Stalin, the moment we cross over the 38th parallel, there will be revolution in South Korea. Everything will turn out just fine. It’ll be very quick.” Stalin would refuse him permission to do that time and again. The reason for that is pretty obvious — Stalin was worried about American intervention. He was a very cautious individual in this particular instance.

Then he changed his mind. It’s not exactly clear when he changed his mind, but we know that in late January 1950, Kim Il-sung was at a reception. Kim was a little bit drunk, went up to Shtykov, who was the Soviet representative in Pyongyang, and said, “We want to reunify South Korea.” Shtykov reported that to Stalin. Stalin sent a cable back to Shtykov for informing Kim Il-sung that “this matter requires preparation.” That is already a yellow light, not a red light.

Subscribe now

Later, Kim Il-sung went to Moscow, and they effectively agreed that the invasion would start. He changes his mind around late January 1950, or at least he tells Kim Il-sung around that time.

Various theories have been advanced. The most obvious is that at that point, there’s already going to be an alliance with China, so as a worst-case scenario, the Chinese could fight this war if Kim makes a mistake, or if the Americans intervene. But I think if Stalin thought that the Americans would intervene, he would never have authorized Kim Il-sung to do that.

The question is, why does Stalin change his mind from thinking that the Americans might intervene to thinking that they will not intervene? That is where it becomes complicated.

First of all, we have Dean Acheson’s remarks in the press conference, which are straightforward, where he says, “America has a defensive perimeter, which does not include Korea.” That is probably the most misguided statement ever made by an American foreign policymaker. That, in retrospect, was a very bad idea.

Annihilate the American aggressors
A propaganda poster in support of North Korea. The title reads, “Annihilate the American aggressors!” ca. 1950. Source.

Beyond that, there’s been speculation by historians that Stalin had spies who had access to debates at the center of American power in Washington. The problem with that is, of course, we didn’t have any evidence. We can just say, “Stalin has spies everywhere. That’s how he knew about certain things like NSC-48, for example.“

What I found was very interesting. I was reading a discussion between Mao Zedong and Anastas Mikoyan in 1956 during the 8th Party Congress in Beijing. Mikoyan comes to Beijing in September 1956. They’re having discussions. Stalin is dead for more than three years, right? Suddenly the discussion turns to North Korea, and Mao says, “Why did you agree to let Kim Il-sung cross over and start the war?“

When I saw this, I thought, “That’s a moment right there. That’s very interesting.” Even Mao himself did not know what was going on. Stalin did not inform him.

Then I see Anastas Mikoyan’s response, which is, “Our intelligence intercepted cables by the Americans that said that they would not intervene in the conflict.” That is super interesting because you get that one little piece, one acknowledgement, one little piece of information that you can weave into the narrative and say, “The role of intelligence was very, very important.” This is basically from the horse’s mouth.

Mikoyan says that three years after Stalin died, “Actually, the reason we did this was because we misinterpreted the intelligence that we intercepted.” That is where it all goes wrong.

The nature of the intercept, I couldn’t find that in the archives. By the way, this document of Mao Zedong’s conversation with Mikoyan is actually from the Chinese archives, not the Russian archives. When I went back to the Russian archives, I couldn’t really find anything about that. It’s probably still somewhere locked away in the KGB vaults.

Jon Sine: My thinking is this is one of the places where there might be a lesson for today that screams at you most clearly. Do you think Joe Biden was playing 4D chess, learning from Dean Acheson, when he kept making these mistaken statements about Taiwan that his staff would later clarify by saying, “That’s not really what he’s saying,” while he maintained, “We’re going to protect Taiwan”?

Sergey Radchenko: That’s a whole different conversation regarding strategic ambiguity versus strategic certainty and which approach is better. The discussion has typically gone like this — If Biden remains unclear or maintains ambiguity about what the United States would do, that might potentially trigger Chinese intervention because they would underestimate American resolve to defend Taiwan. If, on the other hand, he stated very straightforwardly, “We will actually defend Taiwan,” then this would tempt the Chinese to test American resolve by invading Taiwan.

You can twist this argument in many ways. The bottom line is that you have internal debates and external communication between leaders. You have statements that can be misinterpreted and often are. This is how we slide into conflicts and wars — sometimes by misinterpreting what the other side wants us to do, or will do, or will not do, by underestimating the other side’s resolve.

The best example from our present situation is that Putin underestimated the resolve of Western powers, particularly the United States and Europeans, to help Ukraine. He underestimated the extent of their commitment, although that commitment has now started to fade away in some instances. Certainly in February 2022, he did not anticipate that level of commitment. Why? Because Crimea was a different story and nothing meaningful happened after Crimea, so he learned from that experience. It turns out the response this time was very different.

This parallels the early Cold War. There would be one type of strategy, and then the Soviets would do something completely reckless, like initiating the Korean War, which would reinforce the thinking in the United States that the Soviets must be confronted and pushed back against, even in a place like Korea. From Korea not mattering to anyone and nobody being able to find it on any map, suddenly it becomes the center of American attention, with people going there to die on behalf of the free world. Who would have thought? That’s a change of attitude, and it happens over weeks and months.

Jordan Schneider: This is one of the big lessons I took away from your book — the leaders are playing this weird game. Dean Acheson said it and probably believed it. Then when it happened, the politics changed, and everyone decided that it’s one thing to project it out, but when you’re faced with the reality, the news, the headlines, and the threat to prestige that all these people weigh so heavily, then perspectives change.

It’s also remarkable how you illustrate, by delving deeply into these leaders and their profiles, that they often don’t even know what the plan is. You see them winging it frequently. So even the idea that you can try to guess what your adversary is doing and play 3D chess when they themselves or even their lieutenant is thinking, “I don’t know what we’ll do next” — it’s wild to internalize that there are so many unknowns where one path could lead to the end of humanity.

Sergey Radchenko: Exactly. That’s the weird thing about Cold War history. Things could have taken terrible turns. They could have also turned out better than they ultimately did. We did have terrible crises during the Cold War, and at those particular moments, we should be grateful that leaders, despite all of their confusion and inability to understand what the other side was thinking — or maybe because of it — decided to de-escalate and pull themselves away from the brink. That is something that should be acknowledged.

Jon Sine: One of my lessons, and this might be the bridge from Stalin to Khrushchev, is how much these prestige battles mattered. It makes me think much of the Cold War was really just a sad, pointless endeavor of people jockeying over imaginary status points and where they sit in various social groupings. They’re getting pulled into games by third-party players while imagining it’s part of a broader struggle they have defined for themselves among their own group.

The Korean War is an interesting point because this is where America comes to push back, and what ends up happening is we now have a divided Korea. Fast-forward to today, and it seems like one of the most worthwhile interventions from the Cold War. But then you look to other cases, Vietnam being an example of an atrocious intervention. I’m wondering how you think about that — when America decides to intervene and when it becomes a really costly and pointlessly costly endeavor versus something that was worthwhile?

Sergey Radchenko: That’s a difficult question. It entails a moral judgment, and it’s difficult to make because, for example, the Korean War could have ended up as a nuclear war. Didn’t Douglas MacArthur want to use nuclear weapons, or at least threaten to do so? It could have happened but didn’t, and so if it did and we ended up having a nuclear catastrophe in Northeast Asia, would we be better off “dead rather than red“? It’s difficult to say.

In the end, it kind of worked out despite millions of casualties. It was a terrible, destructive war with so many deaths, but we can say in retrospect, “Look, the Americans stood up to defend freedom” — not quite democracy, frankly, because Syngman Rhee’s regime was not exactly democratic, but years and decades later it all worked out. Was it worth it? Probably.

By the same token, you might even say the same thing about Vietnam. One of the things I discovered while looking at Soviet documents — and the Soviets had remarkable access to the Vietnamese Politburo. Somehow they had spies there because in the Soviet archives you have speeches of Vietnamese Politburo members. So internal documents and all sorts of things from Vietnam that you wouldn’t expect. The Soviets clearly knew a lot more about what was going on.

You read these Vietnamese documents and recognize that actually Eisenhower’s domino theory was not wrong in the sense that that’s exactly what the Vietnamese wanted to do. They wanted to reunify their country and then promote revolutions. They obviously wanted to control revolutionary movements in Laos and Cambodia, but in the late 1960s they were discussing starting a communist revolution in Thailand, and I don’t know why that particular project didn’t work out for them.

When you read something like this, you think, “Wait a second, this is the domino theory. That’s what they’re talking about.” The Americans got involved and that was very tragic. You can turn to almost any historian of the Vietnam War and they’ll tell you, “This was a terrible mistake.” But you could also raise this counterfactual: “What if the Americans did not get involved and how would it have worked out for Southeast Asia?” Which is, by the way, very nice and prosperous now. All of that is to say that these are hard judgments. Very hard judgments.

Jon Sine: When reading your book, I was thinking that there is a fine line between describing these players battling for prestige and simply narrating that one side actually believed in a domino theory while the other side correctly presumed there was one. Where do you draw the line between objectively analyzing that each side is engaged in this game? Is there a place to step out and say, “You can analyze this correctly but you don’t necessarily need to engage in the game“? I guess it goes beyond the realm of a historian, but it was something I was trying to think about for lessons today. We’re not in any sort of domino theory situation now, but...

Sergey Radchenko: Exactly right. You might ask, “The domino theory — Vietnam was interested in taking over Southeast Asia. So what if they did? Would American global interests still be deeply undermined or fatally eroded somehow?” That is where you have to make the strategic judgment about what’s important and what is not important for America, and what’s worth paying the price for.

We all have to agree that having communists take over Southeast Asia was perhaps not the best idea for freedom and democracy, but is it worth sacrificing all those lives and basically getting America stuck in this quagmire far away from its shores? What would have happened if the Vietnamese were victorious earlier? Would that not have worked out differently?

Perhaps the Vietnamese had so many problems that their hopes to dominate Southeast Asia would have run aground anyway for internal reasons, also because the Chinese would be upset with that as well — as what ultimately happened in the 1970s with the Sino-Vietnamese conflict. Maybe the Americans should not have gotten involved precisely because this was far away from their core interest.

But what about Korea, then? You might come back to Korea and say, “That is also far away from America. Should the Americans have gotten involved, or should they have just followed Dean Acheson’s advice and stuck to their defensive perimeter?” It’s all about strategic judgments.

This matters today as well. Let’s say the Russians overrun Ukraine and establish a puppet regime or annex Ukraine. Would America’s global interests or core national interests be fatally undermined? You can make the argument both ways. You could say yes, because American credibility would be at stake — others will say, “America is a paper tiger, they cannot be trusted, they cannot even defend Ukraine, so how will they defend Taiwan?” and other things start falling apart.

Or you might say, “Actually, it doesn’t really matter.” Ukraine is far away from America’s core interests. America’s core adversary is China, that’s where it should be focused. Some people have been arguing exactly that, and so Ukraine is a distraction.

As a historian, what can I say about this, apart from noting this was always a problem and always part of the discussion? The best we can do really is see how historically things have worked out. Sometimes they have worked out well, and sometimes they have not. But there was always a price to pay. That is a very unsatisfying answer from a historian.

Jon Sine: It’s not unsatisfactory, but I would also draw in the domestic aspect as well. You go to pains to show how much it undermined the Soviets to spend so much of their effort and attention in these Third World competitions that ultimately were of little significance. We still feel the effects in the United States from our voyage into Vietnam in terms of undermining faith in the government. I think that’s another aspect that needs to be considered.

Sergey Radchenko: That’s right. The Soviets were ultimately the losers of the war in Vietnam. You’d think the Americans were the ones who lost out, but as I argue in the book, it’s actually the Soviets who lost out. They thought they won, but what they got was an ally they constantly had to pay for, and that was a gigantic drain on Soviet resources. They just kept paying for the Vietnamese, and the Vietnamese said, “Sorry, you have to forgive all those loans because we cannot do anything.” That actually contributed to Soviet over-extension. Who’s the real winner then? That’s a big question.

Vietnam's army and people are fighting well
A 1965 poster depicting Vietnamese soldiers firing Chinese-made Type 24 heavy machine guns. Caption reads, “Vietnam is fighting well!” Source.

Jordan Schneider: Let’s reel it back to the Stalin-Khrushchev transition. I love how you painted how the fight to kick out Beria ended up creating a fork in the road where Germany could have ended up in a different place, but everyone kicking out Beria ended up making that path toxic. What should people reflect on when understanding that power transition?

Sergey Radchenko: That is really an interesting period. We still don’t know all that much about it. We know enough from the memoirs and recollections of people like Nikita Khrushchev about what actually happened. We have the film “The Death of Stalin.” That’s where most students learn about Beria’s demise from.

The problem for that period — late Stalin up to maybe ’53, ’54 — is that we don’t have sufficient documentation. It’s not that it’s unavailable; I think it just doesn’t exist. Things like Politburo records of what they were discussing in the Politburo — the Presidium, as it was called — we don’t have that. There are some gaps in the record.

There is an interesting story about Germany in all of that. Around this time, the Soviets get the idea that East Germany should perhaps not try to build communism. It’s not working out well. They try to communicate that to the East German leaders, who are very hard-line, committed orthodox communists. These leaders keep pushing policies which result in a full-fledged uprising in Berlin in June 1953, which has to be suppressed with tanks, tragically.

The Soviets get this idea of maybe having a confederation or something similar. What they actually say is they could have a bourgeois Germany — a bourgeois Germany that is not aligned to the West necessarily, but is also not a communist country.

Now, the question is who’s pushing this? That becomes difficult to answer because you have people like Beria who seem to be in favor of this. Stalin’s hideous henchman comes across almost as a liberal, letting people out of the Gulag, saying to the Germans, “Look, you don’t need to build communism.” Beria is one. Malenkov seems to be supporting this as well, so it’s very difficult to say that it’s just Beria, because Malenkov also entertained similar ideas.

What I argue in the book is that Beria is arrested, and in the process where they try to implicate him or pin all kinds of terrible things on him, they say, “Well, he’s a Western spy, he’s a British spy. He was trying to undermine communism in Eastern Germany.” This is where internal politics and external foreign policy overlap. At this point, Malenkov could have said, “I also was in favor of the same policy,” but notably, he doesn’t. That’s how it works.

Beria is executed, and East Germany remains basically communist. Later, West Germany becomes admitted to NATO, and at that point, this idea of the neutral Germany that is bourgeois but not aligned to the West falls by the wayside. The Soviets then commit to maintaining the regime in the GDR, which is very costly. It’s a very costly proposition and difficult to do because people keep running away, which is one of the reasons that ultimately they have to build a wall.

What I’m trying to do with this book is show that you can actually have a variety of policies. At different times, like in ’53, they thought, “What about this approach?” It didn’t work out for different reasons, and then a different policy emerged.

Jon Sine: Your research basically finds the same thing as Joseph Torigian, that especially in those moments of transition of leaders, it’s really not about policy discussions at all that determine what’s going on. It’s internal struggling, what he calls a knife fight with weird rules.

Sergey Radchenko: That’s right. Joseph’s book is fantastic on this. In fact, we work from some of the same materials.

That’s what you get. Once you go into the real depth with these records, what you see is really just people above all. You see people positioning for power, for influence, backstabbing each other, and so on. Big questions of policy get drawn into this in weird ways, but sometimes in ways that you would not expect necessarily.

This is something interesting because I don’t know how you would explain that from a theoretical perspective. It doesn’t really work out. Acquaintance with archival materials like Joseph Terian and myself have been able to do is very useful.

Jordan Schneider: It’s like the theory is just schoolyard politics, which is kind of incredible. You have these moments in this book of big statesman-like times where Khrushchev feels really good when he meets with Eisenhower, and Brezhnev feels like he can do business with Nixon. But those are the times where your grasp on power is most secure.

When it’s not secure domestically or internationally, then it becomes about the unseemly parts of human nature that are evolutionarily the things you would see 10,000 years ago, not what you would necessarily expect humanity to be able to pull off and aspire to in the 20th century.

Sergey Radchenko: Jordan, I would actually argue against this. People don’t evolve so fast. In some ways, we do behave like monkeys in the forest because that’s evolutionary behavior, the struggle for “I’m the king of the hill” and “Everybody defer to me.”

If we see that with monkeys, why wouldn’t we expect people to have the same tendencies? Fundamentally, those psychological motivators of human behavior are really important. This is where over-reliance on some grand narratives about ideology or realpolitik starts falling apart. Yes, of course, all of those things matter. But in the end, we are monkeys in the forest.

You see these behaviors historically repeating themselves again and again for centuries. There must be something fundamentally human to this kind of behavior — this struggle for power, struggle for recognition, the desire for legitimacy, the desire for prestige. All of those things seem to be deeply rooted in human psyche.

Jon Sine: Once upon a time, I wrote an essay that I was going to use to get into my PhD program on reorienting how we think about international relations around these Evolutionary Psychology ideas of status and prestige. Needless to say, I’m a fan of your narrative.

But you can also root it in something else — philosophy, or at least great works of art, which you do. You quote from Dostoevsky in describing, I think it really applies most of all to Khrushchev above anybody else, “Am I a trembling creature, or do I have the right?” That just jumps out at you as something correct when you look at Khrushchev.

Maybe let’s fast-forward to the Stalin speech because in a way, he’s not so much a trembling creature when he’s up there thoroughly denouncing Stalin. This will be the thing that he’s known for. You look at some of the Marxist historians who say after reading “China Today,” this is where they say it all started to go wrong. Tell us about that.

Sergey Radchenko: Khrushchev decides to denounce Stalin — this is known as the Secret Speech — denounced Stalin for his various crimes, for repressions, for the way that he conducted the Second World War. The party is absolutely shocked. When this verdict is delivered, the party faithful are there, and they cannot believe it. Although for a time, things were kind of going in that direction anyway, so you had silent de-Stalinization already shortly after Stalin’s death. But in February 1956, it becomes official. It’s not publicly announced because it’s a Secret Speech, but it’s circulated inside the party.

It’s a very interesting question as to why Khrushchev does that. When you read about this, you almost feel sympathy for Khrushchev. You think, “Well, here’s a good guy.” But then you also have to remember that he did all kinds of other things, like ordering the Soviet invasion of Hungary, which serves as a counterbalance to this very positive view of Khrushchev.

Why did he do that? Was it for political reasons in order to consolidate his power, or did he really think that Stalin was a terrible person? I think probably both. It helped him consolidate his power, but there’s no reason to think that he was lying when he was saying that Stalin was a criminal.

Now, some people said, “Well, where were you? You were also participating in all this stuff. You benefited from this because your career overlapped with absolutely horrendous purges. You built your career on the bones and blood of your predecessors.” There’s something to that as well.

But when all is said and done, you look at this Secret Speech of February 1956, you cannot help but feel that this was the right thing to do. You cannot help but feel like this is one of the good moments for Khrushchev. There are some others, like the way he decided to back out from the Cuban Missile Crisis, where you feel, “He did the right thing.” That’s how I felt when I was looking at it.

