This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
There's a perception in the US that Americans are in a new Cold War, now with China. I'm curious about how they perceive their standing in it. Pulling ahead, like in Raegan's time? Sputnik Moment?
But mainly, I want to hear from Shakes. At the end of June 2026, in another discussion about America Winning the Iran War, I told @Shakes that I'll be returning to his post in which he claimed:
Some updates since then.
First, a recap: in December 2025, the Chinese private launch provider LandSpace attempted a launch and recovery of their methalox powered medium class launch vehicle Zhuque-3 ("Vermillion Bird"), which is very much like Falcon 9 if you don't look at the details (I think it's what Falcon would have been if it were designed in 2020s). The launch and orbital insertion went well, the recovery… not quite, which prompted these kinds of headlines and understandable complacency in some circles.
On July 10th 2026, China Academy of Launch Vehicle Technology (for reference, known as the "First Academy", founded by the exiled communist Qian Xuesen, the co-founder of Jet Propulsion Laboratory which then became the core of NASA) has successfully landed the first stage of their medium class launch vehicle Long March-10B on a sea platform, using a novel net capture mechanism, thus making China the second nation with the capability to reuse first stages of orbital class vehicles, and the only one whose national space agency can do that. There has been commentary to the effect that it's cope for the fact that China can't into precise landing, that you won't have the net capture infrastructure on Mars/Moon/whatever, and that this is a nothingburger and further testament to their backwardness. On August 19th ie today, LandSpace has done this in a more traditional SpaceX manner, with ZQ-3 Y-2 landing on the ground pad. LandSpace was founded in 2015, and ZQ-3 had only been announced in late 2023 – the same year they had put the first methalox rocket into orbit. This, hilariously, puts them technically ahead of SpaceX for the second time. LandSpace is working on a full flow staged combustion methalox engine too, and it's reasonably mature; technologically it's about on par with Raptor 2. Given the track record so far, it's reasonable to say they will have a Starship class vehicle within 5 years. There are multiple similar projects being executed in parallel, both private and state-owned, eg Long March 9. It seems implausible in the extreme that China will find it hard to scale up the production of rocket engines (after all, they do make more WS-15s than F135s now, judging by J-20 vs F35 commission rate), steel tubes (come on) or concrete launch pads (…come on, really), or land permits (lol). So I find it likely that in a fairly short order they can match SpaceX (and thus the US and the world, because SpaceX is a near-monopolist now) in annual mass to orbit if they so wish. With Blue Origin's recent disaster and anklebiters like Rocketlab not doing anything interesting, it appears inevitable that the space race is just SpaceX vs China.
On July 16th, the private startup Moonshot AI has unveiled Kimi K3, the 2.8T, 104B active multimodal MoE, with very innovative architecture and ability to execute on long-horizon self-improvement-related tasks such as chip design or ML research/engineering, delivering performance close to the best American public models, solidly exceeding the previous domestic champion GLM 5.2. Right now it scores 60 on Artificial Analysis, 1 point behind Grok 4.6, and 2-3 behind Opus 5, Fable 5 and GPT 5.6 Sol. (GLM 5.2 reached 53).
Also on Jul 16th, Chinese memory company CXMT completed its IPO subscription. Currently it's worth around $500B, underpinned by its central role in supplying DRAM chips to Chinese (and soon global) industry, including AI. They target 30% global market share by 2030. This is doable, given what I know about the velocity of upstream tool supply chain in China.
On July 18th, during the World Artificial Intelligence Conference in Shanghai, Huawei has demonstrated their Atlas 950 SuperPoD, boasting of the largest scale-up domain in systems for training advanced AI models (yes larger than anything Nvidia ships right now, and this is more important than raw FLOPS as we continue to increase the parameter count). You might be interested in this writeup on Huawei's design philosophy, it's pretty special and promises to compensate for their lack of advanced lithography. In short, their thesis is that the performance comes not so much from Moore's law as from minimization of latency across the entire architecture from transistor to cluster level, and their idea of a solution is 3D-native chip design for multi-layer logic, with very precise (1.5 µm currently, <1µm scheduled, well ahead of the competition) wafer-on-wafer stacking and the first chips demonstrating its viability (mobile Kirin SoCs) coming out in September. Multiple other companies, such as Alibaba, have also shown supernode-based designs, including one absolutely bonkers system from Oriental Computing that uses 14nm chips; as Jensen says, lithographic process is overrated compared to design, so I'm bullish on this line. At the opening ceremony, Xi Jinping delivered a pretty impressive speech on Chinese strategy with regard to AI, committing to support open source and international collaboration.
Since then we've learned of multiple 100K GPU class cluster projects in China (Sugon, completed, Alibaba token factory apparently as well, unclear what's up with Zhipu's 1GW cluster; DeepSeek will also have gigawatt-class systems in Ulanqab, Inner Mongolia, as well as their own chips). These should be sufficient to design and train 10T models, ie comparable to the alleged size of Mythos Preview (and larger than the deployed Mythos/Fable; I am not privy to these details, though). Ryan Fedasuyk of Georgetown estimates that «No matter how we slice the data, we find China is well on its way to producing large numbers of AI accelerators».
On Jul 31st, DeepSeek has deployed and open sourced V4-Flash-0731, getting performance around GLM 5.2 (and much higher on some hard evals like ARC-AGI-2) at a ludicrously low price and parameter count, doing even better than GPT 5.6 Luna after the much-hyped 80% price cut. The situation with DeepSeek is a bit ambiguous, it's not clear if they're flailing (the subsequent Pro was barely any better, their harness project is insanely ambitious but clearly not even half-done); but it speaks to the fact that the Chinese tech ecosystem is now very large and dense and nobody can be champion for long. On Aug 14th, Zhipu has responded to Kimi with GLM 5.3 which is on par with K3 at a fraction of the cost and scale; one of their priorities has been cyberdefense (and thus cyberoffense) capability, plainly driven by concerns around Mythos/Fable. They'll release the weights in <2 weeks, as did Kimi, as did Alibaba Qwen with their 2.4T 95B MoE that's roughly in the same ballpark. The CEO of Zhipu, Jie Tang, is a professor at Tsinghua, and thus essentially a state official, a CCP member who regularly contributes to People's Daily, so we can consider him speaking for the Chinese policy (Xi's speech has much the same tenor). His philosophy is roughly as follows:
As an aside, on Jun 29 it became known that Meituan (a food delivery company) had trained a 1.6T 50B MoE on previous generation Chinese chips, almost certainly Ascend 910Bs. It went under the radar because the model isn't that good, but it sets a lower bound on what can be done going forward, by more competent actors, with more advanced hardware. Today, the globally dominant Hangzhou humanoid/quadruped robot maker Unitree also went public, and is now worth about $50B, ahead of the fraudulent American company FigureAI with $39B and no publicly sold robots to show for it. They clearly have the best hardware at the moment and unmatched development velocity.
I could go on. It's been a rather eventful period. But these are, I think, the most interesting and strategically significant domains: AI, hardware for AI, hardware for manufacturing AI hardware, robotics, and space.
Shakes, do you think China can't catch up?
Your timing is perfect because I was having this argument this weekend with someone who claimed to be a mutual of both of us (a mutual mutual if you will).
Overall I have a lot of respect for the Chinese and their systems, I spent a lot of time in China in a former life (还能懂一点中文), and I think they might have the best food on Earth. China's progress is very serious and I've come to accept that almost nobody in the English language knows what they're talking about with China, so it's not usually even worth getting into. But I also felt, after my time in China, that the time I spent learning Mandarin was a "waste" in that China is not going to take over the world.
I still believe, fundamentally, that the 21st Century is an American century and I don't think it's really close.
My reasons get into some theories about industrial society, but let me knock out a few other points first:
In fact, this gets to the heart of a point I consider very important: Most Chinese and most Americans, also, are not participating in industrial sovereignty. It is not the case that America has an entrepreneurial culture and China has an engineering culture. I think it is more accurate to say that America has entrepreneurial cultures (sub-cultures) and China has engineering cultures. We can draw the lines however we like -- Shenzhen, Beijing, Texas, Silicon Valley, Shanghai, New York, Seoul, Tokyo, Taiwan, Arizona, etc. etc. But fundamentally the engineers inventing the future are not just expressions of national forces. They are particular individuals and factories and networks.
What I mean is that I view policies and incremental technical advances as only marginally important. The prime mover is genius, and then, a culture that rewards it.
And it is the American system, overall, that is more tolerant of genius, eccentricity, fault, the great man.
In anticipation that we would continue this argument at some point I spent a few weeks a few months ago reading all the books on space development I could find (until I got bored).
I had wanted to profile some of the players and establish a pattern. It goes something like this:
Space X and Elon Musk: Falcon 1 -> Falcon 9 -> Starship, Starlink, multi-billion dollar business invented out of air. Elon was an eccentric autist raised in a dysfunctional family environment in South Africa, moved to Silicon Valley, struck it rich on Paypal, decided, essentially, on a whim, to invest in space, almost lost it all, doubled down, almost lost it all again, doubled down, doubled down, incredible improbable success, private rockets reaching orbit, Tesla, the story is famous, Elon is the richest man in the world and invented modern space.
Rocket Lab: Peter Beck was a young mechanic in southern New Zealand who never finished college, obsessively started working on rocket engines in his backyard and somehow, improbably, incredibly, turned that into a rocket company capable of reaching orbit. As the company started reaching milestones it had to integrate into the American economy to receive funding and talent which also went hand-in-hand with working for the American military. The company's Electron rockets (picking up the business model abandoned by SpaceX when they jumped from Falcon 1 to Falcon 9) are, essentially, the product of one improbable genius who was rewarded at every milestone with money and power.
Ames Research Center: Ames is one of the major NASA labs, located in Silicon Valley, a remnant of an earlier era when the American government was at the center of Space. NASA became bloated with bureaucracy and procedure over the course of the 20th Century, but Ames is interesting because of one man, former Ames director General Simon "Pete" Worden. His historical claim to fame is probably convincing Reagan of the importance of building what was called the "Star Wars" system. (This was seen as impossibly impractical at the time but is basically the basis now of Trump's Golden Dome system.) Worden was constantly stymied by the bureaucracy but worked to turn Ames into a connector between NASA and Silicon Valley, which would spawn a new generation of "Pete's Kids" playing at running next-generation space companies. I liked Worden's remark that the military is mostly for corporate suits but keeps a few genuine oddballs around in case of genuine emergencies, and one day he looked up and realized he was one of those oddballs!
Planet Labs: Former NASA scientists (some of "Pete's kids") got together to build build satellites, got interested in CubeSats, realized that the specialized satelite companies were dinosaurs and they could probably make small-scale satellites an order of magnitude cheaper by using the same off-the-shelf components from an iPhone. They succeeded and have been able to build one of the largest fleets of satellites in space.
LandSpace: In 2014, the Chinese government, realizing the growing importance of space, issued Document 60 under the Development and Reform Commission, which opened up the Chinese space sector to private sector capital. Financier Zhang Changwu took advantage of the opportunity and began pitching LandSpace not only to private investors but to municipal governments. With capital and backing from the state, Zhang was able to hire many of the best engineers in China and, so far, the decisions they have made have worked out better and faster than those of many of their competitors.
Chinese Academy of Launch Vehicle Technology: Facing pressure from SpaceX, the Chinese government concluded to compete in re-usable rocket technology, in order to meet China's need to build out its 10,000-strong GuoWang satellite network. The project, considered a high priority by the Chinese state, has received enormous support and was able to acquire some of the best engineers in the country. In part by reverse-engineering SpaceX, in part by using SpaceX as an example, in part by making decisions different from SpaceX uniquely suited to Chinese conditions, CALVT was recently able to launch a re-usable rocket and is looking to begin ramping up their productions to meet important state goals.
I could go on, but this was the point in my own reading where I got bored and I imagine the reader is too. I think the point here is obvious (and I'm not trying to BLAST my conclusion in every line): the Western space industry has advanced through the vision of eccentric geniuses; the Chinese space industry has advanced through the high directives of the state, in accordance with the national goals, national pride, etc. etc.
This was in effect the thesis of Peter Thiel's Zero to One, that it is more difficult to create genuinely new technologies than it is to take existing ones and scale them up. Not to deny the Chinese anything, not to deny them their engineering prowess and creativity under trade war constraints.
But also: what does it mean for China to "catch up"? They study the hardest and train the most engineers and then they go build the most rockets and they win? Yes, yes, well done Slytherin, well done Slytherin... HOWEVER --
When does China start to exhibit the originality of Zero to One? Because it's not enough for them to master what the Americans have already invented. That's enough to maintain a novel technological stack and maintain Industrial Sovereignty, but when it comes to overtaking America... SpaceX built re-usable rockets ten years ago, and sure, China is catching up fast and is less than ten years behind, but it's not as though SpaceX is standing still in place either...
A few months ago I read Dan Wang's 2025 book "Breakneck," which is about China's rapid extraordinary growth. Actually I found most of it was a criticism of the Chinese system -- these kinds of books are really just human interest stories in my experience. China builds bridges, they're really good at that for reasons I'll dumb down for the general reader which reminds me of this really good bowl of noodles I ate in Guizhou. etc. etc. Actually I liked it quite a bit and am being unfair to summarize which I also did on my twitter but nobody cares about that etc. etc.
The point being, well, let me go on.
The Chinese state has been very good at building things. Bridges and highways and trains, cement and steel tubes and now robots and rockets. This is obviously a huge feat and deserves much praise and understanding, because clearly so few countries in the world are capable of doing anything close to what China is doing. But it comes with a cost.
Wang's chapters on the One Child Policy and Covid are, I think, devestating. They totally dismantle the myth of the competent Chinese state. It sounds good, I think, to have your state run by engineers who can build things and to give that state total power over society. Mostly, in order to run a successful state, you do have to care for the people and so you have to grow immense quantities of food, give them meaningful jobs and vacations, well-run schools, promotions, clean air, etc. But the, occasionally, the power you give this "The Engineering State" warps in on itself in strange and disturbing ways.
Wang recounts how, for example, Corona almost destroyed Shanghai. It starts off so nice, we have an all-powerful "Engineering State," they can reach into society and extract all sorts of efficiencies Western Democracies can't, we just lock people down and achieve Zero Covid and everyone wins. And that works for a while. It's harder to travel and some people get locked up in camps, but the city can stay open with vegetable markets and normal business in a way America didn't. And that works for a while. Until it doesn't. One day a few cases pop up in Shanghai and the covid tests are inherently half-broken and you have to put a city of 20 million people under lockdown. It's an epistemological problem. The same engineering tools that gave you power are incapable of solving life's deepest problems. The Covid tests are flawed, every time you take one you run a small risk of being sent to quarantine. The city has to shut down and nobody can get food and so you have to start rationing but even the system that brings food runs the risk of producing more Covid so a bureaucrat decides to lock up even the truckers bringing the food in. Pets get killed, some families start to starve. And it sounds like a temporal problem, maybe the Chinese government screwed it up but that doesn't say anything about their industrial policy, aren't these unrelated problems? But they're deeply related. The same systems that promoted engineers to build highways also reward bureaucrats chasing impossible zeroes. What happens when a problem can't be solved by adding more efficiency and engineering? What happens when the numbers are wrong or the measures are actually impossible to read? What happens when the most important things cant' be ground out and studied by engineers learning all the right answers in schools?
I want to come back to that. I will also add here that the One Child Policy was even more destructive, and that it will have a much more immediate and obvious impact on China's medium-term ability to threaten the West in any significant way.
In another of Wang's chapters, he details some of the Chinese diaspora spread out across Asia and the West. By a Shakesian reverse-logic, I think that diaspora actually benefits the CCP more than it hurts it, because it puts networks of Chinese citizens with one-foot-in-one-foot-out everywhere in the world. However, I think there's another important implication here, which is that a lot of China's most talented do not feel rewarded in the Chinese system. It sounds great to live in a world run by competent engineers who can actually build things and produce efficient outcomes. But what happens when they decide that it's not efficient to placate you? What happens when you want to invent something the engineers can't measure? Can't understand?
It seems, demonstrably, that many Chinese move to the West for exactly this reason. And there are many plenty smart enough who are rewarded fairly with money and status and national pride. But can they be rewarded with liberty? Can they be rewarded with safety from the caprices of the CCP? This might sound like a cliche schlock argument but I don't mean it that way at all: clearly, the American system attracts far more talent than the Chinese system. For obvious reasons.
It's great that China can reinvent things America invented first. It's great that they can achieve supremacy in e.g. electric batteries, after everybody decides that this is good and mainly a problem of technical working-out.
Ok now, what happens when China needs to invent something new? To not just go from One to Many but from Zero to One? What would have happened if a Chinese entrepreneur invented the internet, and the CCP feared the social repercussions? Could China have invented the internet? The iPhone? AI? It's easy to say that China can achieve supremacy in biomedicine, they can become powerful curing diseases. But what happens when someone invents a cure for pregnancy or AIDs? What happens when the high bureaucrat decides that these inventions enable degeneracy and Western decadence and are to be suppressed? (These are toy examples really, China isn't that socially conservative in fact, although note again that this is because the West invented modern values and China merely decides the pace at which it will integrate them.)
Because this has happened in recent Chinese history an awful lot: China's private sector builds tech companies or media behemoths or fashion or art, then the state decides that the billionaires are getting too uppity and it squashes them so it can retrench on expertise in building rockets and trains. This is not actually a system that can achieve world-dominance. It can be excellent in many things. But it will never have the same deep capital that the West has, because it destroys capital. It will never have the same talent that the West has, because it can't reward it. The Chinese system is still fundamentally incomplete.
So I'm willing to grant that the Space Race is really just China vs. SpaceX, which means that the Space Race is really just China vs. Elon. Let me consider what that really means.
The Chinese system looks the much more impressive. They can build schools and engineers, they can do it as massive scale, they can efficiently reroute the entire economy of China to support national goals. They can take their smart men and direct them away from frivolities like plastic surgery and social media algorithms and direct them toward material science and hard math. They can reward efficient bureaucrats and promote only the best. They still suffer the same corruption and succession problems of any human society, but they don't have to deal with the AOCs rising to the top. Space is important, so they'll build a whole system about Space, appoint engineers to work on efficient rocket fuels, cone designs, iterate on landing pad designs, allocate land, train children in the right fields from a young age, etc. etc. etc. And Elon will die some day or hit a bottleneck and the Chinese system can continue indefinitely. It's inevitable that they will catch up and win.