But the Chinese were not very happy. Mao Zedong was saying, “Wait a second. You did not consult us.” This is part of the problem. Mao Zedong himself did not like Stalin, but he certainly did not like being surprised like that by Khrushchev. He said that, “Stalin was a great sword, and Khrushchev had discarded the sword, and now our enemies will seize it and try to kill us,” or as he would also say, “Stalin was a great rock, and Khrushchev lifted it only to drop it on his own feet.“

Mao predicted problems for the communist movement, and of course he was right because problems broke out in places like Poland that summer of ’56. Then ultimately, in the fall of 1956, we had the situation in Hungary that just went off the rails very quickly. All of that is connected to de-Stalinization.

In my book, I talk at length about it. I don’t focus so much on the domestic impact of de-Stalinization. There are other people who have written fantastic books about it, but what I look at is how it impacted countries like Poland, Hungary, North Korea, and then I also talk about China’s role in all of that.

This is where Mao Zedong, for once, feels like he’s the great strategist who’s determining the fate of global communism because he sends a delegation to Moscow led by Liu Shaoqi. Liu Shaoqi says in conversations with Khrushchev, “What you’re doing to Poland, trying to pressure Poland — this is called great power chauvinism. You should stop doing that. On the other hand, in Hungary, that is a kind of revolutionary rebellion, so you should basically suppress that.“

e know there is no Soviet intervention in Poland, and there is Soviet intervention in Hungary — by the way, for reasons that had nothing to do with Liu Shaoqi’s advice. It was for Khrushchev’s own reasons that simply overlapped with what Liu Shaoqi was saying. But the Chinese conclude that this is because of their advice. So they have finally got to a position where they’re determining the general line of the communist world movement, and that, I think, is consequential for the Sino-Soviet split, for all the things that happened afterwards.

Jordan Schneider: Two quotes, some of my favorite from the book, where you found in the Politburo that Mao talked about Khrushchev’s Stalin speech as if Khrushchev had broken the incantation of the golden hoop. Of course, the brocade that the Monkey King wore and now everything’s going to go crazy and Mao’s going to be able to be free. The superstitions of what can and can’t be done to challenge the Soviet Union are now blown wide open.

I also love the detour that you had of these incredible sources of the pro-Stalinist folks in Denmark and Norway. The story’s been told a lot in the US, but at some point, I think it was the Danish guy who was so pissed off — when he was getting the readout, he was saying, “Khrushchev, what were you doing? You should’ve arrested this man, Stalin. What a monster.“

The ability for some folks to see through the cognitive dissonance of it and say, “Okay, we should just give up. This is a completely bankrupt thing.” And then other folks in different situations are more or less bought into the system, or their incentives are different. It’s funny that the players with the least actual political power, the intellectuals in the West, were often the ones who clicked most quickly to the fact that a communist system able to bring about someone like Stalin is just a bankrupt endeavor on its face.

Sergey Radchenko: It took a while actually, and obviously ‘56 had a dramatic impact on the world communist movement. In many ways I would say 1956 was the beginning of the end for communism.

It was very clear. If you needed to use tanks to suppress a country or to promote your vision of economic development, you’ve got a problem. That is where you have a collapse of party memberships in communist parties across Western Europe. So many things can be linked back to 1956.

It was such an eye-opening moment for Western intellectuals, many of whom just could not believe that Stalin could be such a hideous figure. But that’s a consistent problem. I was struck reading Martin Sherwin and Kai Bird’s biography of Oppenheimer, American Prometheus, where they discuss how nuclear scientists working out in California were just absolutely blind to the brutalities of the Soviet system. You think, “How can smart people who are obviously geniuses in physics be so uncritical as to think that there’s this worker’s paradise and everything is going wonderfully well?“

For Oppenheimer himself, 1939 seems to have been the turning point — the Soviet-Nazi pact. After that, he finally started thinking that perhaps this Stalinist paradise was not so wonderful after all. But Soviet intellectuals continued to believe until 1956.

In 1956 you have the double shock. First, Khrushchev himself says, “Look, Stalin is not the father of humankind, not this great genius or demigod that we held him for, but actually a brutal criminal who repressed innocent people and was just a terrible person.” Then you’ve got the Soviet intervention in Hungary which nails the final nail in the coffin of the Soviet project for many Western intellectuals. That’s why you have this chaos in Western European Communist Parties from CPUSA to the Norwegians, the Danes, Great Britain, and so on.

Jon Sine: You just said in 1956 you could see the beginning of the end of communism. When Deng Xiaoping was writing at the Sixth Party Plenum in 1981, rewriting the history of what happened in China, he’s very careful not to reject Mao in total, seeming to draw a lesson from 1956.

This is a very prominent lesson today among certain Marxist historians who look back at the Soviet Union and how it fell. 1956 is a crucial year, but you didn’t just root it in the Secret Speech. You pointed to tanks going into Hungary as a crucial thing. Do you agree with their analysis? Was there a way to save Stalin’s reputation? Did the Chinese keep a 70/30 Stalin type of interpretation? What are your thoughts on that?

Sergey Radchenko: That’s obviously how Mao would have liked to see things, which is why he proposed to discuss Stalin in those terms. That Stalin committed some mistakes — mainly not believing in the revolution in China and mistreating Mao. But generally speaking, he was a great Marxist-Leninist. That would have been a better approach.

Stalin was a horrible person, and saying this frankly and openly was a better thing to do. In fact, the problem with the Soviet Union and with Khrushchev’s de-Stalinization was that they did not go deep enough. They deposed Stalin but the system remained, and I would argue the system still remains. Is that a good thing? I don’t think so.

There needs to be a different approach to people like Stalin in places like Russia and to people like Mao Zedong in places like China. But of course the Chinese will say, “We are careful. We don’t want to undermine the legitimacy of the communist state. That’s why we still print Mao Zedong on our money and that’s why his portrait hangs over Tiananmen Square, because he is the founder of the Chinese state. If you say that he was a horrible criminal responsible for the deaths of 35 million people, well... that might delegitimize the Communist Party.” Are they wrong? I think they’re probably right — of course it will delegitimize the Communist Party.

Jon Sine: Nothing delegitimized the Soviet Communist Party more though than their failing economy. You mentioned not de-Stalinizing. There was the Kosygin reform which could have radically restructured, gotten rid of the ministerial planning apparatus. I feel like if they were going to go down that route, they needed to get away from the Stalinist planning system at some point. It could take you to maybe the early modernization, but they never seemed to get out from under it.

Sergey Radchenko: Khrushchev’s reforms brought about a lot of chaos in the USSR. Constant restructuring, compressing ministries, taking out whole parts of the apparatus. I say in the book that Khrushchev was like this guy who felt the communist project was a really great project and you just have to tinker with it here and there. Take out a pin here, put something there and then it would work just fine. The problem was that it fundamentally wasn’t working, so that realization took some more time to creep in. That’s already after Khrushchev.

We can see the key point, as you rightly pointed out, is that the Soviet project was not delivering for the Soviet people. It was not delivering enough because the Soviets set themselves up not just with the claim that they would improve the standard of living — because actually the standard of living was improving — but that they would outstrip the United States.

They claimed they would be better. We have the kitchen debate, that famous encounter between Nikita Khrushchev and Richard Nixon, of which we have a video recording. I was so happy to find the Soviet version of it in the Russian archives, which was hilarious to see what was not captured on tape. They go around and Nixon shows this and that and says, “Well, look, this is an American house.” And Khrushchev responds, “No, this is made of wood. This is rubbish. We now make stuff of plastic. This is so much better.“

You have this competition with the United States. The idea was that it would deliver for the Soviet people and it simply wasn’t doing that. In some areas, there were breakthroughs, but by and large, the Soviets had meat riots. Think of the 1962 riot in Novocherkassk where the KGB had to be deployed — the military police — to break up a protest with casualties. In 1963, the Soviets were spending their gold reserves to import grain for the first time. That is already where you’ve got some serious warning signals coming.

That’s why they launched the Kosygin reforms. But then that kind of goes off the rails. At the same time as all these problems in the economy multiply, they also discover oil and gas in much larger quantities in western Siberia and they think, “Okay, that will save us.” To a certain extent it actually helped them for another decade or so but ultimately, of course, the whole thing fell apart.

Jordan Schneider: I want to back up because I don’t want to move too quickly past the second half of the 1950s. You actually have this incredible moment of optimism with Khrushchev, where you have Sputnik, of course, and then this techno-utopian vision that Khrushchev has. He believed that in the future, the Soviets would have a bright future, a fairytale-like socialist abundance with people having so much to eat that they would, quote, “be careful not to overwork their stomachs” as robots did all the work. People would just hang out, only working for an hour or two per day.

A 1953 poster. The caption says, “Study the Soviet Union’s advanced economy to build up our nation.” Source.hhhh

Share

You make this fascinating connection between what Sputnik does and how it changes Khrushchev’s psychology, leading to some of the most dangerous moments in human history. Could you explain that connection? Khrushchev has this big technological breakthrough, decides he’s going to catch up and surpass the US in 15 years, and what does that lead to in terms of the risk of war in various hotspots around the world?

Sergey Radchenko: As you say, Jordan, these are interrelated processes. The Soviet defense program, the missile program, and the Sputnik program were obviously intricately related. In fact, Sputnik was launched on an ICBM.

What was happening in terms of defense priorities? The Soviets got their bomb in 1949, still under Stalin. There was a modest buildup of nuclear bombs in 1950, but then they really put a lot of effort into the development of missile technology. By around 1955, there were some serious breakthroughs.

First, you had a proper Soviet thermonuclear test — not the one in ’53, which wasn’t the later design — but the proper thermonuclear test in 1955. That made Soviet nuclear weapons much more powerful, giving Khrushchev this sense of great power: “Look, we’ve got these massive weapons that can destroy whole cities.” That was an empowering feeling.

Then you had breakthroughs in missile technology. By 1956-57, missiles increased their range. In 1957, the Soviets conducted the first test of an ICBM — intercontinental ballistic missile. Khrushchev realized this basically meant the Soviet Union was invincible as a power. It could reach the United States. It could destroy the United States.

For the first time, he developed this feeling — later shared by all Soviet leaders and even today by Russian leaders — that Russia was a mighty power precisely because it could destroy the world. That’s where its claim to greatness came from. It could destroy the world, and nobody could mess with it. For the first time, Russia could be safe because nobody would dare to invade a nuclear power.

Khrushchev understood that, which is why he began massive cuts in the armed forces and shut down projects like the battleship program that was Stalin’s prestige project. Why? Because he believed in nuclear weapons. He felt they gave the Soviets might and a new voice in international politics.

Although he didn’t plan to use nuclear weapons, he used them for blackmail and discovered the usefulness of that already in 1956. During the Suez Crisis, Khrushchev threatened to destroy Great Britain and France. Those powers ultimately backed out from Suez for reasons that had little to do with Khrushchev’s threats — mainly American pressure on the British. But that’s not what Khrushchev thought. He believed, “I was so successful in doing that. Now, let’s play this card again and again.“

That, Jordan, got him into dangerous situations, ultimately leading to the Cuban Missile Crisis, because he felt that the Americans or anybody else would have to retreat when faced with Soviet might. It gave him this trump card to play.

On the Sputnik side, you had the launch on October 4th, 1957. That’s why we talk about the “Sputnik moment” — a term we now use to describe breakthroughs by America’s adversaries. The original Sputnik moment was a great breakthrough which symbolically showed that the Soviet Union could outcompete the United States in a key area of science and technology.

Unfortunately for the Soviets, they weren’t able to capitalize on this. Though they could capitalize on Sputnik itself, they couldn’t provide sufficient investments in scientific and technological terms into their own economy where their economy would compete with the United States. There were structural problems with the economy as well, and it simply wasn’t competitive.

The Sputnik moment faded away fairly quickly, but what remained was the Soviet nuclear threat, which was real and which became ever more serious as time proceeded.

Jon Sine: The Cuban missile crisis remains one of the most hotly debated historical occurrences. From the US perspective, it’s incredibly important. The Sputnik moment you mentioned factors into that. Another element is the menage a trios between China, the US, and Soviet Union — a constellation of powers competing for greatness.

The Soviets seemed to have a trickier game to play because they were balancing their great power aspirations between themselves and the US while also competing for leadership over the communist revolutionary world with China. They were attempting this double dance, which came to a very dangerous head in the Cuban missile crisis, with that process playing into Khrushchev’s decision to try to win over Fidel Castro.

How does bringing in the Chinese and the competition for prestige dynamic change how we think about the Cuban Missile Crisis?

Sergey Radchenko: That’s a great question. The Cuban Missile Crisis is a very overstudied episode of the Cold War. What I contribute in my analysis is precisely this Chinese angle, which is fascinating. We haven’t really been talking enough about that.

If you think about the Cuban Missile Crisis, what questions are we still trying to answer? The first question is why Khrushchev sent missiles to Cuba in the first place. There are different competing theories about it.

The prevailing American theory for years was that it was strategically necessary for the Soviet Union. Khrushchev realized that his ICBMs — for all the Sputniks of the world — weren’t particularly good. They weren’t accurate; they were problematic. So he had to put intermediate-range ballistic missiles in Cuba to target America more accurately and reliably.

That has been the traditional explanation for Khrushchev’s decision. Looking at the Soviet and Russian records, there isn’t a single moment where Khrushchev actually raises this as an explanation — not once. I didn’t find a single piece of evidence to support that. It’s more about political scientists trying to figure out why Khrushchev would do this than Khrushchev actually explaining his reasoning.

In the 1990s, Sergey Mikoyan, who was the son of Anastas Mikoyan whom I previously mentioned, proposed a different theory. He suggested it wasn’t about addressing strategic problems at all. It was about saving Cuba because of the Bay of Pigs and the fear that the Americans would take over Cuba — a realistic fear, let’s be honest. Khrushchev wanted to save Cuba.

A 1960 Soviet poster. Caption reads, “The Cuban people do not break!” Source.

Is there evidence for this? If you look at the Russian records, including the Presidium discussions, you actually find evidence for it. Khrushchev says on at least two occasions, “We just wanted to save Cuba from American invasion.” This is very interesting because it supports this particular theory.

In my analysis, I ask why he was so obsessed with Cuba. What was that about? I argue that it was because he was in competition with China, where the Chinese were saying, “The Soviets have betrayed revolution. They’re weak. They’re not standing up to American imperialists.” Under those circumstances, losing Cuba — which was basically a socialist country run by a real revolutionary, a wannabe communist — would be unacceptable to Khrushchev.

He was really worried about the Chinese angle. There’s substantial evidence for this because throughout and after the crisis, he was very sensitive to Chinese criticism. He kept telling the Cubans, “Don’t believe the Chinese. We don’t want to sell you out. We are actually the only ones helping you,” and “Look what the Chinese haven’t done for you anyway.“

The Chinese had made their own inroads into Cuba because they were talking to Che Guevara, who had a very close relationship with the Chinese ambassador. This presented a big challenge for Khrushchev.

Bringing us back to the angle about prestige and the Soviet desire for recognition and greatness, there’s another interesting piece of evidence that connects to my theory. Khrushchev wanted to be treated as an equal to the United States. When he was making the decision to send missiles, it was in the context of a discussion about American missiles in Turkey.

From his position, if the Americans had nuclear Jupiter missiles located in Turkey, why couldn’t the Soviets have missiles in Cuba? That wasn’t fair. Khrushchev commented, “We will effectively give the Americans a little of their own medicine.“

If you psychoanalyze this phrase, what does it mean? Does it mean he was really concerned about strategic problems or reliably hitting Washington? He never wanted to use nuclear weapons to begin with. No, it was more about equality — why were the Americans allowed to have missiles near Soviet borders, but the Soviets weren’t allowed to do the same? This brings in the question of equality, status, and greatness: “We are on par with the United States. We can destroy them, therefore they should not expect special treatment.“

That’s how I describe the opening phase of the crisis. The relevant chapter also discusses how the crisis itself unfolded. I was fortunate to have access to remarkable materials that add to our understanding of how Khrushchev ultimately decided to back out from the situation.

The key piece of evidence is that Khrushchev really thought Castro was going off the rails at one point, especially after Castro proposed to nuke the United States in a first strike. Castro later denied this, saying he never meant anything like that, but that’s how Khrushchev understood Castro at the time. Khrushchev was shocked, thinking, “What is he talking about? This guy’s crazy.” At this point, Khrushchev became frightened.

One of the things you find in the Russian archives is how early Khrushchev decided to back out. Kennedy gave the quarantine speech on October 22nd, 1962, and then Khrushchev dictated a letter to Kennedy on October 25th, already effectively backing out of the crisis. It only really lasted for three days.

Analyzing Khrushchev’s language and concerns is fascinating. All of that material is now available in Moscow for researchers — though I don’t recommend going there.

Jordan Schneider: The dynamic you see in the Cuban Missile Crisis is one that plays out over many crises, where we have lots of influential actors say, “Let’s just send some nukes. It’s the path of least resistance. It’ll solve our problem.” We mentioned this earlier in the Korean War context, and it happened here as well.

But you do see this other very human reaction where the person who will actually make the decision often seems to be the only one truly weighing what the second-order consequences mean, beyond just the narrow military, theater-level advantage you might get by being the first to pull the trigger.

Sergey Radchenko: That’s precisely the point. When you have the responsibility, you must make that final decision. The Soviets had extensive plans for using nuclear weapons in Europe to wage nuclear war. They practiced for it and carried out military exercises with nuclear weapons.

The military planned for it, but they didn’t decide on this matter. The leadership had to make that decision — Khrushchev ultimately, because the responsibility was in his hands. When he considered what he could potentially be authorizing, he was deeply concerned.

This is where personal psychology becomes so important. I discuss a fascinating episode in the book about how Khrushchev contextualized his decision on the Berlin Crisis by recalling his experiences from the Second World War and the story of Nikolai Voshchugov.

Back in 1941, after the Germans invaded the USSR, Voshchugov, one of the commanders on the Soviet side, came to Khrushchev and said, “I’ve lost my tank army.” Khrushchev asked, “What do you propose to do?” Voshchugov pulled out his handgun and shot himself in front of Khrushchev.

What’s interesting is that Khrushchev recounted this story when discussing his decision — making during the Berlin Crisis. Why is that important? Because Khrushchev realized that human rationality had limits. Why would this man take his own life like that? Was it rational? No. Similarly, starting a nuclear war isn’t rational, but people might still do it. This realization added to his reservations.

Despite all the nuclear plans, exercises, training, and available weapons, Khrushchev did not plan to fight a nuclear war. He reasonably concluded that he didn’t want to engage in nuclear warfare.

The same applies to the American side. Consider President Eisenhower’s approach to the second Taiwan Strait Crisis in 1958, when the Joint Chiefs of Staff presented the nuclear option, saying, “Use these weapons.” Eisenhower refused to consider it because he was the decider, as George Bush would later say. They were the deciders who had to make these choices, and the weight of responsibility was immense.

Jon Sine: What’s interesting is that Mao at times seemed unconcerned about nuclear weapons. I believe he once said — perhaps to Khrushchev — “If they used a nuclear weapon on China, they could take out half the population but we’d still have 300 million Chinese,” or something to that effect.

This brings me to the Sino-Soviet split, which is perhaps the most important aspect of the Cold War. There are many contenders for that title, but it really comes through in your book and is one of the most interesting aspects. I’d like to ask about its origins. You mentioned potential ideological debates but don’t find them convincing. The split started under Khrushchev but reached its most critical point — a literal war — under Brezhnev. Could you talk us through that?

Sergey Radchenko: This is another topic historians have debated for years, myself included. I wrote another book on the subject called Two Suns in the Heavens where I presented a particular view on why the Sino-Soviet split happened. We also have scholars like Lorenz Lüthi from McGill University, a good friend of mine, who wrote a different book arguing for the importance of ideology.

This debate goes back to the 1960s. In line with my general skepticism about ideology, I argue that ideology wasn’t really what mattered most, which is counterintuitive. There was extensive propaganda and ideological rhetoric on both sides — the Chinese accusing the Soviets of revisionism (revising Marxist-Leninism), while the Soviets called the Chinese dogmatic.

So much material was produced in various proclamations, statements, and letters that you might think ideology was obviously important, but I don’t believe that’s what truly drove the relationship or caused the split. Fundamentally, the Chinese didn’t want to be underdogs in this alliance, particularly Mao Zedong. He believed he deserved a better position and wanted to lead the alliance forward in ways he felt were correct. He didn’t want to defer to people like Khrushchev, whom he didn’t consider particularly bright or insightful.

It was partly a conflict of personalities and partly a conflict over leadership. Khrushchev gave a remarkable description of the reasons for the split in a conversation with Castro, who asked him when visiting Moscow in spring 1963, “What’s going on between you and the Chinese?” Khrushchev replied, “I also don’t know what it’s about. They say they’re against world war, we’re against world war. They say they’re for revolution, we’re for revolution, so none of that makes sense.” Then he advanced his theory: “Actually, the Chinese want to play the first fiddle.“

Then he completely goes off the rails in one of the most fascinating snippets of Khrushchev I ever found. He launches into a theory about how even in a circle of friends, some are naturally smarter, with various degrees of intelligence, different colors of hair, and so on — and it becomes quite explicitly racist.

Basically, Khrushchev felt like, “Look, who are the Chinese? We are the ones who had the communist revolution. We are the ones who won the war against Nazi Germany and launched Sputnik into space. So why are the Chinese trying to claim leadership from us? It’s unreasonable. We are the natural leaders.”

From Khrushchev into the Brezhnev era, the Soviets were determined to maintain their primacy in this relationship. They also hoped to bring the Chinese back to their “proper place.” They felt the Chinese had erred and would eventually recognize their mistakes, repent, and return under Soviet leadership.

That’s one reason this conflict continued for so long. Only under Gorbachev did things change, when he finally said, “We don’t want you to be younger brothers. We’re not interested in that.” Then they rebuilt the relationship on a more equal basis.

Jon Sine: One of the key points of criticism was when Khrushchev realigned toward saying, “We’re going to have peaceful competition or peaceful coexistence,” while the Chinese were saying, “No, we should heighten tensions, especially when it comes to the US.” This becomes even more ironic 10 years later with the Mao-Nixon rapprochement.

Sergey Radchenko: That’s precisely the point. That’s why I’m somewhat skeptical of ideological explanations. The Chinese felt the Soviets weren’t revolutionary enough, and then what did they do? They invited Nixon to Beijing, and Mao said, “I like rightists.” That’s an actual quote from his conversation with Nixon. What does that make of the various ideological disagreements they had? That’s a good question.

Stay tuned for part 2!

跨越欧洲骑行记之八—梅莱拉的鸡汤

23 March 2025 at 08:12

罗马尼亚境内骑行的最后一夜,我在吉尔久镇度过,从那里骑过多瑙河上的友谊大桥,就是保加利亚了。吉尔久的客栈在一处有葡萄藤和果园的院落中。我推车进了大门,一位老者坐在院墙下的椅子上打招呼,让一位年轻人把我带到房间。我把车子倚在墙角,门厅内一位清洁工打扮的老妇人在擦地板,收拾床单和杂物。

收拾停当,我出门找饭吃,老者仍然坐在院墙下的桌子旁,另一把椅子上坐着一位看上去30几岁的女子。她能讲英语,说自己名叫梅莱拉,那位老者是他父亲,名叫马瑞安,那位清理房间的妇人是她母亲,名叫马瑞安娜。她告诉我餐馆的大体方位。我吃完晚饭回客栈时,一位消瘦的中年人正在院子里干活,手持电磨具打磨褪色的绿色篱笆墙。他没戴任何防护,电磨具嗡嗡作响,他脸上和身上落满一层绿色油漆粉末,夕阳下像电影中的绿人史瑞克,只是比史瑞克瘦小得多。

莱拉正在跟他父亲喝酒。见我回来,她说:“时间还早,跟我们一块儿喝点吧。”我说:“我街上买了拉普提(罗马尼亚语“牛奶”)”。梅莱拉大声说:“不要拉普提,要卡尔巴顿!”卡尔巴顿是桌子上摆的酒瓶子上印的名字。梅莱拉把她坐的椅子让给我,自己去搬来一把,坐在我和他父亲对面。她给我倒了一玻璃杯卡尔巴顿,用小刀切下一片橙子,放在酒中,然后给自己也满上。

Read more

Rickover’s Lessons

20 March 2025 at 18:36

Charles Yang is the executive director for the Center for Industrial Strategy, a bipartisan think tank focused on industrial policy. Previously, he served as an AI and Supply Chain Policy Advisor at the Department of Energy and was an ML Engineer at an AI hardware startup in San Francisco. Today, he’s here to present some excerpts from his research into how Admiral Hyman Rickover built the nuclear navy.


Strategic competition demands more than technological innovation — it requires building industrial power. The U.S. is realizing the damage done by decades of underinvestment in the nation’s industrial base, which now jeopardizes its ability to compete on the global stage. Today, the production capacity of Chinese shipyards is over 200 times that of US shipyards, and China has used its chokehold on critical mineral processing as leverage to retaliate against US sanctions.

A new bipartisan consensus is emerging around the need for industrial policy — from the passage of the CHIPS and Science Act, to the recent bipartisan introduction of the SHIPS for America Act and the Critical Minerals for the Future Act.

As Congress steps into this more active role, policymakers should learn from the successes of our past. Nearly 75 years ago, Admiral Hyman G. Rickover, “Father of the Nuclear Navy”, pioneered a bold program to develop and operationalize nuclear power in the Navy. Under his leadership, the U.S. government harnessed the power of the atom, building the world’s first nuclear-powered submarine and the world’s largest fleet of nuclear reactors for civilian power.

Lessons From the Past

Rickover spent his entire career in the Navy and is still the longest-serving naval officer in US history. He spent the first 20 years of his career as an electrical engineer, where he honed a strong technical foundation and unique management style. In 1946, he was assigned a 1-year tour of duty at the Oak Ridge site of the Manhattan Project. Rickover immediately recognized the transformative potential of nuclear technology — he spent the rest of his career building the “Nuclear Navy,” which ensured US strategic dominance of the high seas for the rest of the 20th century.

Within the span of 10 years, Rickover created an entire office dedicated to nuclear propulsion, and successfully launched the world’s first nuclear-powered submarine without cost overruns. He conclusively demonstrated the strategic importance of nuclear propulsion in a timeframe no one thought possible and helped the US beat the Soviets to nuclear propulsion for submarines by 3 years. His institutional legacy is the US Navy’s safe construction and operation of nuclear reactors.

As the US gears up for another strategic competition, Rickover’s story can offer helpful lessons for aspiring technocrats. Oftentimes, industrial policy is framed in terms of legislation, but Rickover demonstrates that industrial policy is as much about policy as it is about strong leadership.

USS Enterprise, the world’s first nuclear-powered aircraft carrier. Source.

Talent, Training, and Management

Rickover spent an inordinate amount of time focused on interviewing personnel — he made the final hiring decision for every naval officer who applied to serve on a nuclear submarine until he retired. And he was an unorthodox interviewer, screening for high agency individuals who could think on their feet — literally! To test their composure, Rickover famously made candidates sit in a chair with the front two legs shortened as he loomed over them during questioning.

One interview account:

For one interviewee who said he liked hiking, Rickover asked him if he had ever hiked the nearby “Goat Mountain”. When he said he had not, Rickover told him to bring him proof he had climbed it by tomorrow morning and he would be hired. It turns out that Goat Mountain was the peak of a structure for mountain goats in a zoo. He went to the zoo, asked a tourist to take his picture, jumped into the enclosure, and climbed to the top. He’s hired the next day!

But it didn’t end at the interview process — Rickover believed in continued technical training for his staff and in building out a talented workforce base for this new technology:

While Rickover worked to staff up quickly in the short term, he also set out to build a deep bench and a long-term pipeline of talent. He required each officer and engineer he hired to submit a self-study plan demonstrating mastery of advanced texts in metallurgy, physics, and chemistry, along with field trips to AEC facilities, totaling 854 hours of study or 16 hours per week. He also worked with MIT to develop a survey course on nuclear physics and a master's degree in nuclear engineering, with a curriculum drawn up and agreed to by Rickover, starting in June of 1949. Rickover also worked with Oak Ridge National Lab to develop a 1-year curriculum in nuclear science and technology, a program christened “Oak Ridge School of Reactor Technology (ORSORT) with the first cohort starting in March 1950. Westinghouse, GE, utilities, naval and private shipyards, and Naval Reactors all sent students to ORSORT and the program started turning out ~100 graduates a year, providing another training center to develop a nuclear industry.1 Finally, Rickover had his engineers provide training lectures to a variety of audiences, ranging from senior officials in BuShips to junior technicians, as well as to explain shipboard problems and applications to scientists at Argonne, Oak Ridge, and Westinghouse/GE.

He was also known for his unique style of management. Not only did he interview every naval officer in his office, he also maintained direct lines of communication with every nuclear sub commander and project officer on-site with contractors, giving him early awareness of every issue. The demanding oversight he extended over his technical staff under his command pushed them to have greater awareness of their own direct reports:

Rickover was also an intensely demanding and scrutinizing manager. As most writing then was done on carbon copy paper, every night Rickover would collect the “pinks” of every piece of writing from his various teams i.e. the carbon copied half, and read over them at home, including drafts. When his officers protested as to how they should be expected to keep track of everything in their purview, including drafts reports from staff below them, Rickovers responded “It’s up to you to see that I don’t know more about what’s going on in your shop than you do”. By enforcing tight lines of supervision over his officers, Rickover ensured that he maintained full visibility into each team, including the project facilities at Knolls, Bettis, and the shipyards, allowing him to catch problems early on. It also enforced a culture of direct accountability and oversight across the organization.

Rickover’s focus on hiring, training, and close project management represented his philosophical approach to how to build complex systems managed by humans.

Near the end of his career, Rickover testified to Congress after the Three Mile Island Reactor accident. He spent the vast majority of his testimony talking not about regulatory reform, but about the lack of training and inadequate culture of responsibility among the operators.

“Human experience shows that people, not organizations or management systems, get things done. For this reason, subordinates must be given authority and responsibility early in their careers…

Complex jobs cannot be accomplished effectively with transients. A manager must make the work challenging and rewarding so that his people will remain with the organization for many years. This allows it to benefit fully from their knowledge, experience, and corporate memory.”

~ Hyman Rickover, 1982

Industrial State Capacity

Rickover’s scrutinizing style of management extended to the private companies he worked with. He pioneered the practice of project officers, who lived on-site at the projects and who would report directly to him any delays or unforeseen issues, so that Rickover could escalate immediately and ensure the project remained on track.

Government contracting was, and still is today, a largely passive and administrative activity. While Rickover acknowledged that the government was the “customer” and the contractor was the one responsible for delivering, Rickover’s unique approach to program management was exercising tight oversight over the contractors. Rickover hired technical experts into his office and then sent them out as project officers to oversee the various contractor sites. There, the project officer was expected to be the active representative of the Naval Reactors Office, reporting directly to Rickover any issues with contractors and ensuring the contractor was on track to deliver the product as expected. In every sense, Rickover’s project officer was to be his eyes and ears on the ground. Rickover took great pains to ensure there was no customer capture, telling one of his project officers, “Don’t go to dinner with them. Your wives must not get friendly with their wives. You’re not even to let your dogs get friendly with their dogs…when you do that, you become one of them…you don’t represent me anymore”.

Rickover’s success in scaling industrial technology was demonstrated early on with Zirconium production. In 1949, the world had only produced a shoebox worth of purified Zirconium, but the material showed promise as a fuel cladding material due to its durability under high temperatures without blocking the emitted neutrons needed to enable fission reactions. AEC opened up a simple contract for private companies to bid to produce Zirconium, but none of the companies were able to scale up production. Rickover took over production a year later, applied his practice of close project management with the (now defunct) Bureau of Mines, and only then passed it off to industry:

But by 1949, when Rickover was looking to scale up promising fuel cladding material production, the AEC had already decided to run contracts through another AEC division. Unable to exert the centralized control over the contractors, the AEC manufacturers were slow to scale up a high-quality production process. In 1950, after a year of delay, Rickover finally received permission to have the Westinghouse Bettis site directly manufacture Zirconium metal and worked with the Bureau of Mines (BuMines) to purify the Zirconium. Under Rickover’s scrutiny, Bettis scaled a novel purification process to thousands of tons of production capacity. Rickover opened up contract bids for Zirconium only after having derisked this novel technology. When the Secretary of the Navy later asked Westinghouse how they managed to scale up this process, the response he got was “Rickover made us do it”.

“The man in charge must concern himself with details. If he does not consider them important, neither will his subordinates.”

~ Hyman Rickover, 1982

Bureaucratic Innovation

Building big things requires lots of people. Rickover was not only an exceptional manager of people and deeply technical, but his 20-year naval career before Oak Ridge taught him how to wrangle government bureaucracy — and discern which rules mattered and which didn’t. For example, Rickover was interviewing an officer who thought the monthly reports on the gasoline usage of his base’s motorboats were pointless and wasteful. Rickover told him to simply remove the tickler file that tracked the reports from the boss’s secretary file and to send over a note the next day alerting Rickover that the task had been completed. The interviewee did and was hired.

Subscribe now

Rickover’s bureaucratic skill is exemplified by his success in rallying the Navy behind the nuclear-powered submarine. He believed this was a feasible, near-term project, despite widely-held convictions to the contrary — including those of the Atomic Energy Commission (AEC). Even Robert Oppenheimer (who served as one of the first AEC commissioners) doubted nuclear propulsion early on.

In light of initial resistance from the civilian AEC, Rickover formulated a unique bureaucratic innovation to position himself within two chains of command — one within the Navy and the other within the AEC.

Rickover was also able to realize his bureaucratic innovation to occupy a spot on the org chart both at AEC and in the Navy BuShips, something he first formulated while at Oak Ridge. This way, if the AEC refused something, he could respond that “this is a priority for the Navy” and vice versa. Similar to how the Manhattan Project reduced risk by pursuing parallel technological approaches, Rickover would reduce his bureaucratic risk by pursuing parallel chains of command. This unique structure lives on to this day, with Naval Reactors shared between the semi-autonomous National Nuclear Safety Administration (NNSA) in the Department of Energy (DOE) and the Navy.

“The status quo has no absolute sanctity under our form of government. It must constantly justify itself to the people in whom is vested ultimate sovereignty over this nation”

~ Hyman Rickover


Rickover firmly believed that the right team and the right culture could build incredible industrial technologies at scale, even within the government. While discourse in Washington DC often focuses on regulations or money, Rickover’s life brings a uniquely human-centered view of industrial policy: one that recognizes the importance of state capacity, technical personnel, and most importantly, public leaders with the vision and drive to build technology.

You can read the full story of Rickover and how he built the world’s first nuclear-powered submarine on Charles’s substack.

Moonshot AI's AGI Vision

19 March 2025 at 18:49

Founded in 2023, Moonshot AI is one of China’s four new “AI tigers” that’s attracted massive valuations and big-name investors including Alibaba and Tencent. The firm is known for its chatbot, Kimi, whose most recent release highlights improved math, coding, and multimodal reasoning capabilities.

The following piece is a translation of an interview with one of Moonshot’s founders, Yang Zhiling. With a bachelor’s degree from Tsinghua University and a PhD from Carnegie Mellon University, Yang boasts an impressive resume that includes time working at Google Brain and Meta AI. He was also a technical contributor to some of China’s earliest large models, including Pangu 盘古 and Wudao 悟道. His research publications are numerous, and he is the first author of two highly-cited papers in the natural language processing (NLP) field: Transformer-XL, which proposed a method that extends the context length of transformer models, and XLNet, which introduced a way for models to better understand complex data relationships.

This interview was published on the official account of Overseas Unicorn on February 21, 2024. In it, Yang outlines his vision of Moonshot AI as a combination of “OpenAI’s technology idealism” and “ByteDance’s business philosophy.” He covers a couple of key points:

  • Moonshot’s goals and how it plans to compete with OpenAI,

  • The pursuit of AGI,

  • Data challenges and the potential of multimodality and synthetic data,

  • Personalized AI models,

  • Yang’s approach to leadership and his vision for a global tech future.

Yang Zhilin of Moonshot AI: How Can a Newly Founded AGI Company Surpass OpenAI?

Original article, Archive link.

Interviewers | Tianyi, Penny, and Guangmi. Editor | Tianyi. Typesetter | Scout.

01. AGI: AI is essentially a bunch of scaling laws

Overseas Unicorn: We compare training LLMs to landing on the moon, and the name “Moonshot AI” [literally “dark side of the moon”] is also related to moon landing. How do you view LLM training by startup companies? Under conditions of limited GPU and computing resources, is it still possible to achieve a “moon landing”?

Yang Zhilin: “Moon landing” has several different production factors. Computing power is certainly a core one, but there are others as well.

You need an architecture that simultaneously satisfies both scalability and generality — but today, many architectures actually no longer meet these two conditions. Transformers satisfy these two conditions in the known token space, but when expanded to a more general scenario, they don’t quite work. Data is also a production factor, including the digitization of the entire world and data from users.

So among many core production factors, by changing other production factors, you can make computing power utilization more efficient.

At the same time, regarding “moon landing,” computing power will definitely need to continue growing. Today, the best models we can see are at a scale of 10^25 to 10^26 FLOPs. This order of magnitude will certainly continue to grow, so I believe computing power is a necessary condition. This is because machine learning and AI have been researched for 70 to 80 years, and the only thing that actually works is the scaling law, which is the expansion of these various production factors.

We are actually quite confident that, within a one-year time window, we will be able to achieve a model at the scale of 10^26 FLOPs, and that resources, ultimately, will be reasonably allocated.