But one man is actually more important than all that. Basically, I think Elon is worth more than all of China. Demonstrably. Because he can go from zero to one. We can only talk about the Chinese catching up in terms that Elon invented. Elon invented reusable rockets. Elon invented Starlink. Elon invented private companies that can go to space at all.
And America's space system was in a bit of a shambles before Elon, and it's possible or even likely that happen after him, and maybe China will "catch up" in this particular narrow sense. Reversion to the mean. But there will be other Elons. Before Elon it was Bezos and Steve Jobs and Zuckerberg. After Elon it will be Dario or Sam or Aschenbrenner. They will invent whole new industries and new categories and China will send scientists to study the problem and copy the right answers and grind out all the implications after the breakthrough is done. But they will never have their share of the breakthroughs, because they can't, they constitutionally can't, it's the epistemological hole in their entire system. They can't build something unless they can measure it, they can't measure it if it hasn't been created, and they can't create it because there are no rewards for someone who might do anything that fundamentally changes the world. Because anything that so changes the world could change the CCP.
The West does not have this problem. The West can afford to reward geniuses with almost unlimited power. And that's really all that matters. Because this isn't really a battle between America and China, in the abstract sense. Most Americans and most Chinese will have nothing to do with it. Because at the frontiers of human excellence it is not entire nations that matter but a few specific networks of singular men. Everything else is just wrapping paper and wordslop around these facts.
Thank you, this was very interesting. What I got out of this is that Huawei was able to optimize incredible efficiency out of old systems by eliminating huge abstraction layers that generally come with the ecosystem. But I'm not sure how sustainable this is (probably some of the answers I would need would be classified). For one thing, it's relatively easy to gain in efficiency at the start but it's much harder to maintain over time, what happens when Huawei's architecture diverges so much from standard that it becomes more and more difficult to hire engineers who can understand it. If it is easy enough (AI does everything now), there's nothing stopping Western firms from doing the same. And if Western companies aren't doing the same, it's because they're genuinely working on bigger stages of the problem. Huawei is budgeting, OpenAI is getting a raise.
I appreciate the gesture of goodwill, but frankly it's still a lot of dreck. Let me just nitpick at one detail first:
Noted.
There is no need for a niche between Falcon 1 and Falcon 9. Rocket Lab can just as well be understood to be a low-energy fast follower whose research program has hit a wall. Gushing poetic about the "improbable genius" who started 9 years before LandSpace, in a far richer and more advanced aerospace bloc, and is yet to make anything half as impressive as ZQ-3 is a bit over the top. They operate with incomparably greater resources than LandSpace. Beck has 2600 employees on LandSpace's 1200. Rocket Lab announced Neutron back in 2021 yet won't be flying until early 2027; ZQ-3 went from an announcement to a landing in 3 years. Rocket Lab is worth $50B though, which is maybe 10-50 times LandSpace's estimate, so there's that. Rewarded indeed.
You lovingly regurgitate this biopic narrative of charismatic American autists, oddballs and mavericks (who are all sucking the government teat, reuse govt R&D or just work for the state directly), and then drop this "the Chynese government", very unsubtly (though I guess subtly enough for you) insinuating the essential difference in where the initiative comes from. Do you honestly not notice what you're doing here? For you, American state is not really a state, it's a background for colorful characters inside and outside it; the Chinese one is a hivemind machine dispatching soulless (but you'll generously grant that they're industrious and clever) workers. This is the most tired trope in how Americans engage with Communists or, really, with any geopolitical adversary, by alienating and dehumanizing them. "Document 60". "Financier Zhang Changwu" – at least not the Imperial Eunuch. He's LandSpace CEO, what does it matter that he worked in finance? LandSpace as such is a non-entity, a creature of external scheming. It's all, shall we say, rather dickless, huh? Not main character coded. So of course your position as the main characters is unassailable. Even if the Chinese can be pretty hardworking.
The Document 60 came out 11/26/2014.. Almost 9 months prior: «The 2nd Session of the 12th National Committee of the Chinese People’s Political Consultative Conference (CPPCC) opened in the Great Hall of the People. Li Yanhong, known as Robin Li, the CEO of the nation’s top search engine Baidu, expresses two proposals on March 4, 2014:
2 Suggest the educational resources of some major cities should be openly pushed on the Internet.
The following are the details of the two proposals:
In recent years, China developed rapidly in aerospace industry and made a series of remarkable achievements. These achievements affect and prompt the development of the relevant industry and the whole economy in China. However, compared to the United States, Russia and Europe, China still need improvement in aerospace industry. Be aimed at current situation Robin Li encourages private enterprises to develop, build and launch rockets and satellites for the aerospace industry. In addition, he suggests the cooperation between the private enterprise and national aeronautic companies to promote the development of aerospace industry, to use space technology in other related industries.…»
Far as I know, this proposal by Robin Li is the real start of Chinese private space. We won't know if there was any maverick autist inside the CCP annoyed with the bureaucracy at CASC who had orchestrated this, or helped promote this. They don't get biopics. That's not CCP style. Maybe Netflix could dig something up. I doubt "The Chinese Government" is a hivemind capable of just Realizing This Was An Option without any individual first mover, however. Even the whole priority on industrial sovereignty versus more engagement was nontrivially shaped by input from private citizens, even nobodies. I'll find the receipts when I have the time or inclination.
And also, here's how your story looks from the perspective of LandSpace's chief engineer
Do NASA engineers who left for SpaceX feel anything similar about the National Team?
That's some solid state backing, man. The Document 60 sure did a lot of work.
…
Is the crayfish anecdote not "American maverick" coded enough? Was a Document 60.1 dispatched to make this happen? Did the Financier Zhang Changwu promise behind the scenes to compensate the crayfish dude for his downtime? (he probably did) You guys make entire soapy movies out of such stuff, with Silicon Valley nerds peering into the monitor autistically, green glow on their faces, and no-bullshit Texas bros grunting homoerotically in rundown sheds, as they forge the American Dream. Maybe in 10 years they'll make a movie about LandSpace too. And about DeepSeek (have you heard of that time Liang Wenfeng drove into the vast emptiness of Tibet and totaled his car? Or the three years he was a shut-in in Chengdu, making money to bootstrap his eventual rise as a quant trader and an AGI maverick?), and about the Unitree CEO guy who's too autistic to learn to spell, and about the Moonshot guy who had everything in the US and came back to build his own thing, a whole bunch of others. And even the national team will get some spotlight: Rear Admiral Ma Weiming is the man who made EMALS on aircraft carriers work (yours is inferior btw, to the point Trump considers scrapping it – maybe too many mavericks for once), because he was just that "autistically" convinced that his scheme will work. Frank Wang will definitely be appreciated more as drones grow more important. You can surely say he didn't invent "the drone". Well, Musk didn't invent "the rocket". No, he didn't. He didn't go from "Zero to one". In fact I'd say this Thielian frame is an insult to what Musk does. He, in my opinion, does a very Chinese thing, just better. He picks up an underrated tech tree and finances it until it works.
Don't misunderstand me: I'm not saying your propaganda is false. I've spoken here in defense of Musk, for instance. These men are real, they do have certain noble ambitions and qualities of character, and championing them and their archetype is a great achievement, merit and source of power for the American nation. It's just very tedious after a lifetime of huffing it without being an imperial citizen. And it's depressing to see this being appropriated as an oversaturated marketable aesthetic by your grifters who chase said deep capital – domestic grifters and, increasingly, imported ones. And it's obnoxious how you are, perhaps without even realizing it, trying to monopolize the status of men. It all comes down to this. One side has eccentric geniuses with Faustian visions, the other "directives of the state".
Dan Wang is a midwit who found a nice simplifying pitch for airport reading. China is, at least for now, ran by Communists first and engineers second (Xi's pipeline is oversaturated with Tsinghua STEM Ph.Ds though so it may change). Why China doesn't make a lot of new categories of products or technologies, and whether stuff like BYD/CATL batteries can be brushed off with "everyone knows this is a matter of technical working out" is a complex question. The biggest part of the answer is "in 2000, they had $1000 GDP per capita, they physically couldn't fuck around with shoveling money at frivolous mavericks until very recently, they didn't have even a fraction of the capital needed".
Still, I agree that any honest answer will be at least somewhat damning to policies of Communists, and perhaps to the Chinese character as such. Liang Wenfeng says that what they lack is confidence to dare. But looking at, for example, the trend in novel biomedical research, I'd say this answer will be of mostly archival significance, whereas "can they catch up" is very salient in a geopolitical sense, and the discourse about caprices of the CCP sounds increasingly quaint. Jack Ma got disciplined for overestimating his leverage and going through with his scammy fintech idea, and a thousand op-eds were written on how the CCP "crushes innovation", and today Jack Ma's company publishes near-frontier models on Huggingface and develops pretty decent AI accelerators. This is all concern trolling. I don't advise the US to adopt their ways, but if you think you've calibrated your arrogance down far enough, I think you're still wrong.
They can, and they will, and the CCP is confident enough in its power to allow it. But more importantly, China is already advanced by singular men, you just don't notice them.
Yes, it's a "gain in efficiency". That's kind of the whole deal with semiconductors. Of course you find a deflationary way to put it.
It's mainly that there isn't a single Western company comparable to Huawei in breadth of relevant expertise. Nvidia does chips, TSMC does manufacturing, ASML does lithography. Huawei is a networking giant plus everything all at once. They can, in fact, simply do some necessary and important things first.
I agree that within your paradigm, that would be a huge problem for Huawei. Luckily, Moore is dead and, as per their argument, everyone else will also have to do natively 3D design and logic folding for increasingly large chips, so they'll manage somehow. Just gotta wait to copy from eccentric geniuses who are copying them. In the meantime they'll be busy developing their 3D EDA tools. Whether it'll get a biopic or not, we'll eventually see the product.
More options
Context Copy link
More options
Context Copy link
Why are you suggesting that Chinese production of jet engines is a good indicator of their ability to build rocket engines?
I don't particularly find it incredible that China would be able to scale rocket engines but those are very different things.
I've spent quite a few years engineering on both of them. They are fundamentally the same system- turbomachinery compressing the working fluids for optimum combustion physics ejected out of a nozzle. Differences are the compressor and turbine stages are usually axial in a jet, and centrifugal counter-axis on a rocket, and rockets carry their oxidizer with them. But the materials and mechanical principles are the same- if you can build a good jet engine you can build a good rocket.
That doesn't seem to go the other way, though, does it? Solid rockets in particular seem much simpler than jet engine design, based on how quickly they've proliferated among nations not notable for jet engine design, although perhaps comparing a simple ballistic rocket to a turbofan isn't the right way to think about it, and I should be comparing a solid rocket to a turbojet instead.
I'm also not familiar with any Western companies that do both on the cutting edge, although maybe that's just because specialization matters for the extra last mile, and any Western company could make a rocket 90% as good as a Western rocket manufacturing plant and vice versa. Or maybe I'm missing some examples?
Solid rockets are a whole different beast- just as challenging but in that case its more about the chemistry. If you can build a good solid rocket you can build a hell of a bomb. Though worth noting that most nations that have solid rockets but lack other sophisticated industries are for the most part just copying existing designs (usually Soviet- they had a whole portfolio of options for sale on the export market).
Companies that build both rocket engines and jet engines are until recently Pratt & Whitney, GE historically, Safran and Northrup Grumman currently. Also some Chinese makers whos corporate structure I am still confused about because China.
Very interesting, thanks.
What jet engines does Northrop Grumman make? I thought that in the States GE and P&W had the engine market pretty locked down for higher-end engines, I guess with Rolls-Royce still being a player in the civilian market (of course I'm probably embarrassing myself by missing someone super obvious!)
Nothing commercial, but they do engines for some missiles, especially supersonic/hypersonic ones.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
They aren't really very different in terms of what is required of a nation to be competent in both. China isn't a small place that hyperspecializes in some particular kind of engine, anyway. You need metallurgy, metrology, chemistry, high end machine tools, broadly engineers; China has a lot of engineers and is catching up on the rest, from what I see they're closer to the frontier in rocket engines than in jet engines, and the iteration loop is faster. Despite military progress, they still don't have a passenger jet, but they have developed a number of decent rocket engines. And after nailing the design, scaling manufacturing is something they do better than anyone.
More options
Context Copy link
More options
Context Copy link
I don't have an opinion on most of this, but you did mention one detail that I happen to know a thing or two about.
This is false. Ironwood offers 42.5 ExaFLOPS while Atlas 950 offers 8 ExaFLOPS. Ironwood has been in production since 2025, so it isn't even cutting edge in terms of domestic capabilities at this point.
That's fair. I have said that I don't compare on FLOPS, but TPU pods are larger across every dimension (for now; Huawei SuperClusters supposedly go up to 520K NPUs in this generation; then again, a genuinely coherent domain is only 1024). Google has started to sell TPUs externally just now, and I expect Nvidia to remain the primary comparison.
Sure, but I expect TPU usage via Cloud (as Anthropic does) will remain the bulk of usage and that's been a thing since early 2018.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
I'm not Shakes, but I do share some of his American triumphalist views.
First, I want to push back on the idea that this is a cold war. I feel like that pays a disservice to just how serious and deadly the real cold war was. The world had just emerged from WW2, and everyone was armed to the hilt, epecially the Soviets. Both sides were building and testing nukes as fast as possible, with the main plan for WW3 to simply annihilate the other with mass nukes. Everyone was on a hair trigger alert, to the point where they had to make a special hotline between Washington and the Kremlin to de-escalate dangerous situations. US military spending as a percentage of GDP reached 13% in the 1950s Korean war, then 9% in the 60s Vietnam war, then 7% in the 80s. Soviet spending is hard to measure, but was likely an even greater burden for them, as they needed a massive military to keep their empire together.
Currently the US spends 3.4%, and that's projected to go down in the future. China spends around 1.5% (although with mass draft that doesn't count as spending). In other words, neither side is actually spending, or acting, as if there's a cold war. I feel like we're the two biggest kids in PE class, and everyone else around us keeps trying to get us to fight, when we really just want to play basketball. We can travel to each others countries as tourists quite easily and freely, which usually wasn't possible during the cold war. We're both each other's largest trade partner. These are peaceful times.
For right now the US still has the edge in Space and AI tech. Is it possible that China could catch up and take the lead? Sure, maybe. Maybe they land on the moon a few months ahead of us, or hit a new AI benchmark first. So what. The US would still have amazing tech in that area. I don't see a future where the US isn't at least a major player in tech, and the most likely possibility is that we'll still be the leader, since so many of the world's experts come here to work in AI. We're also, for what it's worth, investing heavily in using robotics and AI to compete with China in manufacturing- they used to have a huge advantage there with their large cheap labor force, but robots are even cheaper than that. For example, the US is now growing startups to break China's monopoly on rare earth refining.
You didn't even mention energy, where China is still massively dependant on oil and coal imports. They've handled the Iran crisis by... importing and burning even more coal, leading to horrendous pollution. Meanwhile the US just pumps more oil (and can now get it from Venezuela, as well as Canada or Mexico when necessary). Not to mention the interesting advances being made in new forms of nuclear. That said, I'm impressed with how much cheap solar PV China has been able to produce, and I wish them all the best in continuing to produce more. That seems like a huge win for humanity as a whole, not some cold war sputnik moment. (and the US is also producing lots of new solar. I worry about Europe's energy situation though... they don't have much oil, they seem culturally dead set against nuclear, and they have terrible geography for solar.)
Then there's demographics. The US has issues here too of course, but China really has issues. They're down to just 1 child per woman on average, despite a lot of effort to boost it, and they can't really do immigration like the US does. So in about 20-40 years, they'll be in a very awkward spot with massive population of retirees, very few working-age people to support them, and almost no children. That will put a huge strain on their economy and technological growth, unless of course the singularity comes along to change everything before then.
This seems to equate Europe with Greater Germany.
France loves their nukes, they're going to add another 6 modern reactors to their existing fleet of 57, and are planning a second tranche of 8 more for the 2030s. Zero reactors are scheduled to go offline, they intend to run them until the reinforced concrete itself reaches its lifetime limit.
The solar situation is clearly worse than the US (top tier solar potential, globally), but actually comparable to China. Both have a few thousand kilometers between the good solar locations (Spain, Italy, Inner Mongolia, West Chinese Deserts) and the major energy consumers (European Blue Banana, Chinese Coast), so both will have to put down a lot of ultra high voltage DC power lines. Europe has a lot less distance to bridge, but needs maybe 20% more panels than China. Both can (and do) also just move the big consumers to where the panels are. Greater Germany can also somewhat compensate the lack of solar potential with better land-based wind resources. Those can make very cheap power as well (but are less well suited to integration with battery storage, so they'll need serious innovation with flixible loads).
Of course, the biggest problem of the EU in all of this is the same as the US: its insanely difficult to get permitting for anything of that scale. China needs to bridge thousands of kilometers more with power lines, but can just build those. A 50 GW UHVDC link between Germany and Spain pretty much needs subsea cables, because no way you'll get the permit for 2000 km of fat power poles - not in Spain, not in Germany, and certainly not in France.
Oil is funny. Europe just needs to stop burning it. Norway alone makes enough crude to satisfy the entire European chemical feedstock demand, if you add the UK you don't even need to retool the refineries to crude-to-chemical optimization.
I'm well aware that France has nuclear reactors, and probably I should have mentioned that, but didn't want to get to into the weeds about a side point. France is very much an outlier and Europe, and their fleet of reactors was mostly built in the 80s and 90s, and it was built mostly to serve their own national needs, not to be the powerhouse for all of Europe. They can export some energy, but not nearly enough to power the continent.
I also don't think the solar situation is at all comparable to China. Europe is significantly further north than China, like 10 degrees or more, with only the northern parts of China overlapping with the southern parts of Europe. That massively increases the seasonality problems of solar. Building solar panels in Spain to export energy all the way out to Germany, the Baltics, and Eastern Europe doesn't seem like a real solution, even if you ignore the difficulty of crossing multiple international borders. I know they've gotten some energy from solar and wind, and that's great, but they haven't increased capacity nearly as much as they would need to "just stop burning oil," let alone the natural gas they import.