Overseas Unicorn: For OpenAI to train their next-generation model, we estimate they have at least 100,000 H100 GPUs, with single clusters reaching 30,000 GPUs. OpenAI is clearly pursuing the “moon landing,” with the possible shortcoming being that they don’t focus as much on user and customer experience. Where will Moonshot AI’s path differ from OpenAI’s? What can Moonshot AI do that OpenAI won’t do?

Yang Zhilin: A key point in the short term is that everyone’s tech vision is not exactly the same. Many areas are not OpenAI’s core competitive strengths (for example, image generation); DALL·E 3 is at least one generation behind Midjourney. GPT’s long-context capabilities are also not state-of-the-art. The lossless long-context technology we recently developed performs better than OpenAI in many specific scenarios because it uses lossless compression technology. You can use it to read a very long article, and it can effectively reproduce specific details and make inferences about the content. Users will discover many scenarios themselves, such as chucking 50 resumes at it and having it analyze and screen them according to their requirements.

To achieve differentiation, I believe we need to look at how large the tech space is: the larger the tech space, the greater the differentiation that can be achieved at the technical, product, and business levels. If the technology has already converged, then all anyone can do is follow the same path, resulting in homogeneous involution.

And I’m actually quite optimistic, because there is still a huge tech space. AGI technology can be divided into three levels:

The first layer is the scaling law combined with next-token prediction (this foundation is the same for everyone, and the catching-up process is gradually converging). On this path [of scaling law with next-token prediction], OpenAI is currently doing better because they have invested the right resources over the past four to five years.

The second level has two core problems. The first is how to represent the world in a general-purpose way. True “general-purpose” representation is like a computer using 0 and 1 to represent the entire world. Transformer-based language models can represent a book, an article, or even a video — but representing a larger 3D world, or all the files on your hard drive, is still difficult. They haven’t achieved token-in-token-out, and are actually still far from the so-called unified representation. Architecture actually solves this problem.

Overcoming the bottleneck of data scarcity through AI self-evolution is another issue at the second level. Today’s AI is actually like a black box, and this black box has two inputs: a power cable and a data cable. After inputting these two things, the box can produce intelligence. Subsequently, everyone realized that the input from the data cable is limited — ie. the so-called data-bottleneck problem. The next generation of AI needs to unplug the data cable, so that as long as power is continuously input, intelligence can be continuously output.

These two core problems lead to enormous space at the third level, including long context, cross-modal generation, the model’s multi-step planning capabilities, instruction-following capabilities, various agent functionalities, and so on.

These higher-level elements will all have enormous differentiation, because there are two important technical variables in between. I believe this is our opportunity.

In addition to the technical level, our values differ somewhat from OpenAI: we hope that, in the next era, we can become a company that combines OpenAI’s technology idealism with the business philosophy of ByteDance. I believe the Asian mindset towards commercialization has certain merits. If you don’t care about commercial value at all, it’s actually very difficult to create a truly great product, or to make an inherently great technology even greater.

TikTok owner ByteDance sees valuation drop a quarter to US$300 billion ...

Overseas Unicorn: What kind of story should AI model companies tell? Should they frame their narrative around the pursuit of AGI, like OpenAI, or focus on becoming a super app? Are these two narratives in conflict, and how should they be balanced?

Yang Zhilin: The way a company tells its story depends on investors’ mindsets. For us, the more important question is understanding the relationship between these two goals.

AGI and product development are not a means-to-an-end relationship for us; they are both goals in themselves. In the pursuit of AGI, I believe the so-called "data flywheel” is crucial, even though it’s a somewhat old concept.

Products like ChatGPT haven’t yet fully established a continuous evolution loop based on user data. I think this is largely because base models are still evolving — when a new generation is developed, previous user data becomes less useful. This is tied to the current development stage — today, progress is driven by the scaling laws of base models, but in the future, there could be a shift toward leveraging the scaling laws of user data as a source of progress.

Historically, almost all successful internet products have ultimately relied on scaling user data. Today, we can already see signs of this with MidJourney. By leveraging the scaling laws of user data, it has managed to outperform simple base model scaling. However, when it comes to language models and text generation, the scaling effects of base models still far outweigh those of user data. That said, I believe this will eventually shift towards user data scaling — it’s just a matter of time.

This is particularly important now, as we face data bottlenecks. Human preference data, for example, is extremely limited but also indispensable. I believe this is one of the most critical challenges for every AI-native product today. A company that doesn’t care enough about its users may ultimately fail to achieve AGI.

Overseas Unicorn: What’s your view on MoE (Mixture of Experts)? Some argue that MoE isn’t truly a form of scaling up and that only scaling up a dense model improves a model’s capabilities.

Yang Zhilin: You can think of models with MoE and without MoE as following two different scaling laws. Fundamentally, a scaling law describes the relationship between loss and parameter count. MoE changes this function, allowing you to use more parameters while keeping FLOPs (floating point operations per second) constant. Meanwhile, synthetic data changes a different relationship — it allows for data scale growth while keeping FLOPs unchanged.

Following a scaling law is a predictable path, and people try to modify specific relationships within these laws to achieve greater efficiency. That extra efficiency becomes their competitive advantage.

Right now, many believe that simply implementing MoE is enough to achieve something like GPT-4. I think this view is too simplistic. Ultimately, the more fundamental challenge is how to establish a unified representation space and a scalable data production process.

Overseas Unicorn: If compute were sufficient, would anyone build a trillion-parameter dense model?

Yang Zhilin: That depends on how fast inference costs decrease, but I definitely think someone would. Right now, inference costs are too high, so everyone is making trade-offs. However, if compute weren’t a constraint, training a trillion-parameter dense model would undoubtedly perform better than a model with only hundreds of billions of parameters.

Overseas Unicorns: Anthropic has been emphasizing model interpretability, which has sparked a lot of debate. What’s your perspective on interpretability?

You just mentioned that models are a “black box,” and we still don’t fully understand how the human brain works either.

Yang Zhilin: Interpretability is fundamentally about trust. Building a system that people can trust is important, and the applications related to this might be quite different from something like ChatGPT — such as integrating long-context models with search.

If a model never hallucinates or has an extremely low hallucination rate, interpretability wouldn’t even be necessary, because everything it says would be correct. Also, interpretability itself can be seen as part of alignment — for example, chain-of-thought reasoning can be considered a form of interpretability.

Hallucinations can be addressed through scaling laws, but not necessarily in the pre-training stage. Alignment itself also follows a scaling law, meaning it can be solved as long as the right data can be found. AI, at its core, is just a set of scaling laws.

Overseas Unicorn: What are your expectations for AGI? At its core, isn’t the transformer still a statistical probability model? Can it lead to AGI?

Yang Zhilin: There’s nothing wrong with statistical models. When next-token prediction is good enough, it can balance creativity and factual accuracy.

Factual accuracy is usually a challenge for statistical models, but today’s LLMs can exhibit highly peaked distributions. If you ask a model a question like, “What is the capital of China?” then the model can assign a 99% probability to the character “Bei” (as in Beijing). At the same time, if I ask it to write a novel, the probability distribution of the next word would be much more evenly-distributed. Probability is really a method of general-purpose representation (通用的表示方式). In this world, there is a vast amount of entropy. We need to capture the deterministic elements while also allowing the inherently chaotic aspects to remain chaotic.

To achieve AGI, long-context will be a crucial factor. Every problem is essentially a long-context problem — the evolution of architectures throughout history has fundamentally been about increasing effective context length. Recently, word2vec won the NeurIPS Test of Time award. Ten years ago, it predicted surrounding words using only a single word, meaning its context length was about 5. RNNs extended the effective context length to about 20, LSTMs increased it to several dozen, and transformers pushed it to several thousand. Now, we can reach hundreds of thousands.

If you have a billion-token context length, then the problems we face today would no longer be problems.

Additionally, lossless compression is essentially the process of learning determinism from chaos. An extreme example is an arithmetic sequence — given the first two numbers, every subsequent number is deterministic, meaning there is no chaos, so a perfect model can reconstruct the entire sequence. However, real-world data contains noise. We need to filter out this noise so the model only studies the learnable content. During this process, we must also assign appropriate probabilities to uncertainties.

For example, if you generate an image, its loss will be higher than that of generating text because images contain more chaos and information. However, the key is to capture only the aspects you can control, while treating the remaining uncertainty probabilistically. Take a water cup as an example — whether its color is green or red is a probability-based variation, but the shape of the cup remains unchanged. Therefore, the priority is learning the cup’s shape, while its color should be treated probabilistically.

Overseas Unicorn: What patterns exist in the increase of context length? Is there any technological predictability?

Yang Zhilin: I personally feel that there is a Moore’s Law for context length. However, it’s important to emphasize that accuracy at a given length is also crucial. We need to optimize both length and accuracy (lossless compression) simultaneously.

As long as we ensure the model’s capability and intelligence, I believe the increase in context length is very likely to follow exponential growth.

02. Multimodal: most architectures aren’t worth scaling up

Overseas Unicorn: Everyone anticipates multimodal technology will explode in 2024. Compared to text, where do the technical challenges of multimodal lie?

Yang Zhilin: Currently, state-of-the-art video generation models actually use at least an order of magnitude fewer FLOPs than language models. It’s not that people don’t want to scale them up; it’s that most architectures aren’t worth scaling.

In 2019, the most popular architecture was BERT, and people later asked why nobody scaled up BERT. The truth is that architectures worth scaling need to have both scalability and generality. I don’t think BERT lacked scalability, but you can clearly see it lacked generality — no matter how much you scaled it, it could never write an article. Multimodal has also been stuck on architecture issues for the past few years, lacking a truly general-purpose model that people are willing to scale. Diffusion clearly isn’t it — even if you scaled it to the heavens, it could never be AGI. Today, auto-regressive architectures have brought some new possibilities, sacrificing some efficiency to solve the generality problem.

Auto-regressive architectures themselves are scalable, but tokenizers might not be, or eventually tokenizers won’t be needed at all. This is a core problem for 2024.

Overseas Unicorn: If tokenizers aren’t scalable, do we need a completely new architecture beyond transformers?

Yang Zhilin: Just talking about transformers themselves, I don’t think there’s a major problem. The core issue still is solving the tokenizer problem. The transformer architecture has actually already undergone many changes — today’s implementations for long-context and MoE aren’t standard transformers. But the spirit or ideas behind transformers will definitely exist for a long time. The key is how to solve more problems based on these foundational ideas.

Overseas Unicorn: If context length becomes infinitely long, we wouldn’t need tokenizers anymore?

Yang Zhilin: Correct. Essentially, if a model is strong enough, it can process any token, pixel, or byte. With infinite context length, you could directly input everything on your hard drive to it, and it would become your real new computer, taking actions based on all that context.

Overseas Unicorn: Leading model companies like OpenAI and Anthropic think a major bottleneck in 2024 will be data — so they have high expectations for synthetic data. What’s your view on synthetic data?

Yang Zhilin: A scalable architecture is the foundation — and this architecture must first support continuously adding more data before data truly becomes the bottleneck. The data bottleneck we’re talking about now will be encountered in the text modality in 2024, but introducing multimodal data will delay this problem by one to two years.

If the bottlenecks in video and multimodal can’t be solved, then the text data bottleneck will become critical. We’ve actually made some progress on this — if the problem is constrained, such as mathematics or code writing, data is relatively easy to generate. For general-purpose problems, there isn’t a complete solution yet, but there are some directions worth exploring.

Overseas Unicorn: Will the bottleneck in 2025 be energy? Because by then, individual clusters will be very large, which will bring energy challenges.

Yang Zhilin: These problems are actually connected. Eventually, multimodal might solve the data problem, and synthetic data might solve the energy problem.

By the GPT-6 generation, players who master synthetic-data technology will show clear advantages. This is because there are two types of data: “pre-training” data and “alignment” data, the latter of which is more costly to obtain. If you master data-generation technology, the cost of alignment might decrease by several orders of magnitude, or you could produce several orders of magnitude more data with the same investment, changing the landscape.

I think 2025-2026 might be an important milestone: most of the model’s computation will occur on data generated by the model itself.

By 2026, the amount of computation used by models for inference might far exceed training itself; you might spend 10 times the cost on inference, and then one-tenth of that cost on training. A new paradigm will emerge: inference becomes training, and this inference doesn’t serve any users — it only serves to generate synthetic data for itself.

If this happens, the energy problem is also solved, because inference can be distributed. It doesn’t violate any laws; it’s essentially energy conservation. I’m just changing the computational paradigm to allow energy to be solved in a distributed way.

03. Super App: Model Fine-Tuning May Eventually Not Exist

Overseas Unicorn: The search and recommendation systems behind Google and Douyin have strong flywheel effects: their algorithms can provide real-time feedback based on user behavior, continuously improving user experience. LLMs, however, currently can’t provide real-time feedback on user behavior. What will the flywheel effect of AI-native products be?

Yang Zhilin: I’ve thought deeply about this question. The ultimate, core value of AI-native products is personalized interaction, which is something previous technologies haven’t implemented well. So this question is actually about personalization — how to enable users to gain highly personalized interactive experiences the more they use your product. For many products today, the degree of personalization is almost zero. Previously, we could only do personalized recommendations, but now users can interact with products. This interaction is highly anthropomorphic and personalized. How do we achieve this?

I think this is fundamentally a technical issue. In the traditional AI era, achieving personalization required continuously updating models, using small models to solve specific problems. In the large model era, one way to achieve personalization is through fine-tuning — but I believe fine-tuning may not be the fundamental method and may not exist in the long term. Why? When your model’s instruction-following ability, reasoning ability, and contextual consistency ability become stronger, everything only needs to be placed in memory. For example, your large model’s memory can have a bunch of prefixes to follow, reducing costs dramatically. Ultimately, the process of personalizing a model is actually your entire interaction history — which is a collection of your preferences and feedback. This feedback is more direct than products from previous eras because it’s generated entirely through conversational interfaces.

Based on this judgment, the next question is: how to achieve long-context-based customization at the technical level to completely replace fine-tuning?

I believe we’re moving in this direction now. Future models won’t need fine-tuning but will instead solve problems through powerful contextual consistency and instruction-following capabilities. The long-term trend should be personalization of the underlying technology, which will be a very important change.

For example, GPT-4 brought a new computing paradigm where creating GPTs doesn’t require fine-tuning. Previously, customization was achieved through programming, but today it’s achieved by making the model’s prefix very complex, and extracting what you want from this general-purpose set. Personalization achieved this way is truly AI-native personalization, and a traditional recommendation engine plug-in will definitely be eliminated by this new approach.

Overseas Unicorn: How did you make the decision to first develop lossless long-context?

Yang Zhilin: I think the most important thing is to begin with the end in mind. Large models as new computers definitely need large memory, because the memory of old computers has increased by at least several orders of magnitude over the past few decades, and old computers also started with very little memory. The second point is that the ultimate value of AI is personalization.

Overseas Unicorn: OpenAI also has some long-context capability now.

Yang Zhilin: But they haven’t truly viewed the user interaction process as a personalization scenario. For example, if we prompt ChatGPT with something, regardless of whether it’s today or tomorrow, as long as the model version is the same, the effect is basically the same. This is what I mean by a lack of personalization.

Ultimately, everything is instruction-following. It’s just that your instructions will become increasingly complex. Today, your instruction might start with 10 words, but later it could be 10,000 words or even 1 million words.

Overseas Unicorn: Chatbots have always been the ideal for AI scientists. If each user has hundreds of conversations with a chatbot daily, and the chatbot system can collect and understand more user context, will it ultimately far exceed the matching accuracy of search and recommendation systems? Like interactions between colleagues or family members, where just one sentence or even a glance is enough to understand each other.

Yang Zhilin: The key is crossing the trust threshold.

I think the ultimate measure of an AI product’s long-term value is how much personalized information users are willing to input into it, and then lossless long-context and personalization are responsible for turning these inputs into valuable outputs.

New hardware forms may also be needed — but I think models and software are still bottlenecks. To dig deeper, the prerequisite for users to input a lot of information is trust — you need a sufficiently engaging and human-like AI. You can’t say, “I’m setting up product features specifically to get your information.” The end result should be that users and AI become friends, so users can tell the AI anything.

Inflection Pi’s motivation is actually good — wanting to establish strong trust — but Pi may need to take another step forward. How to build trust with users? Human society probably won’t accept being assigned a lifelong companion; that seems somewhat against human nature.

Overseas Unicorn: Moonshot AI wants to create a super app. What does your ideal super app look like? How big does it need to be to qualify as “super”?

Yang Zhilin: It’s about breaking out of niche adoption. When all your relatives are using it, only then have you truly become a super app. And I believe that improvements in AI capabilities will lead product adoption. For example, if character.ai were a perfect multimodal model today, I think its chances of breaking out of its niche would be at least 10 times greater. Ultimately, an application’s ceiling is reflected in the year-over-year increase in connections between AI and humans.

04. Moonshot AI: People with the ability to unlearn make the best talent

Overseas Unicorn: What does the ideal CEO for an AGI company look like?

Yang Zhilin: On one hand, there needs to be a tech vision. You can't just keep doing things that have already been proven to work by others. A real AGI company must have its own unique technical judgment, and this judgment should influence the overall direction of the company. If the top leader can't make decisive calls, that won't work either. At the beginning of the year, we were already working on auto-regressive multimodal models and lossless long-context, but these only became extremely popular in the last couple of months. Even today, lossless long-context is still not widely accepted as a consensus. If you only start noticing these trends now, there won’t be enough time to iterate, and in the end you'll just become a follower.

Another point is having a profound understanding of AI-native product development and then adapting the organization to this new mode of production. In the past, product development was about understanding user needs and designing features accordingly. But in this new era, design needs to be completed during the manufacturing process. ChatGPT’s design was finalized through its creation — it wasn’t built by pre-defining a bunch of scenarios and then finding corresponding algorithms. Similarly, Kimi users uploading resumes and using it for screening was a completely untested use case before we launched, yet it emerged naturally from real-world usage.

Resource acquisition is also crucial, with compute power being the primary cost driver. In the early stages, funding is key, but later on, product commercialization becomes necessary. However, commercialization cannot simply copy mature models from the previous era; it requires innovation. A good CEO and team should have some experience but also possess strong learning and iteration capabilities.

Overseas Unicorn: But maybe some investors can’t tell whose “tech vision” actually leads the pack.

Yang Zhilin: I’m not too worried about this problem. What we have now is the best distribution mechanism: it’s close to a real free market and we will end up with the most efficient resource distribution. What we need to prove to others is not the value of our vision, because a vision is an abstract thing. We need to prove our worth through delivering real models and products. Anthropic received much more funding immediately after it released models like Claude. The market is fair.

Overseas Unicorn: From the perspective of building product- and company-competitive moats, the industrial era relied on economies of scale, and the internet era emphasized network effects. Will there be a new paradigm in the AGI era?