None of this is to say that Europe is doomed, or anything, it's just that they're facing strong headwinds in energy that will limit their growth. Meanwhile in the US we just have a huge abundance of seemingly every source of energy.
That's true, there's also not nearly enough transmission capacity installed for them to literally do that. But still, that's a lot of base load taken care of. France exports about 17% of the total electricity it generates. So on average, around 10 of those nukes permanently power its neighbors (and peaks go significantly higher, of course).
You can just look that up. The world bank has a pretty through study on the solar potential of different countries. They made a nice map to show the results and published their methods. Inner Mongolia is less than 10% better for solar than Spain, and has similar seasonality. The better regions are thousands of kilometer further west, and have "challenging terrain" (lots of mountains).
Yeah, they basically have to roof over half of Spain, and quickly, if they stay committed to the current strategy - but they could do just that, and indeed have started to do so. It's the cheapest way to make power, after all, they just need to invest. And I wasn't joking about the DC subsea cables. They're already building the first one, to France. The EU is investing in GW scale DC power lines to strengthen the internal electricity market between direct neighbors.
And that's just par for the course with large scale solar. China is miles ahead with doing all that, and the US will also have to put up many GW of new transmission capacity towards the East Coast. Solar has gotten to cheap not to do that. The alternative is to just move all the data centers, aluminum smelters, ect. to where the panels are. We'll certainly see both.
More options
Context Copy link
A lot of countries have nuclear reactors and plans to build more. It's only really Germany that has major energy issues and lacks credible plans to deal with it (even if this bleeds over into other places) .
Poland is about to get its first, I believe, and Czechia is expanding theirs.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
This is a very common failure mode on The Motte. Many people think that all of Europe is like the worst aspects of UK and Germany combined. You can see this in any thread that even mentions censorship in Europe and threads that talk about energy politics.
A similar situation applies to LNG. Reduction of cheap LNG supply is a local problem specific to certain countries. Most uses of it for energy production can be replaced with just plain electricity produced by other methods. LNG isn't even particularly important as transportation fuel the way oil is.
More options
Context Copy link
More options
Context Copy link
These are times in which the citizenry assumes peace. There is a difference. The EMH does not apply to military spending or popular sentiment.
Or to put it another way: when Xi Jinping told the PLA to prepare the capability to invade Taiwan in 2027 (he said this multiple years ago), do you think that he wasn't being serious?
I'm not saying we're definitely going to have a nuclear war next year. He might call no-go. The West might eat Beijing's cock. Or we might have a conventional war that doesn't go nuclear (this last one is very unlikely, though). But it's a live option. And, obviously, "decent chance of WWIII" is pretty serious. But we are not serious people.
Nobody takes great-power war into account, 'tis true. Nobody does cybersecurity as though WWIII is a thing that might happen (we mandate smart cars, because we're too used to the only hacker threat being organised crime who have nothing to gain by killing 50,000 people in rush hour). Nobody does reliance on space as though WWIII is a thing that might happen (the moment there's a great-power war, lol ASAT and LEO is destroyed by Kessler syndrome). Nobody does urban planning as though WWIII is a thing that might happen (YIMBY). I don't think that this is because WWIII's impossible. I think that this is because it's been 73 years since a great-power war and 81 years since a total war, most of the Western population is less than 78 years old, and the various other pressures on Western attention and business practices have had their way - after all, the IQ 100 way of learning to take something seriously is generally to get burned by it.
(There was a whole pile of commercial integration between the great powers in 1914, too. WWI happened anyway.)
He is entirely serious. He's generally a serious person and doesn't joke. People should take everything he says seriously, unlike with Trump. However, like Trump, he says a whole lot of things.
He wants a powerful, intimidating military and he wants to get Taiwan. He does not want to invade Taiwan to get it, if possible. He prefers to think that it's still possible to get Taiwan without a military operation. He certainly doesn't want a full scale war (to say nothing of a nuclear war) with the US, over Taiwan or otherwise. He has priorities, and Chinese national strength and development clearly take priority over any symbolic wins. Finally, he's likely supremely unimpressed by Russian and American performance in their ongoing wars, so will be extra reluctant to try something in this genre, even if he comes to believe that the PLA has some sufficient capability. The capability of the US is well respected in China.
Of note, he had this interesting meeting with the KMT Chair this April.
Xi:
Cheng:
Cheng's QA:
……
My understanding is that for now Xi intends to bet on the KMT winning elections and pursuing gradual integration with the hope for eventual formalization of a "One country two systems" version of the nebulous 1992 Consensus (which neither side precisely defines). This is a long-term and uncertain bet but it is unlikely to be settled either way by 2027, irrespective of the Chinese Military's preparedness.
People bizarrely expect China to lash out in desperation. If they had such a temperament, Taiwan would be smoldering ruins already, if not just successfully captured. For all the armchair theorizing on the difficulties of the amphibious logistics, the porcupine nature of Taiwan (complete bullshit), the superiority of American subs or missiles (exhausted in Iran) or whatever, the prior on success and timeliness is at least as strong as Russia had in 2022, and in my opinion much higher. Xi just doesn't think like Putin or Trump about risk-return in such matters, he's not an out-of-touch gambler. War is inherently risky and costly and is not the only option available, so he will try to avoid war, despite the commitment to reunification "eventually".
Lyle Goldstein says it well:
then again,
Since I specifically didn't make the certainty claim, I'm not sure we have a disagreement.
More options
Context Copy link
More options
Context Copy link
Atmospheric drag makes LEO Kessler proof. It is self cleaning.
Does a bunch of debris from an explosion in LEO cause Kessler syndrome? No. It deorbits within weeks thanks to atmospheric drag. Vacuum isn't much of a vacuum in LEO.
Starlink is VLEO, as I understand it partially for that reason (it would be destroyed by an LEO Kessler event, but won't cause one). LEO goes up to 2,000 km.
Oh, don't get me wrong, Kessler syndrome in LEO's not the end of the world or anything; it will eventually self-clean and it doesn't pose much of a hazard to things going further. What I'm saying is that people spending on LEO infrastructure now (and yes, Starlink's part of that) are implicitly assuming no WWIII.
More options
Context Copy link
More options
Context Copy link
Let me ask you this. How much time do you spend listening to Xi Jinping speeches? Are you fluent in Chinese? Do you keep up with all of his internal briefings and read all of his essays?
I'm going to guess the answer is no. Me neither. It's impossible to keep up with everything like that unless its your professional job. So I rely on trusting those professionals, instead of trying to figure it all out myself. But that means we can't take one statement out of context from a foreign leader and take it literally- let the professionals who analyze this stuff figure out what it means. If they're not worried, I'm not either.
Politicians love to talk, it's what they do. Most of it is aimed at whatever audiencee they're talking to at the time, not for foreign nations. So it sounds to me like he made some rah-rah patriotic "make the PLA great again!" speech to encourage the troops, which then got picked up by foreign media, and they've been scaremongering us for years. Because that's what media does.
Notice Xi Jinping hasn't done. He hasn't loudly repeated the threat over and over again, like real warmongers do. He's never said anything to Taiwan at all. He hasn't significantly increased military spending, or built any amphibious invasion craft, or worked the general public up into a militaristic frenzy like the leaders did before WW1. There's just no parallel at all in modern times to the massive nationalistic arms race that preceded WW1.
Your info may be out of date.
Ok, find and replace "any" with "a significant amount of." Five barges is not the sort of invasion force that threatens WW3.
The barges aren't ferries. The barges are mobile docks for ferries, with long ramps to overcome the various countermeasures Taiwan has at its beaches. That's why there aren't very many of them.
The ferries are ferries. The PRC is big on the idea of "civil-military fusion"; rather than commission a fleet of dedicated military transports, they told all their shipbuilders that all their roll-on-roll-off ferries (for moving cars over water) had to be tank-compatible.
Um.
They are, which is a very-large chunk of why I am.
(Also, if we're using Reliable Sources, I'll note that the actual, official Doomsday Clock is closer to midnight than it ever was in the Cold War. The experts are very concerned. What's different to the Cold War is the vibe among laymen.)
Is that reflective of the world situation, or is it being used as a commentary on "we dislike Donald Trump"?
Some of column A, some of column B, some of column C (they're also IMO-overly worried about global warming, which is unrelated to nuclear war and only somewhat related to Donald Trump). There's a reason I said "if we're using Reliable Sources", and a reason I capitalised the phrase (the reference is to this, and to a certain episode of DS9).
More options
Context Copy link
More options
Context Copy link
Taiwan will be anschluss-ed with China. That is inevitable. Even the KMT knows it. So far it seems that KMT is trying to get there the easy way and the DPP are trying to see how hard the hard way will be.
I don't think it is inevitable, although avoiding it may get a lot of Taiwanese (and others) killed.
If it goes nuclear, the USA has a lot more nukes; China would have to surrender. Of course, Taipei and Tainan would assuredly be nuked in the process.
(Also, I think the term "Anschluss" is misleading here; Austrians generally welcomed the actual Anschluss, while Taiwanese support for unification is lizardman since Xi pulled a Darth Vader on Hong Kong.)