Yang Zhilin: In the short term, changes in organizational structure drive technological advancements — better technology is achieved through better organization, which then directly translates into a superior product experience.

In the long term, network effects are still likely to dominate. The question is: how will they manifest? Traditional two-sided networks from the internet era may still exist, but not necessarily in the form of users and content creators. For AI-native products, the two-sided network effect may be reflected in personalization, where users and the AI engage in a co-creative relationship.

So right now, I see two key areas worth exploring: the continuous improvement of model capabilities and the development of two-sided network effects. These will shape new paradigms in the AGI era. Midjourney has already seen explosive growth through its two-sided effect, while Stable Diffusion, as an open-source model, faces the challenge of being too fragmented on a single side, instead relying solely on base model improvements.

Overseas Unicorn: From the hiring perspective, how do you define strong talent?

Yang Zhilin: I break it down into experience and learning. The ability to learn is a general-purpose capability, which not only includes learning but also unlearning — especially unlearning previous experiences of success. Let’s say you built YouTube from 0 to 1; you might find it harder to work on AI products now than other people do, because you have to unlearn a lot of things. Learning is more important than experience. Maybe in 5 years, the AI industry will cultivate a large number of so-called mature roles. Currently, I don’t actually think that dividing people by roles is all that meaningful, since every person needs to be multi-faceted.

Overseas Unicorn: What kinds of researchers possess “tech vision”?

Yang Zhilin: The core ideas are twofold: focusing on the big picture while letting go of the small details, and maintaining an endgame mindset. I’ve worked with many researchers, and a common issue is over-optimization — getting caught up in refining details while missing the broader perspective. For example, we saw that transformers solved the context length limitations of LSTMs, but if we take a step further back, we realize that each generation of technology is fundamentally about extending context length.

Overseas Unicorn: How many more of these people do you think Moonshot AI still needs?

Yang Zhilin: Objectively speaking, the real limit for us is still supply. Currently, experienced AGI talent is very rare, but there are lots of people with the ability to learn.

But from a demand perspective, the organization cannot become too large — if it turns into just another Big Tech corporation, many of its organizational advantages will be lost. So we will definitely maintain a lean and highly efficient structure. One key judgment is that AGI does not require that many people. In the long run, once we truly “unplug the data,” models at the level of GPT-6 and beyond should be able to evolve on their own, breaking through the limits of human capability.

Overseas Unicorn: How do you assess the difficulty and timeline for catching up with GPT-4?

Yang Zhilin: Hitting benchmark scores on par with GPT-4 is very easy, but achieving its actual performance is definitely challenging. It’s not just a matter of resources — Google has already demonstrated this. In fact, the training cost of GPT-4 isn’t that high; several tens of millions of dollars is not an intimidating figure. This is positive news for us, and we’ve even already made substantial progress.

The most critical factor is having a strong tech vision to anticipate what GPT-5 and GPT-6 will be, and then executing and building the necessary foundations ahead of time. Otherwise, it’ll never be possible to surpass OpenAI. Much of OpenAI’s advantage comes from its foresight — by 2018, it had already committed to what it believed was the right path and spent years building deep capabilities.

Overseas Unicorn: If you were to develop an image-generation AI, how would you approach it? How would you balance language comprehension and image quality?

Yang Zhilin: Midjourney has already done exceptionally well in the single task of image generation. If I were to develop a similar product, I would want it to handle multiple tasks, while still excelling in certain key areas. This is actually the same approach OpenAI attempted, but they didn’t quite succeed.

An AGI company should focus on becoming the default platform — the primary way users interact with AI. Meanwhile, niche user groups will still have specialized needs and ultra-high standards for performance, which is why there’s room in the market for companies like Midjourney. However, if AGI becomes powerful enough, many users will migrate. For example, if I were to repackage all of Photoshop into a single prompt — essentially turning it into an outsourced all-in-one designer — then fewer people would use Midjourney.

Midjourney’s current dominance comes from its first-mover advantage, which enabled it to kickstart a powerful data flywheel. The tricky part is whether such a time window will exist in the future — if not, general-purpose models may eventually outcompete and overtake it.

Overseas Unicorn: Following the strategy of becoming the default platform, how many key user entry points do you foresee in the future?

Yang Zhilin: At least two — one for utility, the other for entertainment.

The way we access information today may become obsolete because, at its core, searching for information is just a means to an end — we do it to complete a task from start to finish. In the future, AI-driven interfaces will likely replace search engines as the primary way users interact with information. Retrieving information is never the end goal; it has just been artificially framed as one. Sometimes we want to accomplish a task, and other times, we want to learn something new. The ideal AGI interface should directly help users complete tasks, rather than simply helping them find information.

Overseas Unicorn: From today onward, how much investment do you think it will take to realize your vision of AGI?

Yang Zhilin: Achieving a fully realized AGI will require investment on the scale of tens of billions of dollars. However, it won’t be a one-time expense — it’s about setting up a self-sustaining loop where the business can generate the necessary resources to fuel further development. This multi-billion-dollar estimate is based on the need to scale up by at least two to three orders of magnitude. Of course, costs will be optimized along the way.

Overseas Unicorn: What should the business model of an AGI company look like? Will it still be seat-based or usage-based?

AGI delivers varying levels of value depending on the task it completes. It may operate more like an outsourced service, pricing each task individually. Beyond that, advertising will undoubtedly play a crucial role. With deeply personalized interactions and conversational engagement, ad monetization could become significantly more efficient than it is today.

Overseas Unicorn: If training costs for models like GPT-4.5, Claude-3, and Gemini-2.0 are around $300 million today — and future models in 2025 could require tens of billions of dollars — does that mean the pursuit of AGI is a trillion-dollar gamble? Have you considered its ultimate impact on human society?

Yang Zhilin: One impact that’s almost certain is a real and tangible increase in productivity. Today, a single piece of software might function at the intelligence level of 1,000 programmers, but in the future, applications could be powered by the equivalent of a million programmers, continuously improving through iteration.

Thinking about the possibilities, everything we take for granted today could change. Training models on a vast range of languages and cultures will inevitably influence values and perspectives. The way people allocate their time will shift — fewer people may work purely for money, and more of human life may be spent in digital or intellectual spaces. Ultimately, we may see the emergence of a massive virtual cognitive ecosystem. To truly build the Metaverse, we may first need to perfect AI.

Additionally, I firmly believe AGI will be inherently global.

Overseas Unicorn: Right now, leading AI models are both powerful and relatively inexpensive, leading to a strong Matthew effect [self-reinforcing cycle in which early winners keep accumulating advantages]. Doesn’t that mean the final market landscape will be highly consolidated?

Yang Zhilin: Within a five-year window, top players might still dominate. However, in 50 years, I believe AGI will be fully commoditized — it will be no different from electricity today.

ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.

新一年的春天,我尝试继续恢复活力和精神

为全球华人游荡者提供解决方案的平台:游荡者(www.youdangzhe.com)
这世界的辽阔和美好,游荡者知道。使用过程中遇到问题,欢迎联系客服邮箱wanderservice2024@outlook.com.

【和放学以后永不失联】订阅放学以后Newsletter,每周三收到我们发出的信号:afterschool2021.substack.com 点击链接输入自己的邮箱即可(订阅后如果收不到注意查看垃圾邮箱)。如需查看往期内容,打开任一期你收到的邮件,选择右上角open online,就可以回溯放学以后之前发的所有邮件,或谷歌搜索afterschool2021substack查看。

截至目前,放学以后Newsletter专题系列如下:“在世界游荡的女性”系列、“女性解放指南”系列、“女性浪漫,往复信笺”系列、莫不谷游荡口袋书《做一个蓄意的游荡者》系列、“莫胡说”系列”《创作者手册:从播客开始说起》,播客系列和日常更新等。

大家好,本期Newsletter由世界时区东一区的霸王花木兰轮值。

先和大家预告一下:放学以后after school播客三个月的Gap即将结束,新一期播客《在诗意匮乏的时代,我们一起来读读诗》将于4月2日全网上线,敬请期待!

接着和大家分享下我在西班牙独自生活的日常,以下是正文:

说实话,现在的思绪有些琐碎繁杂。窗外正值狂欢过节,时不时烟花炮声隆隆,为了减少连续多日的噪音干扰,我把窗户最外层百叶帘式的铁质窗关闭,里面一层纱窗关闭,第三层推拉玻璃关闭,最后把幕布卷轴窗帘拉下来,如果不开灯,白天的房间也可以黑到伸手不见五指。晚上房间屋顶的灯模糊不清呈现出暗暗的淡黄色,我把桌上明黄色的台灯打开,手机支架上播放着平静心情的音乐,期待杂乱的心安静一下,手指敲击键盘琢磨着写点什么。

3月,新一年的春天来了,正是万物复苏的季节,我想记下来春天感到愉悦和满足的时刻,分享恢复活力和精神的一些方法,却想到前两天网上看到的一则帖子,说这个春天真难熬。恰好此时此刻的我心情算不上愉悦,便去询问Chatgpt为什么春季情绪容易波动,人容易觉得疲惫?给出的回答让我有些讶异,但又觉得有些合理。

Chatgpt解释说一是由于日照时间增加,生物钟被打乱,人就容易失眠或者情绪低落,二是激素水平波动,血清素(幸福激素)褪黑素(睡眠激素)皮质醇(压力激素)变化会带来情绪起伏不定,引发焦虑、烦躁和抑郁;三是春季花粉过敏,温度变化会影响免疫系统和情绪波动。春季这种季节性情绪障碍还有个专属名词叫Spring SAD-Springtime Seasonal Affective Disorder。而令我讶异的答案是第四个诱发因素,人们对春天的期望容易诱发负面情绪:“社会对春天的期望是'万物复苏,充满希望',但如果现实生活并未如预期般改善,可能会让人感到失望,甚至陷入自我怀疑。”这是一个我没想到的角度,所以修改一下上一段的表述:3月,新一年的春天,这是万物可能复苏也可以不复苏的季节。

目前正在打字的我处于不复苏的状态,前段日子有不少复苏的时刻,所以我来分享下那些让我感到复苏,身体有活力,精神有愉悦的时刻:

一、从讨厌跑步到奇妙地有些上瘾:运动真的能改变大脑吗?

我讨厌跑步。

一方面体力不支,沉重的身体根本跑不起来,跑两步就喘不过气,感觉喉咙要冒血,还要担心跑步带来的膝盖压力和损伤。另一方面坚持不了,过去的三十多年有过很多次跑步运动的念头,大部分陨灭在两天打鱼三天晒网中。于我而言跑步是一件乏味枯燥的事情,在跑步机上跑步必须找能够转移注意力的视频,否则会因为枯燥重复而抓狂,在马路上跑步需要靠意志力坚持,这方面我的意志力也经不起考验。

与跑步相反,我喜欢走路。走路对体力的要求远低于跑步,边走边逛也会让走路变得轻松。和莫不谷一起多次的世界游荡中,我们都会选择能走路就走路,主要是通过暴走消食,给下一顿美食腾出胃口。来到西班牙后,经常因为四处游走探索城市徒步10多公里,有时甚至20公里。前段时间我突然发现,人走着走着会想要跑步,甚至身体跑起来感觉更舒服。

这令我觉得奇怪,Chatgpt答案说:身体对于运动是有渴望的。长时间步行后,人的肌肉和神经会被进一步激活,身体本能地会对强度更高的运动有渴求,这也是为什么走路走多了跑步更有愉悦感。获得这个答案的时候,我突然能理解健身狂魔为啥能以一般人无法理解的程度对运动上瘾,甚至越运动越兴奋。

同时想到,莫不谷在最新一期单口播客《美妙人生的关键呀,让我们一起扭一扭它》分享大脑对新知的渴求,原理大概也是如此。体力,越用越有,脑子,越用越好。一个持续运动的人,会激发身体对于更多运动的渴望,一个持续学习的人,会刺激大脑对于更多新知的渴求。因为我持续在走路锻炼,所以有机会感受到身体本能的渴求,那我的大脑活力,是否也因为缺乏主动使用和锻炼,所以无力应对和处理新知,是否也可以像身体这样恢复本能,感受新知刺激的快乐?在运动中,我为自己大脑的节奏找到了可被解释的理由和复苏的可能。

虽然我的身体会觉得跑步更愉悦满足,但我的体力只能支持跑一会,所以下面是不跑步的人关于跑步的一些经验:

  • 一是数秒小跑,解决体力不支的问题。我会在走路中间插一段跑步,大概从1数到100,跑100秒,不会太困难也不会太枯燥。100秒不够,就可以再插一段200秒的跑步,循序渐进,量力而行。

  • 二是设定目标,解决跑步枯燥的问题。在西班牙,我喜欢逛超市,所以我常常在去超市的路上数秒小跑,前方有期待,人就会动力十足。还会在回家的路上数秒小跑,跑的越多,越快回家休息。

  • 三是不拘小节,解决跑步出汗的问题。如果天气凉爽,不必跑完就洗澡,隔一天再洗问题也不是很大。衣服干了再穿一天,反正总会汗湿,攒一天再洗节水节能又方便。(洁癖人士不必参考本条建议)

跑步会让你满脸通红,浑身发汗,衣服汗湿,紧紧粘在背上,头发出油,一整个狼狈,也会让你在一阵风吹过时,感到凉爽宜人,在身体活动起来后,感到兴奋与愉悦。

二、从不会做饭到常常做饭探索美食:喂饱自己真的能自我满足吗?

记得莫不谷曾经分享过,做出想吃的美食能感受到充足的自我认同与自我满足。在国内生活的时候,我很难充分感受到这一点,极其偶尔地尝试自己做饭,也不会因为下厨获得很大的满足。来到欧洲生活后,我常常会因为做出食物喂饱自己感到满意,如果做出好吃的食物自我满足感就更高。我也在想这个问题,这是为什么?

我发现,一个可能的原因是,便利快捷的生活会削弱人们的满足感。在国内生活时,不会做饭对生活影响不大。在劳动力被层层剥削压榨情况下,想吃什么点外卖,20分钟,30分钟火速送达,不想外卖就出门,五花八门琳琅满目的餐厅、美食会让你陷入选择纠结。除了特别热爱下厨和不喜欢外食觉得不卫生的人,大多数情况下,自己做饭是一个耗时耗力且缺乏成就感的事情。由于不需要“亲自投入”就可以通过消费购买各式各样的服务,需求被快速满足的同时需求本身也容易被忽略。对我来说就是,吃饭重要吗?还行。美食有满足感吗?还行。吃吗?不吃也行。

在欧洲生活,做饭这项基本生存技能很重要。首先是饮食习惯差异,来到了一个与中国胃不完全匹配的环境,想吃的不一定有,花钱买也买不到。其次是人工成本高,无论是餐厅吃饭还是外卖吃饭,对一般人来说都不具有长期可持续性。再者是生活节奏慢,欧洲人注重Work life balance,不可能随时随地有餐厅营业,也不可能30分钟外卖火速送达。在这种情况下,做饭这件事成了能不能维持人体生存的关键因素,如果做的好吃,自我的满足感成就感会更高。你能充分感受到人可以不必依赖社会系统,不必假手于人,凭借个体力量满足个体需求的自我认同。我想,这就像鲁滨逊流落荒岛之后,肯定能从亲手造房子、种粮食水果,做饭获得极大的自我认同和满足。

当决定来欧洲生活但又不会做饭的时候,我就在游荡世界的中途和亲朋好友预先学习了好吃的菜谱,还分享到了游荡者平台上(www.youdangzhe.com)。来到西班牙生活的这两个月,从询问Chatgpt厨房小白做饭建议采购什么,到时不时学习研究网上简单快手的美食,莫不谷和其她朋友还会分享美食秘诀,有时候晚上睡觉想到第二天起来做什么吃的,就会开始期待明天,这是过去在国内生活比较少的体验。我还时不时随机开直播分享自己学到的美食:不用去齐齐哈尔就能吃到的齐齐哈尔烤肉,超简单辣椒炒香干、东北沈阳烤鸡架、凉拌东北鸡架、凉拌木耳粉丝和大虾、电饭煲炖鸡腿、简单好吃的摊煎饼,等等。在下厨过程中,我还学到了蔬菜水果肉类的保存方法,认识许多陌生食材,如西班牙小扁豆、西班牙洋蓟,等等。

(我在游荡者平台分享的美食攻略)

从不会做饭到常常做饭探索美食,我想,喂饱自己的自我满足不仅是上述原因,还有一点是,它让生活变得简单、具体,人在与生活非常近距离地接触中慢慢复苏和感受。

由于写着写着发现收不住,所以现在要快速结尾一下。除了运动、做饭,在西班牙生活的我还感到复苏有活力的时候是逛超市。来到西班牙后逛超市变成一件很有意思的事情,甚至像是休闲娱乐活动。由于超市里全部是西班牙语,很多都是陌生的食材物品,每次去超市都像是寻宝探险。连买什么橄榄油、买哪一种黄油(第一次买错买成酵母),哪一种芝士都需要学习研究半天,在欧洲生活逛超市的感受大概是婴儿刚成长发育时期需要有“五感”体验那样,用眼睛看一看,用软件来翻译,用手摸一摸,靠近闻一闻,买回家尝一尝,新奇,有趣,可能会被种草,也可能踩雷,完全盲盒体验。有意思的是,由于西班牙超市众多,每种超市种类价格各不相同,逛起来还有点像木兰替父从军前的准备工作:“东市买骏马,西市买鞍鞯,南市买辔头,北市买长鞭。”。我想,逛超市的满足大概是新鲜环境和新鲜食物对人持续产生刺激。

说完感到复苏的体验,再来分享下感到还没复苏的体验,一是最近阅读不多,有一些想到要阅读的书籍,加入到书单后便没了下文,大概是阅读需要集中注意力和思考力,也需要静下心来,比起阅读,做点体力活动对我来说相对更轻松。

二是学习仍会觉得疲累,这个疲累不是学习本身带来的,而是一想到学习,很多挫败的经验情绪就会上涌,想要“一步到位”和急于求成的心态也成为挡路石。前两天我偶然遇到了一本书《我不原谅: 一个90后对中国教育的批评和反思》,看到标题便忍不住翻看,想从对方的批驳中看看有哪些共鸣之处,又是否能找到启发之处。看了开头还没继续看下去的原因是,你都不需要了解作者,就能从文字里看出性别。

三是自我觉察还不充分。这个小标题写出来感觉好像上班在写报告,实际是在想莫不谷在做终身学习系列第000期《美妙人生的关键呀,让我们一起扭一扭它》时的一个提议:选择自己真的感兴趣,能帮助自己解决最核心问题或者服务自己终极渴望的东西去学习。感兴趣,我可以回答,“真的”感兴趣,回答起来有点困难,核心问题和终极渴望,回答起来也有困难。我想,这里的问题并不是我一定得有终极渴望,而是我对我个人是否坚定,认同,我对自我的选择、当下正在过的生活和未来有可能的生活是否足够觉察、认可,这对我来说还挺有挑战。

最后,祝福大家在这个春天多多在大自然和个体生活里感受复苏和生命活力,扔掉一些社会、集体、Ta人不合理也不必要的期待和包袱,赶不上这个春天,还有下一个春天,生命里每个季节都可以用来慢慢复苏。

为全球华人游荡者提供解决方案的平台:游荡者(www.youdangzhe.com)
这世界的辽阔和美好,游荡者知道。使用过程中遇到问题,欢迎联系客服邮箱wanderservice2024@outlook.com.

【放学以后文章&书籍&其它】

解锁放学以后《创作者手册:从播客开始说起》:https://afdian.com/item/ffcd59481b9411ee882652540025c377

解锁莫不谷《做一个“蓄意”的游荡者》口袋书:
爱发电:https://afdian.com/item/62244492ae8611ee91185254001e7c00微信公众号:《放学以后After school》(提示安卓用户可下载“爱发电”app,苹果用户可把爱发电主页添加至手机桌面来使用,目前爱发电未上线苹果商店)

Newsletter订阅链接:https://afterschool2021.substack.com/(需科学/上 网)

联系邮箱:afterschool2021@126.com (投稿来信及合作洽谈)

为全球华人游荡者提供解决方案的平台:游荡者(www.youdangzhe.com)

小红书:游荡者的日常

同名YouTube:https://www.youtube.com/@afterschool2021

同名微信公众号:放学以后after school

欢迎并感谢大家在爱发电平台为我们的创作发电:https://afdian.com/a/afterschool

播客收听平台:【国内】苹果播客(请科学/上网)、爱发电、汽水儿、荔枝、网易云、小宇宙、喜马拉雅、、QQ音乐;
【海外】Spotify、Apple podcast、Google podcast、Snipd、Overcast、Castbox、Amazon Music、Pocket Casts、Stitcher、Radio Public、Wordpress

川普启动“外国敌人法案”

19 March 2025 at 01:19

3月15日是星期六,我跟一群朋友在山姆·休斯顿国家森林徒步,走了20公里。因为刚下过雨,一些低洼路段成了沼泽,要在泥泞中跋涉,鞋上和裤脚上全是泥水,还有两名队友掉到水里。快结束的时候,大家筋疲力尽,饥肠辘辘,开始聊吃的。我说,现在只想要个芝士汉堡,加一杯冰镇啤酒。

开车回家的路上,一边找汉堡店,一边听新闻。新闻说,几个小时前,川普总统颁布行政命令,启动“战时外国敌人法案”。有国土安全部的两架飞机从得克萨斯跟墨西哥边境的Harlington起飞,分别飞往萨尔瓦多和洪都拉斯。飞机起飞不久,首都华盛顿联邦法院一名法官,勒令飞机掉头返航。萨尔瓦多总统布克尔对媒体上说:“可惜,晚了”。

美国各大报都在显著位置报导了这个事件。下午去社区健身房,墙上的电视新闻也在报导这个事件。“外国敌人法案”到底是部什么法律?川普为什么要启动这个法案?谁在那两架飞机上?法官下令后,飞机为什么拒绝返航?

Read more

Energy: How to Build Compute in America

17 March 2025 at 19:04

Despite leading the world in AI innovation, there’s no guarantee that America will rise to meet the challenge of AI infrastructure. Specifically, the key technological barrier for data center construction within the next 5 years is new power capacity.

To discuss policy solutions, ChinaTalk interviewed Ben Della Rocca, who helped write the AI infrastructure executive order and formerly served as director for technology and national security on Biden’s NSC, as well as Arnab Datta, director at IFP and managing director at Employ America, and Tim Fist, a director at IFP. Arnab and Tim just published a fantastic three-part series exploring the policy changes needed to ensure that AGI is invented in the USA and deployed through American data centers.

In today’s interview, we discuss…

  • The need for new power generation driven by ballooning demand for compute,

  • The impact of the January 2025 executive order on AI infrastructure,

  • Which energy technologies can (and can’t) power gigawatt-scale AI training facilities and why Jordan is all-in on GEOTHERMAL,

  • Challenges for financing moonshot green power ideas and the role of government action,

  • The failure of the market to prioritize AI lab security, and what can be done to fend off threats from adversaries and non-state actors.

Watch below or listen now on iTunes, Spotify, YouTube, or your favorite podcast app.

Constructing Compute

Jordan Schneider: Ben, why were you, as an NSC director, spending time in SCIFs dealing with energy permitting policy on federal land?

Ben Della Rocca: That’s a great question. Spending so much time reading about environmental permitting law and watching law school lectures on how the Clean Air Act works was not how I envisioned spending my time as an NSC director, but there I was doing just that in the SCIF for my AI work.

The United States has a lead in AI thanks to our thriving innovation ecosystem and the immense engineering and other talent. However, that lead isn’t guaranteed. Which country will lead in artificial intelligence will come down more and more to where AI can be built most quickly and effectively.

By “built,” I’m not just talking about the technical engineering challenge of how to do large-scale AI training runs from a computer science standpoint, but also the physical building challenge. This involves developing the large-scale computing infrastructure and energy infrastructure to execute these ever-growing AI training runs.

What became clear to us in the last administration was AI’s significant impact on national security, the economy, and scientific advancement. We also saw an exponential increase in the demand for computing power and energy resources to develop frontier models.

The amount of compute required by frontier AI models is increasing by a factor of 4-5x annually based on publicly available statistics. That’s an exponential pace of growth. Even when you factor in increasing energy efficiency of computational resources and other countervailing factors, you’re still looking at gigawatts of electricity needed to execute training runs at the frontier within the next few years.

There are multiple power-related challenges when it comes to developing and deploying AI. Assuming current trends continue, you need gigawatt-scale training facilities to develop models at the frontier — that’s one set of issues. Then there’s a separate but related set of issues around developing a robust network of potentially smaller-scale data centers around the country to actually use these tools effectively in different locations. This is a significant energy challenge as well.

At the National Security Council, I led the White House’s work around AI infrastructure and developed the AI infrastructure executive order that came out in January 2025. What the executive order tries to do is directly address that first challenge — how to build these large gigawatt-scale training clusters that, assuming the current paradigm of AI training continues, will need to be built for the United States to maintain its lead.

Jordan Schneider: That’s a tall order you set for yourself, Ben. In your mind, what parts of the executive order will matter most when building out AI data centers in the US?

Ben Della Rocca: The executive order includes a wide range of things that address not only how we bring gigawatt-scale data centers online in this country, but also a broader, distributed network of data centers around the country. There’s a lot in there, but I’ll highlight three key mechanisms.

  1. Construction on federal lands — the centerpiece of the executive order establishes a mechanism by which AI data centers can be built in a streamlined and more efficient way on federal sites owned by the Department of Defense and Department of Energy. There’s a huge value proposition to building on DOD and DOE sites because by building on federal lands, you bypass many of the state and local land use permitting requirements that usually make data center construction and similar projects take such a long time. Additional burdens on the federal permitting side are taken on as a result, but the federal government can and will be doing much under the executive order to make those processes proceed as expeditiously as possible.

  2. Regarding bringing new power generation online — operating gigawatt-scale data centers will require a lot of new power added to the electric grid. This is challenging for numerous reasons, including permitting requirements that complicate construction broadly, delays with interconnection to the electric grid, and other required approvals. The executive order directs the Department of Energy to establish requirements to collect and share information with data center developers regarding unbuilt power projects that have already received interconnection approvals to accelerate the power procurement process. It also directs the Department of Energy to engage utilities to reform their interconnection processes for faster progress.

  3. Regarding transmission — the Department of Energy is directed to use some of its powerful authorities to partner with private sector transmission developers in building transmission lines much more efficiently and quickly than business as usual, and to take part in the planning process as well. You also have other actions to bolster the supply chain for transmission and grid equipment, which would be really useful for the long-term vitality of this industry.

Between these sets of things and other actions to expedite the permitting process within the federal government’s authority, the executive order lays out a pathway for building gigawatt-scale facilities on the timelines that we expect leading developers will ultimately need for AI training.

Jordan Schneider: Tim, what's missing from the executive order?

Tim Fist: We highlight two big gaps.

  1. The EO comes along with a clean energy requirement. All the energy that you're producing to power these data centers needs to come from clean energy sources. It allows the use of natural gas with carbon capture, but that technology can’t be implemented on the timescale required.

  2. While building on federal lands owned by DOE and DOD allows you to bypass a lot of state and local permitting issues, it does open up the issue of NEPA, which automatically applies if you’re building on federal land.

Our recommendation around that is using the Defense Production Act to speed up permitting on federal land, as well as resolve supply chain issues.

DPA gives the President broad authority to intervene in the economy where this is seen as necessary to ensure the supply of technology that's deemed essential to national defense. Our claim is that AI definitely fits within this scope.

Because the most powerful AI systems are now being developed by private firms, a lot of the DoD's future capabilities are likely going to come from models that are trained in data centers operated by private firms.

OpenAI recently announced a partnership with Anduril to bring its models to the battlefield. ScaleAI has built a version of Meta’s Llama, which they call Defense Llama, to help with military planning and decision making. Palantir is building platforms for DoD as well.

Arnab Datta: We mentioned two authorities in the DPA. Title I, which is prioritization, would allow the federal government to tell contractors to prioritize transformers, turbines, and other hardware for AI data center use. The other authority in the DPA is Title III, which is a financial assistance authority. There are also authorities within the DPA and Title III where you can streamline some of these permitting issues that Tim described.

Jordan Schneider: Can you explain how DPA authorities would help?  

Arnab Datta: I’ll use a practical example — right now, natural gas turbines are sold out. GE Vernova said they could be sold out past 2030.

DPA would allow the president to say, AI is a national security priority, and those data center turbine contracts need to be fulfilled first.

I don’t want people to think that our idea and what we’re proposing is about getting out of environmental rules to build this infrastructure. The important thing is that we’re talking about using these to streamline the procedural laws associated with environmental review.

This doesn’t necessarily mean that an environmental review wouldn’t be conducted prior to leasing federal land, for example. Finding ways to limit the likelihood of litigation could be very important here, and that’s where some of the national security exemptions are most useful.

The Geothermal Goldmine

Jordan Schneider: Let’s take a step back here, because we got really deep, really fast. The renewable regulations will obviously disappear in the next few weeks, and NEPA was just nerfed by the Trump administration.

The worry for a hyperscaler might be driving their data center through the process and then getting sued four years from now in a different administration. But even then, the data centers will already be built. We’ll have AGI by then. No administration is going to shut that off because you took advantage of a permissive regulatory environment.

I’d like to discuss the hard constraints when it comes to actually building and deploying the electricity needed for these data centers. A fascinating part of Tim and Arnab’s work was going through all the potential technologies that could provide the marginal electricity needed for these data centers, and ranking them by potential. Let’s do overrated and underrated. What are three overrated sources of electricity that get too much attention in the broader discourse about the future of AI?

Arnab Datta: I don’t love the word “overrated,” but there is sometimes an implication that natural gas makes this super easy. The reality is that gigawatt-scale energy projects, particularly off-grid projects that many compute companies are moving toward, represent massive investments.

By off-grid, I mean not connected to the transmission system — you’re building close to your data center and it’s exclusively powering your facility. This is an extremely expensive and risky investment regardless of how much cash you have on hand. Building energy infrastructure at a gigawatt scale is very costly.

When discussing natural gas, yes, it’s a proven technology, but there are stranded asset risks. If an AI data center outlives its useful life in that location, or if something becomes more cost-competitive and a company wants to switch, you face challenges. Natural gas is important and should definitely be part of the mix, but I wouldn’t say it’s an easy decision. There are supply chain challenges, and the notion that “we can solve this with natural gas” oversimplifies the issue.

Jordan Schneider: Tim, what do you think?

Tim Fist: The obvious one is the much more long-term technology — fusion. We see some hyperscalers signing power purchase agreements for fusion energy. A power purchase agreement is a commitment to buy a fixed number of kilowatt-hours at a future point, and the fact that they’re buying this for fusion is rather incredible when there isn’t a viable commercial fusion reactor that’s ever been demonstrated.

This technology is clearly so far off that it won’t matter over the timeframe we’re concerned about, which is how we ensure we can build this infrastructure over the next five years.

Ben Della Rocca: Each approach has real downsides. There are real risks, problems, and inefficiencies that aren’t fully recognized.

To pick one where there’s sometimes optimism that should be qualified — we need to be realistic about nuclear options in particular. Nuclear energy is something I would be very excited about in a longer-term timeframe, such as the 2030s. However, it’s very difficult to see that being part of the solution for the late-2020s challenges we may face regarding AI’s energy use. We should be realistic about the timing involved.

Tim Fist: We’re at this weird point in history. Twenty years ago, the answer would have been obvious because renewables were completely non-viable, and there weren’t exciting next-generation technologies coming along.

Right now, we’re at a point where multiple technologies are approaching the same level of cost competitiveness simultaneously. Large-scale battery plus storage is now better than natural gas in many areas. Advanced geothermal is becoming super interesting but hasn’t been properly scaled or demonstrated yet. Small modular reactors are just coming online and are probably the next big thing once we can scale them up.

In the near term, natural gas seems like the obvious solution. However, it will likely become an obsolete technology within about 10 years. That’s the core problem we’re trying to grapple with.

Jordan Schneider: I was disappointed about dams. I thought they could be a viable option, but you disabused me of that notion. They’re too slow or not big enough. There’s no new innovative dam technology that’s been developed over the past 70 years, which I was disappointed to discover.

Arnab, what are you excited about?

Arnab Datta: I’m incredibly excited about next-generation geothermal energy. This is energy produced from heat in the Earth’s crust — the heat beneath our feet, as it’s often called. We’re pioneering this energy technology because of our experience with fracking and how we developed drilling techniques that led to the shale revolution.

There hasn’t been enough demonstration yet, though some companies are innovating quite rapidly. The scale potential is remarkable, and it’s an area where the US can really lead because we have an oil and gas workforce where 61% of workers have skills directly transferable to geothermal. We have a supply chain for fracking and shale production that’s ready to go and transferable to next-gen geothermal. The potential is incredibly high, and I wish we were doing more to support it.

Jordan Schneider: Can we stay on the technology for a moment? How do I drill a hole and get electricity out of it?

Arnab Datta: Basically, you’re drilling into the Earth’s crust where there’s substantial heat, and you’re pumping fluids down into that heat. The fluid gets heated up, and you’re circulating it back to a steam turbine that generates electricity. That’s the simple explanation.

Jordan Schneider: So it’s just like a steam boiler with the Earth’s core as the power source? That’s incredible.

Arnab Datta: Yes. There are multiple types of systems. Traditional geothermal requires three elements coming together naturally: heat, fluid, and a reservoir. These are natural geothermal reservoirs.

What next-generation geothermal does is create artificial reservoirs. You’re digging and fracking to create cleavages in the crust, and then cycling fluid through it. This is safe, to be clear. It has been tested and demonstrated. This isn’t something that’s going to damage the Earth. That’s the basic explanation for how it works.

Tim Fist: The amount of heat energy stored in the Earth’s crust that you can access via enhanced geothermal vastly exceeds the amount of energy in all known fossil fuels by several orders of magnitude. This is an abundant source of low-carbon energy without any of the intermittency problems of solar and wind that you can also access using many of the same tools we’ve developed for large-scale fracking.

This technology has already been deployed. Google is powering some fraction of its data centers with this. At this point, it’s primarily a scaling problem.

The areas where you can extract the most heat from the Earth’s crust using these methods also overlap substantially with the areas where you have federal land that can be readily leased. It’s a perfect recipe for solving this problem.

Jordan Schneider: The Earth is warmer under Nevada?

Arnab Datta: The heat is closer to the surface. To add one thing to what Tim said as an example: you have to drill three wells to produce about 10 megawatts of energy in something called a triplet. To reach five gigawatts, which is our goal by 2030, you would need to drill 500 of those triplets.

Jordan Schneider: This is child’s play.

Arnab Datta: That means 1,500 wells. We have drilled 1,500 new wells multiple times in the shale region of this country in a given year. This is not that many new wells to drill, if we can perfect the technology. We’re very well-positioned to take advantage of this, if we can get there.

Ben Della Rocca: I would underscore that geothermal is the single energy source I’m most excited about, in terms of technologies that are underrated by the broader public.

AI itself provides an opportunity to create much of the backstop demand that can funnel capital to the industry and incentivize development and technical advances needed to make the United States a global leader in this technology and advance our energy leadership more broadly.

As Tim mentioned, traditional geothermal resources are primarily available in the western United States. This overlaps heavily with places where large amounts of land are owned by the Bureau of Land Management. One of the sources of delay with building geothermal projects on federal lands has been federal environmental permitting reviews, which take time.

The executive order has directed the Department of Interior to find ways of conducting these reviews much more quickly — eliminating redundant reviews at multiple stages in geothermal projects and creating what are called “priority geothermal zones.” These are areas where the Department of Interior will focus its permitting efforts to move the process along as expeditiously as possible.

Ideally, these zones would overlap with places where AI data centers are being built, to ensure that all efforts are moving in the same direction. There’s a lot more to be done, but we’ve seen a valuable starting point to accelerate development in the geothermal space.

Jordan Schneider: What is the environmental consideration? We don’t even have oil gushing out. Are there endangered species in rock 20,000 feet below the Earth’s surface? It’s just ten guys and a drill.

Ben Della Rocca: That’s a great question. Certainly, the environmental repercussions are fewer than with traditional fracking for the oil and gas sector. With any construction project like this, you have to build a power plant, which involves some change to the natural environment. If there’s an endangered species right where you want to build the power plant, that will be a factor in the environmental analysis. Drilling deep down can potentially cause some impacts to the broader region as well.

The environmental burdens are significantly less, which is why there’s potential for the permitting to go more quickly. It’s a question of marshaling the right policy resources to ensure we’re all moving as quickly as possible given the lesser concerns with this technology.

The Qingshui geothermal power plant in Yilan, Taiwan. There’s a nearby park where visitors can hard boil eggs in the geothermal spring water. Source.

Jordan Schneider: What are some lessons from the shale revolution that can potentially apply to the US government helping incentivize the development and production of geothermal?

Arnab Datta: In the 1970s, coming out of the Arab oil crisis, we made a conscious effort to support the non-conventional production of energy. By the mid-2010s, we were the leading producer of oil and natural gas.

How did that happen? There were four key policy interventions that occurred over those decades that I would emphasize. My colleague at Employ America, Skanda Amarnath, and I wrote about this last year.

First, there were numerous research and development and cost-share programs to innovate in drilling and develop new techniques. The Department of Energy worked directly with Mitchell Energy, sharing some of the drilling costs to test non-conventional means of production.

Second, there were supply-side production tax incentives and demand-side price support. On the supply side, there was a Section 29 tax credit — essentially a production tax credit for non-conventional sources. An analogous current example is the Inflation Reduction Act, which supports production for new types of energy. It’s important that those credits remain in place.