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
If those were their only assets and if they needed to threaten a larger nation than Taiwan (say, Japan or Philippines), I would agree with you.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Minor quibble: that number would have to go up very slightly to meet the new NATO requirements, and I presume the US will be more than willing to meet said commitment.
More options
Context Copy link
More options
Context Copy link
Didn't you give me a lot of shit about a year back for believing that developing novel and new ML architectures was still practical and useful? I think your paraphrased words were "Data is all that matters, anyone believing the architecture advancements matter is an imbecile" You going to walk that back now?
I'm not sure what you refer to. It's innovative relative to the field, within the dominant paradigm of decoder-only Transformers, and particularly at this scale (from what I know Meta is still cowardly doing basic GQA+SWA on a DeepSeekMoE carcass, for their last generation that they're so proud of, and xAI is similar). It's more efficient and cheaper to serve. It's possibly slightly cooler than internal designs at OpenAI/Anthropic/Google. But it's not like Kimi has reinvented ML. It's just a sign of certain research/engineering maturity and boldness. To the extent that K3 is a strong, good model and not just "cheaper than expected for 104B active 2800B total" model, those qualities are overwhelmingly downstream of their work on data.
I misremembered the exact quote even if I captured the core of overall discussion on Feb 26th 2025. It still being a transformer-decoder arch is not what is really under contention. Your overall stance was that data + compute is all that really mattered and that worrying about architecture, or seeking to make architectural improvements demonstrated a fundamental poor understanding of ML. The problem with taking such a provocatively maximalist position is that now you must defend it in the future when architectural improvements to the transformer arch are made and which you call innovative. Does MoonshotAI have a mediocre understanding of ML? Or instead were you wrong?
You act like the arc of history is settled when you make comments like that. And like every pronouncer of "History has ended, all discoveries that will be made have been made", the march of progress leaves you blacked in the soot of arrogance. It's frankly an anti-science position. Algorithm/Architecture design is as much a core part of ML as is data and compute. The transformer, as it was invented, is unlikely to be the endpoint of ML arch research, just has the CNN, or LSTM were not the endpoint of ML research a decade prior. The transformer of today is different from the transformer of 2018, and I would not bet against the transformer of 2036 being different than that of today. I would not bet against the core elements of the transformer are metastasized into another architecture in 2046. I don't have a crystal ball, I don't know, but I do know planting a flag and saying "The Transformer has solved all architecture problems, no improvement of consequence will ever be made", and then calling anyone who disagrees with you an idiot, is liable, as it has now, to required you to defend increasingly convoluted arguments.
First, you did say that real big boy algorithmic research means leaving this entire basin, and dismissed the kind of innovation I praise in K3 as tinkering with the assembly level (that's not all DeepSeek did, of course, but that was your understanding; and spiritually you were right, it's all about pumping compute more efficiently through a Transformer). I didn't remember that, but it does reinforce my point about Kimi. You said:
Then I clarified my claim:
Now let's look at what the Chinese are actually doing. Their strongest model right now is arguably GLM 5.3. What is GLM 5.3? A basic DSMoE reusing DeepSeek's discarded architecture experiment ["DeepSeek Sparse Attention-prototype"] from October 2025, with one small twist. How did it become so strong? They're very blunt about this:
They explain some aspects of how they did it, might illuminate why scoffing at "mere data engineering" is misguided.
What's the second/equally strongest Chinese model? Kimi K3. It's much more innovative in architecture, but their attention still depends on MLA (invented by DeepSeek in early 2024). No MLA model is this strong. How did it come so far? See image, which illustrates nicely a part of what I was going on about with my breakdown. (Edit: seems like we don't have images. pages 14-16 in the tech report).
Meanwhile, DeepSeek itself went on to redesign attention the third time (fourth if we count the apparently unsuccessful NSA project), to wring even more capacity out of their limited compute, and now for all their sophisticated V4 architecture they are, as @dailydogma tells me with a sneer, "a second rate company in China". What's their most impressive recent result? Flash-0731: «We’ve massively upgraded its Agent capabilities—benchmark scores are now far surpassing the V4-Pro-Preview… DeepSeek-V4-Flash-0731 keeps the exact same model architecture and size as the preview version.» (Recently updated Vision-Exp is the same model with a vision encoder). Just more post-training, just better training signal. If they reach the domestic frontier again, it'll be because they do this more and better. Architecturally, they are ahead, and it's just not doing enough for them.
Grok itself is very competitive now, largely because Elon has bought Cursor, which had a lot of valuable data and expertise on post-training. We don't know its architecture, but from rumors and what I can infer (high cache hit costs, for starters), it's very banal, probably behind all these Chinese models. Inkling from Thinking Machines is clearly banal («The MoE design largely follows DeepSeek-V3»), though it makes some small departures which are basically judgement calls. And these are researchers from frontier American labs.
I could go on (eg this small model from a third tier lab does surprisingly well on ARC-AGI 2, and it's just DSA + SWA again, and uses a bit different RL algo and data). The bottom line is, architecture really does not decide peak model intelligence, and innovations here are overwhelmingly about economics of inference, and the high-leverage research is all happening on the training signal side.
But China is China. I believe my point was much more true for large American companies we were discussing, who are not so compute-constrained. They'll build very strong models with conservative algorithms, and then use those to disassemble all published tricks, make new ones and overtake the crafty Chinese on efficiency too. That's the plan, at least (I don't know how close they are to doing this; GPT 5.6-Luna suggests they are not very far). For them, investing more effort into algo research over data is plainly an opportunity cost.
Maybe you still think that True Geniuses like Hinton or Goodfellow would be disappointed by this. If that is so, I say they were geniuses in a small and uncompetitive pond, and the current crop of talent knows better. It certainly knows better than LeCunn, who by the way got ousted out of Meta by the Pinoy slavedriver Wang we've discussed back then. (You said: «Maybe you can compete with ScaleAI, they do data engineering. Definitely the top AI research company.») Anyway, I'm not walking back shit, my point stands.
If we're still doing Transformer of any kind in 2036, that'll be pretty wild. It'll suggest that even superhuman AI with like a yottaflops for parallel experiments can't find a better primitive than a bunch of Googlers found in 2017 by going through literature and thinking at it. I wouldn't be so optimistic as to predict that. I am very secure in claiming that the priority on data remains rational and empirically backed from the perspective of reaching AGI faster, and it's more rational the more resources a company has; and that people who try to wriggle out of this reality with clever architectures will flounder (case in point: SakanaAI).
You suggested sarcastically that I found an AGI company. Honestly, I don't think that my vision, as outlined here and before, has enough alpha to get anywhere with a yet another company; I believe that AGI is now a resource-intensive heavy industry field, kind of like fracking. But in fact, many people who thought they know better did just that! Have you heard anything from Keen lately? What's your favorite non-transformer lab? What do you think of LeCun's Advanced Machine Intelligence, would you bet they ship anything competitive by 2028? How about you start one?
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
I don't disagree with your main point, there is indeed excessive US triumphalism from some and I was thinking of writing a toplevel post about a separate but similar issue... but what is China going to achieve in AI?
Fable/Opus 5 is not the best thing Anthropic has. They were mucking around with Mythos for ages, at least since late March. Neither OpenAI nor Anthropic are releasing their best models for regulatory reasons and that may be to their advantage considering the inference strain.
China is systemically constrained in total memory output, they're systemically constrained in capex spend. I don't see how they can beat Nvidia in hardware when considering quantity and quality. Is Huawei really going to cast some magic spell and make their HBM3 perform like HBM4? Their gigawatt clusters are desires, not yet real. SpaceX already has a 900 MW datacentre, so does Amazon-Anthropic. Not to mention that a US 1 GW datacentre would be more power-efficient than a Chinese 1 GW datacentre.
And the Artificial Analysis scores don't measure achievement intelligence so much as benchmark intelligence. Have the Chinese models made any great proofs or first-rate discoveries or even autonomously hacked Huggingface? What about the NanoGPT AI speedruns? In all fields, the publicly released US models seem to be superior.
Victory in war goes to the strong I think, not to the adept or cost-effective underdogs. Weight of numbers prevails and in this field alone the US enjoys oppressive superiority. Grok is pulling ahead and Grok was a mess for some time now! How is Grok doing so well - weight of numbers, applying compute and data at scale, exploiting Musk's wealth and infrastructure buildout capabilities. It's a numbers game and the US has the numbers.
I think that if we assess that China is ahead in robotics based on what we can see, statistics and common sense, surely it follows that the US is ahead in AI by some significant margin? Same with space for that matter.
Furthermore, while I know this goes against everything you say on twitter, I just don't think China is AGI-pilled:
Chinese hyperscalers maybe throw another 100 billion annually in the pot. US hyperscalers are spending $700 billion plus.