On the price support side, there was targeted deregulation in the Natural Gas Policy Act in 1978 that exempted energy produced from non-conventional sources from the existing system of price controls. This essentially created a price support incentive. People describe it as good deregulation, but its ultimate purpose was to create a more competitive price environment for this type of production.

Third, there were permitting changes that altered the regulatory environment. The Energy Policy Act of 2005 established a legislative categorical exclusion, meaning certain types of production with specific geographic footprints could undergo the lowest level of NEPA analysis and get approved more quickly.

Fourth, and often underrated, was a highly accommodative macroeconomic environment. The shale boom and major productivity increases happened at scale in the late 2000s and early 2010s when interest rates were low. Companies could take on cheap debt and iterate, with abundant capital available for them to enhance productivity to the point where production levels increased even as the workforce declined because drilling techniques became so efficient.

These four factors help explain what happened in the shale revolution. We need to figure out how to compress that timeline to just a couple of years for next-generation geothermal energy.

The Exxon Research and Engineering Company (ERE) stage-gate system, used at Exxon to orchestrate fundamental research for new industrial technologies. Source.

Jordan Schneider: How far away are we today from these awesome steam boilers?

Arnab Datta: Fervo is currently building what I believe is a 400-megawatt facility. I don’t know the exact state of where that stands in project development since it’s a private company and I’m relying on public information. They’ve demonstrated that their technology works at a small scale.

Tim mentioned the Google facility. One of their smaller installations is powering a data center at around 40 megawatts. There’s also a company based out of my home city, Calgary, Canada, called Eavor that’s working with horizontal drilling. They have a small demonstration project as well.

The real question is whether we can achieve scale. The major challenge for these companies is securing enough capital to demonstrate that the technology works and can produce utility-scale electricity.

Jordan Schneider: You mentioned capital. Where hasn’t this been coming from, and where should it come from to realize this vision for geothermal over the next five years?

Arnab Datta: That’s a great question. In our report, we discussed the challenge of financing these next-generation technologies — whether geothermal, small modular reactors, or others. The fundamental challenge is that despite the promise, there’s tremendous uncertainty associated with developing these technologies.

The key is finding investors willing to accept this level of uncertainty in project development. They need to be comfortable covering costs when they increase because permits take longer than expected, supply chain issues arise, or interest rates climb, making capital more expensive. All these uncertainties accumulate, making equity investors reticent to invest at a sufficient scale.

There are only a handful of venture capital firms that engage in this type of investing, and they reach their limits quickly. Banks won’t do it because the risk of failure amid such uncertainty is too high. The government is playing a role — if you look at the Office of Clean Energy Demonstrations, they’ve funded demonstration projects for SMRs.

In the shale context, the federal government shared costs with Mitchell Energy for drilling operations. We need some version of that approach now. We also need the federal government to reduce uncertainty.

When you think about technology demonstration, you typically start with a concept paper, hoping to attract investment for the next step — a small-scale project. Once you secure that investment, you aim for a larger-scale project, attracting a bit more funding. It’s a slow, incremental process of building investor confidence. We need to compress this timeline at each stage and reduce uncertainty so companies can invest.

Financing energy projects typically requires three elements — debt, equity, and offtake agreements (meaning someone to purchase the energy once you’ve produced it). Currently, we see headlines about AI companies “investing” in energy projects, but they’re mostly doing this through these power purchasing agreements.

Subscribe now

With next-generation energy, there’s substantial uncertainty from factors that aren’t easily quantifiable — permitting and regulatory timelines, physical feasibility, and potential material bottlenecks. One of those three participants needs to own that uncertainty, and it’s generally not going to be through offtake agreements.

You would need very high-premium offtake agreements to cover a level of uncertainty that would give debt or equity investors confidence that their investment won’t fail. In our report, we emphasize that if the federal government can either own that uncertainty by providing capital or reduce it through streamlined regulatory procedures, that could unlock the tremendous amount of capital these companies possess, enabling them to invest directly in projects upfront.

Currently, there simply isn’t enough upfront capital, and that’s the barrier we’re trying to overcome. Many tech companies are sitting on substantial cash reserves, but they’re not directing it toward upfront energy investment. We believe that through federal government initiatives that reduce or assume this uncertainty, we might encourage companies to invest capital directly.

I should note that Amazon has invested directly in an SMR project in the Pacific Northwest, and that’s a model we’d like to see replicated more broadly and at a larger scale. That’s what we’re trying to accomplish.

Jordan Schneider: You gave me some great new acronyms, Arnab. We have FOAK, SOAK, THOAK, and NOAK — first of a kind, second of a kind, third of a kind, and Nth of a kind. We’re still in the “first of its kind” universe. Let me push back on this hype train a little bit. Three years we’re going to go from some cute demonstration projects? I take your point that if you can figure out the pumping mechanism, you have a lot of people happy to live in Nevada for a while and drill big holes in the earth. But AGI’s coming soon. Is this really going to get us there?

Tim Fist: This underscores the validity of an all-of-the-above energy approach, where we’ll want to take multiple shots on goal. We don’t want to put all our eggs in the geothermal basket. If you think about it, we had this staging approach for different technologies that could work really well under this all-of-the-above strategy.

You deploy natural gas plants because you know you can bring those online. They’re going to provide secure, reliable energy, and we know how to build them within a couple of years. Solar and battery storage is a really promising option to build out as much as we can where we can get it online.

Building out geothermal should come alongside that. Small modular reactors would come a bit after that as well. If you want to start thinking about fusion, maybe you’ll bring that online in 20 years. Basically, you want to invest in as many of these technologies as you can at once to address the inherent technological risk with this next-generation stuff.

Ben Della Rocca: One additional thing I’d say to your question, Jordan, is that even within the geothermal space, we can talk about this layered or staged approach where we lean more on some geothermal technologies at one point and then move on to others.

Traditional geothermal technologies are more tried and tested. We’re not necessarily doing first-of-kind projects for some of those geothermal hydrothermal resource projects. Fewer potential gigawatts can be brought onto the grid for that technology, but that might still be enough if we push people to do the right exploration and resource confirmation as they’re building their data centers. That could still be a meaningful part of the solution by 2028.

Beyond 2028, that’s when it might be most realistic for some of the enhanced geothermal projects, which are still in the first-of-a-kind stage, to come online. There’s a bit of phasing that we can do that way.

Tim makes an excellent point that there are different strengths and weaknesses to each of these approaches. It’s unrealistic to think that any one energy source is going to be the sole answer to AI’s energy needs.

With solar and batteries in particular, combining those two resources can be a way to access firm power, and there are downsides to the ability of that to scale in some cases. But it’s also faster to build solar plants, and there’s already been work done to review the environmental impacts of solar developments on some of the Western land under government management. This could make construction of those projects proceed more quickly in certain cases and be an important part of the shorter-term solution.

People really should be looking at a wide range of different options and leveraging different site-specific opportunities.

Transmission and Permitting Reform

Jordan Schneider: Can we talk a little bit about transmission lines and transformers? Arnab, you mentioned that a lot of this stuff may just end up being off-grid where Google’s responsible for building the power right next to its new data center. To what extent does hanging lots of transmission lines over people’s farms or whatever actually matter for all this stuff?

Ben Della Rocca: Transmission is a big part of the equation. You could certainly imagine a world, as you said, Jordan, where all the power resources are co-located and we don’t need to transport any electricity from one site to another. That’s theoretically a solution, but it’s unlikely that when we’re talking about gigawatt-scale data centers, at least in the short term, that we’re going to find sites where you can truly get three to five gigawatts of co-located power. Not saying it’s theoretically impossible, but it would be unrealistic to assume there’s a world where we just don’t need transmission lines to solve this problem.

At a minimum, having transmission lines provides a number of other benefits. First, it puts a much larger range of resources in place. If you have a data center being built somewhere around a variety of different BLM lands that are amenable to different energy sources, you can tap into more of them if you can deliver power from offsite to nearby locations.

Even if you are building power generation on site, there are still many advantages to interconnecting that power to the electric grid. It provides stability benefits, removes some of the need to build a microgrid or other sorts of redundant electrical facilities on site, and mitigates some of the financial risk of your project. If you end up using less power than expected, you could resell it onto the grid. Transmission lines are going to be important no matter what.

Subscribe now

I’d like to highlight a couple of things that the executive order tried to set in motion that could help us be forward-leaning on building the transmission infrastructure needed for AI data centers going forward. The most important is that the Department of Energy has some very important statutory authorities to address these problems.

One relatively well-known authority worth mentioning is the ability to establish National Interest Electric Transmission Corridors (NIETC), which are areas the Department of Energy can designate to play a backstop role in accelerating certain permits and approvals if they’re taking a very long time in a way that impedes efficient transmission development. This could be useful in the longer term, though it does take a meaningful period to actually establish a particular region as a NIETC and activate those authorities.

Another set of less well-known authorities that should be fully explored are those allowing the Department of Energy to partner with transmission line developers in powerful ways. Several statutes — such as the Energy Policy Act of 2005, provisions in the American Recovery and Reinvestment Act, and the Infrastructure Investment and Jobs Act — essentially let the Department of Energy create public-private partnerships or work with companies to participate in upgrading, constructing, planning, and financing transmission lines.

Based on analysis that the Department of Energy published under the Obama administration, in some circumstances, these authorities might be used to essentially bypass the need for lengthy state approval processes and allow the Department to play a more efficient role in cost allocation and other processes that state public utility commissions typically handle, which can often take many years to complete.

By using these authorities more aggressively, there’s actually a much faster pathway to building transmission lines, particularly relatively shorter ones. We’re talking about dozens of miles to connect a data center to the grid rather than hundreds of miles of interstate transmission lines. We’re not talking about giant transmission projects, but more targeted transmission builds that DOE can develop with the private sector. This could be a really important pathway to getting more gigawatts on the grid by 2028.

Arnab Datta: I would second basically everything Ben said. Transmission is really important. Some of the short-term solutions Ben identified are exciting, and it would be great to see them implemented.

The reason firms are moving off the grid, making that such a significant factor, is because transmission is just so difficult right now. This underscores the need for longer-term reform and broader permitting reform, with Congress actually taking action. There’s only so much you can do through the executive branch, and it tends to be more imperfect. We need to fix that.

Tim Fist: It takes, on average, 10 years at present to build a new transmission line in the United States. That’s mostly due to these holdups in permitting.

Ten years ago, we built 4,000 new miles of transmission lines every year. Now, it’s more like 500 miles. This is decreasing by a factor of eight.

Jordan Schneider: Ben, what’s a transformer and why does it matter?

Ben Della Rocca: Conceptually, a transformer changes the voltage of an electric current. This transformation process is essential to bring electricity from transmission lines or higher voltage environments to voltages that an actual end-user facility can accept. We need transformers within electric infrastructure to make the power usable for its intended purposes.

The challenge with the transformer industry is that there’s limited capacity with current resources allocated to transformer development to supply an adequate number of transformers needed to build all the power infrastructure that AI is demanding.

Much can be done to support the transformer industry through loan guarantees or other financing options provided or encouraged by government. These measures would allow the transformer industry to invest in the capital expenditures needed to expand their facilities or train new workforce and add workers to existing facilities. All these steps are important for reducing transformer lead times, which I believe are currently in the two-and-a-half to three-year range. Transformers are certainly an important part of the supply chain aspect of this problem.

Jordan Schneider: Since we were talking about the Defense Production Act earlier, now that corruption is in and FARA is dead, how far can a president who really just wants to potentially let their main consigliere who happens to be building giant AI... Just get all the gas turbines before everyone else? Is there any recourse here for something crazy like that?

Arnab Datta: Broadly with DPA use, when utilizing any aggressive or assertive legal authority, it’s important to try to get political buy-in, even if you believe the legal authority is bulletproof and you can do what you want. You still want political buy-in.

The fact that Ben’s here discussing how the Biden administration prioritized this because they saw it as a threat, and the fact that the Trump administration thinks this is a threat and that there should be a national security aspect to AI data center build-out shows there is some consensus.

The DPA is up for reauthorization, which typically happens in a bipartisan fashion. This presents an opportunity to take advantage of that consensus and say, “We are appropriating X amount of dollars and authorities for the DPA to be utilized to help our energy infrastructure build-out for AI data centers,” and put some safeguards on it. If Democrats are concerned about the hypothetical scenario you just described, that can be a negotiating chip for what is typically a bipartisan reauthorization.

Generally, if your concern is around corruption with this issue, there is an opportunity here because the DPA is up for reauthorization. There were actions during the Biden administration regarding using the DPA for heat pumps that Republicans didn’t like, and there were hearings last year on this. Building some kind of legislative consensus could be useful here.

Ben Della Rocca: Arnab has made many great points, and I’d add that in this issue space, the role of litigation shouldn’t be underappreciated in terms of the type of check it can play. This has long been the dynamic with infrastructure projects in many different industries.

The way litigation works around the National Environmental Policy Act and other permitting-related statutes is that lawsuits can allege that permitting requirements haven’t been fully fulfilled. This can result in injunctions that ultimately delay projects while court proceedings are ongoing.

In effect, this means there’s significant value in making sure all the T’s are crossed and I’s are dotted when pursuing an option that uses national security authorities. If you do something that goes outside the bounds of the law or isn’t an airtight legal case, the odds of litigation increase, which can ultimately result in projects being delayed.

If you can’t get past the pre-construction stage because you’re dealing with extensive litigation, that can have huge consequences for the timeline of building artificial intelligence infrastructure, and time has a real premium in this space. The need to ensure that laws are followed very closely here shouldn’t be understated, for a wide variety of reasons.

Jordan Schneider: A lot of the dynamics you both just pointed out also apply to all the NEPA stuff we were discussing. Ben was doing some clever things here and there to try to make it easier for firms, and then Trump just cancelled NEPA, which is legally questionable. We’ll find out. On one hand, it’s probably exciting for Google and Amazon. On the other, you’re opening yourself up to a whole new legal attack surface that wouldn’t exist if we were living in a Harris administration that followed the direction the executive order laid out more directly. Anything else to add on that dimension?

Ben Della Rocca: Your point about the uncertainty here is exactly right. Trump’s rollback of NEPA as it has existed for years certainly has the potential to speed things along, but it doesn’t ultimately eliminate the fundamental statutory requirement, which is for agencies to essentially do their best to review the environmental consequences of their actions.

Without clear regulation, there’s going to be significant ambiguity and uncertainty about what that means, and there will still be years of past practice that courts may look to when determining the exact content of the statutory requirement. The actual magnitude of the impact from efforts to rewrite NEPA regulations remains to be seen. It may take years to play out.

Arnab Datta: One quick thing to add is that it’s important to note that while the NEPA regulation passed by the Council on Environmental Quality was rescinded, every agency still has its rule in place for conducting a NEPA analysis, and those remain in effect. Right now, that’s the rule of the road. There’s a long-term uncertainty that Ben is right to discuss here, but in this immediate moment, the regulatory framework still exists for agencies, and that’s important for people to know and continue to comply with.

Attaching Strings

Jordan Schneider: Tim, you wanted to add a requirement to make these hyperscalers take AI security more seriously before they get access to government financial help. What is the market failure here, and what are the sorts of things you think the government should add to their requirements in order to get all of these special dispensations?

Tim Fist: US AI companies are currently building models that they believe could, within just a few years, reshape the global balance of economic and military power. Consider AI systems that can autonomously carry out massive cyber attacks, automate scientific R&D processes, or serve as substitute remote workers for many kinds of jobs. If this is true, we really need to protect these systems against theft by bad actors.

It turns out that many of the security problems you need to solve are at the data center level. Protecting against sophisticated threat actors like nation-state hacking groups is both extremely difficult and expensive. If a company invests adequately in security, they risk falling behind competitors who aren’t making similar investments.

A core part of the executive order ties assistance around loans and permitting to strong security requirements. This creates a set of requirements that hyperscalers and AI companies can follow to raise the level of security protecting their critical intellectual property against these threats. By connecting it to this assistance, you transform something that would put you at a disadvantage relative to competitors into a strong commercial decision.

We outline several ideas for what this could look like. Specifically, it means finding the best existing standards and applying them across the board, developing new standards and guidance specific to the threat model for attacks on AI model weights, and providing government assistance for supply chain security, physical security for AI accelerators, background screening for personnel to protect against insider threats, and counterintelligence playbooks.

The basic idea is to create a strategic partnership between the government and the AI industry to improve security, with incentives on the other end to make it worthwhile.

Jordan Schneider: The piece that has really struck me when reading Dario’s article about his world in which America gets ahead and accelerates toward AGI faster than China, then the incentives for the Chinese government to leave these data centers alone falls to basically zero.

If you can steal the model weights, then maybe you want OpenAI and Anthropic to continue existing to make cool stuff that you can take and deploy. But if this is the technology which is to rule all technologies, you start to get into a US-Iran 2000s/2010s dynamic where stuff like Stuxnet or drone attacks become a concern. There’s water for cooling everywhere — what if there’s just a giant leak that fries all your servers?

The physical security of these tens or hundreds of billions of dollars being thrown into data centers hasn’t really been discussed much. You can see the potential future where that ends up being a critical part of what the US government and the firms themselves need to focus on for safety.

Tim Fist: I’m more optimistic about protecting models than you are, with the caveat that we need to think about the scope of things that it’s useful to protect.

To be more specific, I expect that over the next few years, the most powerful models developed by US frontier labs are going to be deployed internally first. There are three main reasons for this:

First, as capabilities grow, there will be numerous misuse concerns that labs will want to address before wide deployment.

Second, deploying internally before wider release makes a lot of technical and economic sense as you can use the model to accelerate your own R&D before releasing it more broadly.

Third, it makes sense to first train the big expensive model and then distill it down to a version that’s more economical to serve to users. This is reportedly now common practice across basically all the frontier labs.

If this is true, then protecting cutting-edge models can be done in a more favorable security environment where your attack surface is relatively smaller because you’re initially only deploying for internal use cases.

Eventually these models will likely get stolen, but protecting the bleeding edge from immediate theft is still worthwhile as it allows you to use those models to maintain your overall lead by investing your inference compute into AI research and development, and using those models to develop things like AI-powered cyber defense.

There’s some hand-waving in this theory of victory, and there are many unknowns, but seriously trying to predict this is worth it. The alternative is freely handing it over to China — putting all this money into power and chips and then giving the products freely to China to accelerate their own AI research programs.

Preventing denial of service or sabotage operations is also a worthwhile goal. I’ve seen interesting research recently about the susceptibility of current AI data centers to cheap drone strikes as well as attacks on surrounding network and energy infrastructure. I don’t have a view about how expensive this will be to defend against, but it certainly needs to be a huge part of the investments in defense.

Jordan Schneider: One thing I’ll say is that if America is going to win, it’s going to need PRC nationals working in these labs. If we’re doing FBI counterintelligence checks on every AI PhD Berkeley graduate — I’m sorry, we’re just not going to have an AI ecosystem. There’s some middle ground there, but that’s the one piece I was most skeptical of.

I had one random question. There was this very funny chart that Tim and Arnab had where Google, Amazon, and Microsoft all committed to being net zero by 2030, and they’re on this trend line. Then it just starts to go the wrong way once they realize they have to build tens of billions of dollars of data centers.

Do those commitments just go away? In our anti-DEI world, is there anything statutory about it? Is Blackstone going to get mad at them? What’s the forcing function here that would keep them on those trend lines, absent some really amazing geothermal breakthrough?

Arnab Datta: I wrote about this recently. The way to get more adoption of these newer technologies that are firm and emissions-free is for them to become cost competitive and quick to deploy. I don’t know how firm the commitments are from Amazon and Google or how sticky their internal social costs of carbon are.

I’m trying to think about policymakers and what we can do to get to that place — reduce those costs. These commitments are real, but they’re probably not going to stop a company from even putting a coal plant online if they know they can get AGI first. If you care about climate change and decarbonizing, our job is to figure out how to make that happen as fast as possible.

Ben Della Rocca: I agree that making these energy sources affordable is the best way to ensure they’re adopted. The related piece is making sure that the timeline to actually permit them and bring them online is efficient and fast as well.

In some cases, if clean energy technologies or emerging clean energy technologies can be brought online more quickly than other, less clean sources — if the permitting timelines are actually shorter for those technologies — that can provide strong incentives for industries such as AI to choose the faster route. There’s a large financial premium they could earn from bringing their AI models online and operational six to twelve months earlier.

Jordan Schneider: Any final thoughts, Ben?

Ben Della Rocca: There was a lot of work in the last administration to set forth actions that will address AI’s energy needs, and we’ve discussed many potential ways forward in this conversation.

A linchpin to making all of this successful is effective implementation and really prioritizing this work within federal agencies. Ensuring that people are focused on completing this work effectively, fully, and quickly, and making sure that the work starts on time and proceeds according to schedule is going to be extremely important.

Tim and Arnab, in your paper, one of your recommendations was for the White House to have an AI infrastructure czar of sorts to oversee and spearhead this work. This work is ultimately very complex and interdisciplinary. It’s not just a national security challenge — it also includes energy policy and law, and environmental permitting law. It will require strong leadership from the White House and the federal government to make sure that things happen as envisioned.

To underscore a simple but important point — the implementation side of this really matters.

ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.

Mood Music

Best of the Month

14 March 2025 at 19:05

We’re doing a lot more stuff than we used to! To make sure you don’t miss out on our best content, we’ve decided to start publishing monthly roundups.

Chips and AI

**China’s Weird Chip Surplus, Explained

This article examines China’s paradoxical state of AI compute, where reports of both shortages and overcapacity exist simultaneously. China acquired and produced over a million AI chips in 2024, but faces inefficiencies in their deployment. At the same time, a slowdown in foundation-model training has reduced immediate demand, creating a temporary surplus that is expected to reverse as inference demand rises and China pursues the next generation of AI models. The article synthesizes insights from Chinese media, outlining how state-owned telecoms, tech giants, and local governments are investing in compute, the challenges of building true high-performance clusters, and the Chinese government’s recent moves to curb waste and consolidate AI infrastructure. It also assesses the role of Huawei’s domestically produced chips.

DeepSeek: what it means and what happens next

This interview with Kevin Xu, a former GitHub and Obama Press Office staffer, explores DeepSeek’s unique organizational structure, approach to talent, its parallels with OpenAI, and the broader tensions between open-source collaboration and nationalistic technological competition.

Anthropic's Dario Amodei on AI Competition

We interviewed the CEO of Anthropic! In this podcast, Dario discusses his essay “Machines of Loving Grace,” and advocates for stricter export controls to widen the US-China capability gap and ensure AI systems are developed responsibly. The conversation covers DeepSeek’s unexpected advancements, concerns over AI espionage and model distillation, and the technical and policy challenges of preventing China from scaling up to US levels. Amodei also reflects on AI’s potential impact on democracy, suggesting that AI can be used to strengthen governance and public deliberation rather than simply entrenching autocracies.

DeepSeek’s New Burdens: Does DeepSeek need Beijing, or does it need Beijing to stay out of the way?

With the help of an anonymous contributor, this article speculates on the challenges DeepSeek faces as it attracts attention internationally and from the Chinese government. With growing compute needs, we consider whether DeepSeek could partner with a hyperscaler like ByteDance, and how that would shift its culture and business model. Since then, new reports have emerged verifying that DeepSeek is considering such a partnership as well as weighing the benefits of outside fundraising. This essay also considers the extreme case where DeepSeek receives “national champion” status, gaining state funding, privileged data access, and procurement deals — but at the risk of government interference, loss of autonomy, and geopolitical backlash.

SemiAnalysis + Asianometry on the AI Mandate of Heaven

In this podcast, Jordan has way too much fun making an AI tier list, speculating on Tim Cook’s successor, and discussing Google’s $75 billion AI investment with Dylan Patel and Doug O’Laughlin from SemiAnalysis and Jon from Asianometry.

Watch below on YouTube or listen now on iTunes, Spotify, or your favorite podcast app.

China

Friday Bites: China on Dario, DeepSeeking Truth, Ali + DeepSeek, and a Procurement IP Manifesto

This grab bag features translated Chinese responses to our interview with Dario, as well as Alex Colville’s excellent report about how DeepSeek subtly propagandizes even when the model is theoretically uncensored. Other features include how to use DeepSeek to research China, rumors of hyperscaler partnerships, and an anonymous procurement reform call to arms.

Xi’s Hard Tech Avengers

This article guides you through the less-well known guests who attended Xi’s business symposium, with profiles of the propaganda czars, Politburo tech strategists, and neoauthoritarian academics as well as tycoons of semiconductors, chemicals, agriculture, and 3D printing.

US Policy

**Strategic Ambiguity vs. Clarity

We took a break from DeepSeek mania to publish Republic of ChinaTalk editor Nicholas Welch’s comprehensive guide to the debate on Taiwan policy. It condenses arguments from 50 op-eds and academic papers into 12 key points, with five supporting strategic clarity (arguing it deters China, reassures allies, and aligns policy with reality) and seven advocating for continued ambiguity (which proponents say maintains flexibility, prevents provocation, and keeps Taiwan focused on self-defense). The article highlights areas where more data is needed to evaluate these arguments, and highlights historical parallels to past U.S. foreign policy missteps.

American Power in the Age of Economic Warfare

In this interview, Eddie Fishman explores the evolution of American economic warfare and the effectiveness of sanctions as a geopolitical tool. Fishman, a former State Department official, traces the transformation of U.S. sanctions policy, from the failures in Iraq and Cuba to the unprecedented success of Iran’s financial isolation and the mixed results against Russia. He highlights how globalization turned the U.S. dollar into a powerful chokepoint, leveraged through secondary sanctions and financial penalties to pressure adversaries. The discussion also covers the institutional dysfunctions that hinder swift action, the challenges of enforcing export controls on China, and the need for a centralized agency to handle economic statecraft.

The NSF, Seriously? + AI Safety's Death

In this piece, Jordan critiques the Trump administration’s cuts to basic science funding, particularly its impact on the NSF’s Technology, Innovation, and Partnerships (TIP) directorate, which was designed to drive translational research and bolster U.S. competitiveness. Jordan argues that gutting forward-thinking programs like TIP is a strategic mistake that undermines national security, and suggests that meaningful budget cuts should target bloated defense programs instead.

Trade: How Free is Too Free?

We released a transcript of our 2019 podcast with trade historian Doug Irwin in light of Trump’s tariffs on China, Canada, and Mexico. This interview explores the long history of U.S. trade policy, tracing its evolution from the revenue-focused tariffs of the early republic to the restrictive protectionism of the late 19th and early 20th centuries, followed by the reciprocal trade agreements that defined postwar globalization. The conversation also examines the rise of executive power in trade policy, the role of institutions like the WTO, and the shifting landscape of U.S.-China economic relations.

Gridlocked: Transformer Shortage Choking US Supply Chains

Lily Ottinger and Caleb Harding reported about the transformer shortage in the U.S. and its implications for industrial policy, national security, and economic growth. Transformers, essential for electricity distribution and infrastructure projects, face soaring demand due to grid aging, electrification trends, and AI-driven energy needs, yet domestic supply meets only 20% of demand. Supply chain disruptions, labor shortages, lack of standardization, and reliance on foreign manufacturers — particularly China — exacerbate delays, raising concerns over security vulnerabilities from potential cyber or physical attacks. The report explores solutions such as expanding domestic steel production, incentivizing transformer manufacturing, standardizing designs, and aggregating a strategic transformer reserve.

Innovation Emergency with Trump 1.0's Patent Director

This interview with former USPTO Director Andrei Iancu explores the role of patents in fostering innovation, the increasing dominance of China’s IP system, and the future of U.S. patent policy under the Trump administration. Iancu highlights how congressional inaction causes outdated IP laws to struggle in the face of emerging technologies, while China’s streamlined patent process and aggressive IP strategy have given it a competitive edge. He warns that weak patent protections in the U.S. risk stifling investment in high-risk, high-reward industries, potentially ceding technological leadership to China.

In the trades…

In the wake of DeepSeek mania, ChinaTalk earned shoutouts in the The FT, NY Magazine, Al Jazeera, and Foreign Policy (which called our coverage ‘savvy’). Jordan also appeared on NPR’s Here & Now, The Circuit and Hard Fork to talk about China’s AI development.

ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.

Mood Music

Humanoid Robots: The Long Road Ahead

13 March 2025 at 19:00

Today we’re doing an anonymous Q&A with KL Divergence, a robotics PhD currently in industry working on humanoid robots. For an introduction to humanoid robotics in China, see our article here, and for a deeper look into who’s leading China’s humanoid market, see our latest translated interview with the CEO of Unitree Robotics.

This Q&A covers:

  • When and how AI-driven robotics will reach a tipping point in viability,

  • Challenges and solutions for collecting data to build a robotics AI model,

  • Successful strategies for companies to compete in humanoid robotics.

When will AI + robotics reach a tipping point in viability?

This is extremely difficult to predict.

Here's my non-answer: whenever the world achieves a data flywheel for robotics, i.e. accumulate a dataset large-enough (and algorithms to use it) that allows some robots to achieve a diverse set of somewhat useful tasks, with enough reliability that people allow those robots to operate in their factories, logistics centers, homes, offices, etc.

Once a robot has a “reason for being” in a space, and works well enough, the data flywheel will spin, and the robots will get better and better. This is the same process we are seeing play out in self-driving cars, and why Waymo’s early advantage in deployment is such a big deal, making Tesla dance and move forward plans for Robotaxi. I don't think we will see this happen all at once, in all application domains.

Today, this has already arguably happened for a specific application: robotic picking/packing in e-commerce logistics. Amazon, Dexterity, Covariant, Berkshire Grey, and Ocado all have massive robotics datasets for this basic task, and already use them to create their own “flywheel.” This is short of what we want though, because that data is only for one task, and is specific to those companies' unique robots.

What’s the pathway to viability? How will AI + robotics diffuse through different industries?

So the next stage will likely be doing lots of different tasks (10s-100s) in a structured environment. I would guess logistics centers and manufacturing. I think this could reasonably be achieved in 2-3 years research-wise, and 5-7 years to become commercially commonplace. Along the tail end of that period, you might see these robots start to appear in retail, hospitality, and food service back-of-house. Think: robots doing laundry or restocking shelves. Next, offices. And last, homes.

Will we be seeing AI-driven robots in homes?

We’re 10+ years away even as a question of research, if we ever get there.

Homes (and to some extent offices) are much more difficult than commercial/industrial spaces because of 4 factors: lack of structure and wide variation, safety, and cost.

  • Structure and variation: Homes are the ultimate “unstructured” environment. They come in infinite variations, and change from moment to moment as people and stuff move around. One day you might decide to put the cucumbers in the top vegetable drawer, then next you might move them into the bottom drawer. Multiply that by everything a home robot might have to ever interact with and the amount of variation becomes mind-boggling. It is impossible to create a system which quantifies and anticipates it all explicitly. The realization has been the impetus for the move towards learned — rather than programmed — robotics AI systems over the last 10 years.

  • Safety: It’s an engineering achievement to make a robot that can complete tasks and weigh only as much as a smallish human. If that thing falls in a house (dead battery, malfunction, etc.), the stakes are high: it might fall on a pet, break a glass table, or knock over a candle. Contrast with a controlled commercial environment, where people working near the robots can be specifically trained, the environment is arranged so that failures don’t lead to catastrophic danger, and the robots might even be cordoned off behind a cage to minimize the impact of accidents.

  • Cost: Most proposals for robots in the home have them providing typical domestic labor: cooking, cleaning, tidying up, etc. People already pay for these services in their homes, and it is invariably some of the lowest-paid work in any economy. A humanoid robot has a similar part count and manufacturing complexity to an electric car, so it’s intuitive that the most optimistic cost estimates land the price of these machines at similar numbers: $15-50k, depending on the source. How much would a home robot have to do for a family to justify a $25k price tag with a 5 year life span, assuming no recurring service or subscription costs?

So why do we see some players in the robotics/AI space, humanoid or otherwise, proudly touting their goal of putting robots in homes? My best guess is that it’s a more compelling narrative to attract investment and inspire talent — and it’s not too hard to pivot back to industrial robots anyway.

What are the challenges of getting good training/testing data for AI-driven robots?

You already answered this well in your piece. It's because we need to get robots into the real world to collect lots of good data, but they currently don't work well, can be unsafe, take up space, etc. They have no economic reason to take up space and human attention where you would want them to collect data.

How are proposed solutions addressing these data challenges?

Ways of addressing:

  1. Simulation (but this has flaws, as you mentioned)

  2. Spend a lot of up-front capital to collect robot data directly, in hopes of collecting enough up-front to bootstrap a useful robotics foundation model (i.e. vision-language-action model or VLA)

  3. Find a way to re-use data from the internet, e.g. watching human cooking or furniture assembly videos from youtube (this is very active research, but so far the results have been disappointing),

  4. Master one task at a time (using a combination of 1 and 2, and old fashioned engineering), and hope you collect a dataset diverse-enough before you run out of money (if the dataset is not diverse in tasks, you will have a robot which can do a handful of tasks, but is expensive to train to do new tasks).

What's it like on the ground for a factory collecting this data?

This can vary greatly, but typically you have a task in a factory which has already been defined for a human worker, and it's fairly repetitive. E.g. Tesla's first application is to make Optimus, to grab battery cells rejected into a slide coming out of a battery quality control machine, slot many of them into a grid on a purpose-shaped tote, then walk those totes to a different area of the factory when full. It's very simple and repetitive, and today it's done by humans across dozens of machines. You can imagine other scenarios. For example, sorting packages into bins bound for different geographies in a logistics center. Well-defined tasks and lots of existing automation and built environment (e.g. screens, conveyor belts, well-placed bins and racks) to help humans.

What does it look like to collect the data? That depends on the approach. The most straightforward is teleoperation: a human dons a (usually) VR headset, special gloves for capturing finger movements, and other hardware you'd see in a VR space, and uses them to control the robot directly to do the job. This is “robot-in-the-loop.” It’s slow (the human can't move the robot as fast as they would themselves), and costly (it's actually more expensive than having a human do the job, because it's slower).

Another approach is motion capture: via various methods (camera systems in the work area, body-worn suits, even lightweight worn exoskeletons), we can capture the motion of humans who are already doing the job. This is more speculative, as it’s a difficult research problem to turn these motion recordings into instructions for the robot to achieve the task later.

The last major approach is simulation: usually through the help of a skilled human artist or engineer, create a detailed and functional 3D graphics simulation of the real environment in which the robot is supposed to perform. This allows us to use teleoperation, programmed routines, and reinforcement learning, to control the robot in simulation and collect data on its successes and failures. The weakness of this approach is that the model usually cannot be used immediately on the real robot, because it’s extremely difficult to capture all of the important behavior of a real work task, even a very small one, in a simulation. Roboticists refer to this problem as the simulation-to-reality (sim2real) gap.

On the research horizon, there are a variety of approaches that may allow us to generate or make use of data without actual or simulated robots. A “holy grail” of robot learning for the past decade or more, has been to create a robot learning system which can “learn from watching YouTube videos.” What all of these approaches have in common is that they seek to lower the cost of data for robotics models, by finding ways to make use of lower-quality data (i.e. weaker supervision). The key missing technical piece in most of these approaches is to find a way to map from non-robot behavior in one environment to the actions a robot would take to do the same task in a new environment.

What would indicate a successful humanoid robotics strategy?

How many robots does the competitor have in the real world doing tasks and collecting data, and (importantly) how diverse is that set of tasks? Humanoids are a very expensive way to automate just one thing, so the investment needs to be amortized across many different jobs.

Like bodyguard!

What strategies are robotics firms taking to compete in the market? What will determine who succeeds?

Boy, this is a big question. I won't try to answer the whole thing, but I'll give you a framework.

There are a few fundamental assets to look at here: technical talent (people), chips (compute), robots (how much do they cost? what are their capabilities?), data, and distribution (customer relationships, pilots). Any robotics+AI company or partnership effort needs to assemble all of these ingredients to be successful.

Resources which are less scarce:

  • Robots: you have a whole article about how China is commodifying robots. However not everyone agrees that *good* robots will be so plentiful (perhaps because of protectionism), and others (e.g. Figure, Boston Dynamics) believe they can create an edge by having the *best* hardware.

  • Customer relationships: Tech demo deals like the Figure-BMW, Apptronik-Mercedes, and Agility-Amazon partnerships are very low-risk for the larger company and easy to make. CEOs at humanoid companies tell me they have no problem getting 100s of leads.

So a successful strategist will try to gain an edge in the scarcest resources: talent, chips, robots, and data.

  • Chips: notably — every major NA effort has decided that they need to team up with a giant foundation model provider to have the chips and frontier models to compete. Figure-OpenAI, NVIDIA is in-house, Tesla-xAI, and Boston Dynamics has teamed up with TRI's foundation model team.

  • Robots: Most people attempting to make their own, however Skild, NVIDIA, and Physical Intelligence have all taken a partnering or purchasing approach for robots instead. Whether a competitor sees robots as a competitive advantage, or an expense, is a major dividing line in strategy in this area.

  • Talent: immensely cut-throat. Until very recently robotics+AI was a very niche field. An investor told me he believes there are ~25 people in the world who could lead one of these companies well. Even below leadership, the number of people with any training at all in this subfield is in the low 100s. The best-paying outfits in the world with the best reputations take 6 months to hire someone, and are often just waiting for new PhDs to graduate to fill positions. Talent with <1 year of professional experience but relevant education (usually a PhD) can fetch $500k-1M/yr in this field, and/or significant equity, depending on the size of the company.

And finally, data is the most strategic asset these companies seek to accumulate long-term. At the end of the day, the firm who has the best data (or best strategy for getting it) wins the game.

ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.

❌
❌