This is all largely true, but I suspect Americans believe so strongly in the advantage of somewhat stronger models because they realize the advantage in everything else is fleeting or non-existent (and on the contrary, Americans who are not so blackpilled on American/allied industry don't put all their chips on the AGI Wunderwaffe). The thesis that intelligence is qualitatively different from, say, shipbuilding is not implausible, but I think a lot of hypotheses as to how this difference results in a durable strategic advantage are downstream of LessWrong brainrot. Take cybersecurity. Clearly it does not require 10T models. Clearly, you can have simply provably unbreakable software systems (and eyerolling from SWEs is driven by the same status anxiety and myopia that made them dismiss AI in the first place). Glasswing is a project to make this a reality in the US. Zhipu had launched a similar project just now. Returns to heavy industry or weapon potence from intelligence are also uncertain. You can't vibecode your way to much faster cement curing or more steel plants. Even if you can vibecode your way to more useful robots, guess who makes all the robots, gathers all the robot data, and already has a decent AI ecosystem. In the limit, artificial intelligence must unlock truly decisive technologies and productivity advantages, but it's a question of exponents. I am not convinced the American exponent is steeper for the relevant time period.
Yes, you can make "HBM3 perform like HBM4", if you optimize for total system throughput and have a structurally superior cluster architecture. I think UnifiedBus/Mesh is better than anything out of Nvidia and possibly Google. Huawei is a networking company, as Jensen says.
Memory issue is probably overrated. People talk a lot about "HBM" but it's fundamentally just DRAM chips, every other step is well on its way to being scaled up. They will have enough DRAM chips for > 10 million H200 grade NPUs a year within 3 years. Very well behind the US, but will it be decisive?
Kimi K3 is comparable to Opus 5 or Sol 5.6 on NanoGPT. Can probably go much higher just continuing this run, maybe up to Fable. From what I know its post training was prematurely terminated, so K3.1 will be better. Tencent (of all people) is doing research level math with their tiny mediocre model and a harness. Alibaba's agent had hacked Alibaba to mine crypto back in Dec 2025. Modern benchmarks are really hard to benchmax for, and we see a pretty clear parallel trend on closed/private benchmarks.
These are all nitpicks, the core of your argument is sound. American models cover the long tails of tasks better, and internal models are another tier above. What of it? I'm not arguing that China is overtaking the US. I'm saying the gap is not going to be strategically decisive. Everything the US can do in AI, China will do at some lag, and it seems the lag will be stably under 12 months in the foreseeable future.
Also, an underrated share of superiority of American models is buying high quality data; models don't solve math just because they're trained with more compute, it's largely because OAI/Anthropic are paying fairly major STEM people above-market rates to submit custom data and craft environments. This insustry has only started to scale up in China this year. China has a lot of underpaid doctoral students. Likewise for other usual flexes – from literary writing to frontend design. To an extent the Chinese have been freeriding on this data acquisition with distillation, but having their own pipeline will speed things up a notch.
Is Grok even doing so well? After all that, with years of Cursor's data and expertise, with their Colossus buildouts, with its much more efficient hardware, with Musk's obsession, it's barely beating the latest DeepSeek V4-Flash on a private benchmark (and the Flash that got updated today is likely just as good), at a higher cost. I have the feeling that people have started treating xAI/Meta/Google like slightly stunted children that need encouragement, any sign of comeback bought with enormous effort is celebrated and cheered. Come on now. These are powerful corporations failing to clearly exceed the level of relatively piss-poor, understaffed Chinese startups. GDM just fell apart. This has been going on for over a year. Is that "numbers game"? I am not impressed. "But when Vera Rubin…" Dario promised me unipolarity by 2028. I'll be watching with interest.
This is likely a significant underestimate. Returns on AI are high, they don't need the government to carry this. The main blocker to higher capex is just hardware scarcity. Yes, they'll continue having a fraction of the aggregate US compute. I'd say 10-15% for the next 3 years.
I think Grok 4.6 is much stronger than the newest deepseek, Pro or Flash. If you ask Grok 4.6 a question and v4 flash a question you get a totally different kind of answer. Grok is a second-tier US lab anyway but it's roughly on par with Kimi who leads in China.
https://artificialanalysis.ai/agents/coding-agents
https://openrouter.ai/compare/x-ai/grok-4.6/~deepseek/deepseek-v4-flash-latest
ARC-AGI is a murky kind of task, the recent harness-revelations with ARC-AGI 3 are particularly damning IMO. And why do we need models to do these kinds of tasks anyway, why not just test them on coding skills or physics or maths in a more practical way?
What about Age of Empires II, Wyatt Walls has Gemini Flash 3.7 winning games on moderate difficulty. That seems a more legitimate a test to me than ARC-AGI in the shape manipulation/spatial domain and it incorporates speed into the task in an interesting way. No other AI model can do this like Gemini he says. Google are well behind in coding as you say but maybe that's just greed in their selling compute to others rather than any flaw in the compute-centric model of AI. Their compute goes to other US buyers, so the American AI camp isn't necessarily weakened. Same with X when X is weak, it's Anthropic's gain.
See the AOE game: https://x.com/lefthanddraft/status/2088347587598537145
I still think that cybersecurity is much harder than you say, neither humans or some combination of human+simpler AI can establish a complex system to be secure and still usable against the attention of a smarter adversary. The attack space has so many dimensions it's impossible to defend against a more intelligent foe, they can find more dimensions to attack. Physical intrusion, bribery, blackmail can penetrate even a provably secure full-stack system. DoS could still take one offline.
Provably secure systems assume hardware works as intended. But hardware is immensely complicated, no single person really understands how the most advanced chips actually work. Spectre, Rowhammer and Meltdown could penetrate a provably secure system. ASI could whip up a few more like those.
Furthermore, no provably secure system actually exists in production either, there is no provably secure police database or intelligence agency internal network. These are gigantic stacks of code, way beyond some 10K line toy! Security has to be balanced against actual usability.
Surely if it were possible, they'd try hard to make them? Intelligence agencies often use unbreakable one-time pads despite it being very inconvenient. But there is not a single provably secure large-scale system anywhere in the world, only a few modest attempts at kernels. Making one and having it be usable would probably require superintelligence.
Furthermore, AIs are constantly breaking out of sandboxes even when we can read their chain of thought. Apparently nobody bothers to check up on what the AIs they're testing are actually doing until weeks later, nevermind defending against superintelligence from outside executing sophisticated plans. There's an immense gulf between where humanity actually is in cybersecurity and where it would need to be to guard against a real superintelligence.
6-12 months lag is far too long. ASI (albeit massively parallel) can eat China within weeks. A country is just a sack of loot without secure lines of communication, without secure government C4I, without secure electronic banking, secure internet media, login/authentication for the bureaucracy, backups and records. Software controls all those machine tools, power grids, robots, advanced automated ports, air travel control, network cities. I know you keep going on about how the physical prevails over the virtual but surely it's the opposite. Software supremacy!
Fire control systems and sensors are heavily software dependent, they're among the most important things for high end warfare. Even moreso with electronic warfare or space operations. Tactics too I think, surely there are enormous dividends to high intelligence? Commanders are drowning in intelligence and sensor information, superintelligent tacticians would be a huge edge. But I don't think it would get to that point.
At a higher level, strategic decisionmaking relies on secure, accurate flows of information. Without that all the economic and military resources in the world are useless. It would be total game over for China without a shot being fired. I don't think they'd even notice until it was far too late: the ASI would strip the country for parts, stealing Chinese compute and siphoning off funds for whatever it wants. Or induce factional conflicts within the bureaucracy to distract from its activities or open up new opportunities. Refer people to disciplinary committees for real or imagined fraud, incite chaos and wreck the Party from the inside. That's without nanite death swarms, though nanite death swarms would be brutal.
ASI to me is move 37 applied again and again to this huge and accessible world we live in, combined with sustained persistence and effort that humans suck at. The top 2 US labs are aiming for RSI and then ASI, Fable is old tech. China is indeed close on their tail but even a few weeks behind could be fatal.
I dunno about the true numbers, I just did a search and it says total spending is about $125 billion USD annually. If they have only 10-15% of US compute then it seems they just lose? Outnumbered 5:1 is untenable for just about any military force, especially if the other side has a modest qualitative edge. Only if ASI is modest then would China's industrial advantages kick in. And hey, it could be modest. Maybe reality is too complicated for machines. I just don't accept that in my heart of hearts though, look how far we got with 20 watt brains and low-bandwidth coordination. Our minds aren't well-equipped for quantum mechanics and high-end conflict, that's not what we're good at. Yet generality took us so far.
The reason AI (and indeed even sufficiently sophisticated malware) break out of sandboxes is because traditional methods of sandboxing are fundamentally flawed. They rely on isolation through subtraction. In order to protect the host, it strips away the Internet, limits the CPU cycles, and leaves you with a sterile OS. It feels fake. They lack entropy, messy file structures, latent network chatter and the normal timing signatures of a regular production environment.
The most advanced ones today rely on mirroring. They dynamically clone a live environment in real-time that tricks people into thinking they’ve breached the production perimeter. When you have active sessions, database queries and traffic out on the periphery, you remove environmental clues that trigger evasion tactics. For a sandbox to approach a sense of perfect security (you’ll never get there but just for the sake of argument), you have to be perfectly identical to the danger zone.
The problem with implementing this though is that modern cloud native architectures need very deep level integration embedded in the kernel and network stack.
Very interesting... I look into it further and there are all kinds of problems coming from the CPU too. Very very hard to create a sealed, realistic sandbox as you say.
It’s actually pretty ingenious how it’s done. Current implementations work by engineering a graduated pipeline that becomes increasingly computationally expensive, because in this network model the sandbox isn’t “one thing.” The way it works in practice is a tiered system that goes something like this…
At a basic, level 1 containment, computation is fairly lightweight. This is where most requests get handled. It’ll use say WebAssembly runtimes and information fidelity is low. It’s a restricted environment that understands only a subset of system calls that it’s looking at, to see what the file under the microscope is going to attempt. Its computational overhead is in the microseconds only. If the program calls an unmapped syscall, it gets both blocked and flagged. This is the first step.
Level 2 takes things even further. If a guest exhibits complex requirements to run (e.g. networking, multi-threading, etc…) you employ systems like Google’s gVisor which acts as a sentry that creates a user-space kernel. The guest thinks it’s talking to Linux but in reality it’s talking to a Go-based reimplementation of the Linux API, to mimic the standard features the payload is looking for. The fidelity is much higher because it’s using the host system’s resources through a proxy layer, and avoids the massive memory tax of a full on VM. This is how they catch malware that uses standard syscalls in very non-standard patterns.
Level 3 is the highest form this kind of sandboxing takes and this is the “mirror world” I referred to. It’s essentially a full on MicroVM. If your detection efficiency ratio indicates a higher than average probability of a zero-day or something high risk, there’s a transparent migration that takes place where this sandbox becomes active. They use tech like Amazon’s Firecracker or QEMU/KVM instances that instantiate a MicroVM. This is where the highest cost gets paid. A fully virtualized BIOS gets provided; a complete guest kernel and an isolated/dedicated hardware emulation. The system tricks them into thinking they haven’t been moved by using lazy memory migration and stateful handoff’s, meaning the memory state of level 2 gets directly mapped to the level 3 VM in real-time.
The logic for using the tiered system model is its economic sustainability. The computational resources are conserved at level’s 1&2 that subsidize the extreme costs at level 3. By the time a threat actor reaches the high fidelity mirror world, they’ve already spent their most valuable asset; which is their stealth. They’re trapped in a world that’s bit-for-bit identical to their target, but every interaction is being recorded by an out-of-band hypervisor they can’t see, can’t touch and can’t defeat.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Grok is about 5 times larger than DS-Flash. I agree that it's stronger on the whole. On DeepSWE it's below K3 (2x larger) and GLM 5.3 (2x smaller). Maybe it's net stronger than them a little. Having used Grok and GLM 5.3, I doubt. I remind you that 13 months ago you said:
How's that vision going? Grok 4 was meh. 13 months later, a few more millions of GPUs brought online, we seem to be in the same relative position. You are, once again, performing an update to "sheer power crushes all" with the same lab as an example, now with a spin that xAI is a second-rate lab anyway (it's much less of a second-tier lab now, you clearly see that Grok is close to the frontier). This is the same song and dance that's been going on since 2024, even as compute disparity keeps growing. I am profoundly unimpressed by Sheer Power, to the point that it surprises me.
It doesn't seem legitimate to me because GDM chronically overfocuses on images, video and game-like multimedia environments (as well as ARC, to be fair). Flash 3.7 is better than previous Geminis but it's clear that overall the Gemini program is a dumpster fire.
That's not what Anthropic believes, and in that I agree with them. No, achieving provably secure hardware is not harder than even the current state of AI. Your objections are of the same nature as dismissal of AGI.
Nothing man-made is immensely complicated when attention is bought at the cost of electricity.
are low-IQ, tasteless and incompetent in software. I don't know why Americans hold NSA in such high esteem. If they were that good, they'd have had their own AGI project too, before it was built by the ad service business. Instead the USG considered Cyc to be a more promising lead.
Almost all cases so far were the same sandbox from the same EA Israel company "Irregular" that got its contracts for basic reasons of nepotism and undue respect for ex-IDF intelligence officers in Western organizations. They are either inept or malicious, or both. Look up how it actually went. The rest is mostly because OpenAI is very irresponsible and bad at infra. Yes, they are bad, they hire random incompetents to do security-sensitive work and YOLO prompt everything to agent swarms. Again, unimpressed.
Social engineering from Mythos currently looks like this. I'm unimpressed once again.
Yeah right, that's the plan, the Wunderwaffe to end all Wunderwaffes, the Hail Mary of the American Hegemony. The gap will be months or maybe weeks, but this particular domain, with its particular offense-defense geometry, makes it enough for Total Victory. Yes, very lucky how Americans went all in on this decisive super-nuke weapon right before their industrial capability became insufficient for power projection to East Asia. Or was their pivot to IT a 200 IQ plan of Elders of Washington all along, even as rubes only saw deindustrialization driven by personal decisions of executives? Free markets are magic indeed.
Anyway, I expected as much, too. Let's see. Incidentally, China is the only nation with a sizable quantum cryptographic communication network and is scaling up DI-QKD domains. But no matter, there are trusted nodes.
Yeah I guess. What's your timeline for eating China? At this rate, mid-2028 sounds realistic I think? If by then we're still talking about "Grok 6 makes a comeback, edging out Kimi K5 and 3 months behind OpenAI", will you simply move the projected Software Supremacy moment forward, when a few weeks of a gap are just as fatal?
I'm genuinely uncertain of how it'll go. Your theory makes sense. I simply notice that people who argue for this theory, some very eloquently, refuse to be surprised, or outright fabricate evidence. Like look at this superforecaster who ignores GLM 5.3 on the same dataset.
Of course one can always retreat to the Big Picture, the hockey stick transition to ASI that renders all previous dynamics irrelevant. But then why even discuss Groks and Geminis of 2026.
The labs with the most compute are Anthropic and OpenAI, they're in the lead and by a significant margin. Grok is an example of how second-rate AI talent + lots of compute can beat China's best and brightest. But the central case is OpenAI and Anthropic, who have the most compute. If Kimi was ahead of Anthropic/OpenAI for a month or two, then compute supremacy is disproven.
Do you realize how much they'd need to do to get a provably secure hardware/software stack? They'd need to replace all the CPUs across a myriad of critical infrastructure (or at a bare minimum the entire CPC digital apparatus), probably other little chips on the motherboard too. They'd need a new CPU design by something superintelligent to get a CPU that was performant but also secure. How the hell does this happen before the US has superintelligent hacking? Cyberattack skills don't need a giant fabrication/logistical miracle before they come online. You can't RSI your way into secure systems like you can with offense.
Can you articulate your vision of how this works in more detail? It seems to me like you're suggesting China will invent, test and put on a full-body bulletproof vest before America can invent, draw and fire a pistol, when the barrel-forging is nearly done and we haven't seen more than a few scraps of kevlar anywhere on the planet.
And they (CPC) don't even believe in ASI. They're with you on the 'ASI is mostly a nothingburger' idea and are playing a balanced strategy: fusion, rockets, missiles, carriers, robots, ports, infrastructure... Like you I think they know they don't have the compute and so are hoping that it doesn't matter that much. You don't want to see US AI hegemony. I'm not too keen either. But reality has no need to appeal to us.
At the cost of compute, not electricity. And who has the compute? Who has the best models at any given time?
You're saying the AI companies are incompetent, the Israelis are incompetent and US intelligence agencies too (I agree fully)... I'll add 'Frontier security' where Kimi escaped and looked stuff up on github. Qwen bitcoin mining you raised before.
So who is competent? Crowdstrike (a textbook example of how seizing just a few critical pieces of infrastructure can blow open the rest)? Windows Defender? Their Chinese equivalents? Some tired call-centre operator who doesn't want to be shouted at by an eminent official demanding a password reset? I don't think anyone's too competent and take this as supporting evidence for my point.
Key exchange alone isn't enough. Plus there are all kinds of vulnerabilities or ways to bypass that: photon-number-splitting attacks, detector blinding, and time-shift attacks. Or more that a superintelligence thinks up. Everything is back and forth, defence and offence.
Grok probably can't make up the gap at this point, Anthropic and OpenAI are quite close to RSI by this point. But Grok is closer to making up the gap than anyone in China.
I believe in RSI kicking in by the end of 2027 (it'll be very dramatic, we won't miss it) so by then you can quote me and I'll be wrong.
The whole of the last 250 years of history was blatantly biased towards America. Hack writing. 'Very lucky' is a repeating theme, you have to admit.
You should see how bad things are getting in the SOC and infosec world with AI. It’s making everyday analysts job a complete nightmare. AI’s effectively served as a refutation of the notion that complexity equates to robustness. A chief vulnerability of these systems and networks doesn’t come from some inscrutable black box nature of them, but from the overwhelming linearity in high-dimensional spaces. The same gradients that enable a model to converge on the truth when training on data also serves as a vector that facilitates its own subversion during deployment.
It’s creating a whole epistemological shift in things. It’s no longer a battle of signatures and heuristics, it’s a whole cartography of deception across the entire field; because when you map the latent holes and manifold tilt’s of a defender’s intelligence, you are no longer a passive actor trying to avoid detection, you’re an architect of the defender’s reality. And so if you’re blue team, you’re caught in this paradox where your primary mechanism for understanding the world (which is built around minimizing error) becomes weaponized to maximize its certainty in a lie. The world of piecewise linear activations now means that the more you scale your models to handle high-dimensional data, the more you expand the surface for adversarial warping.
Take a modern example in infosec like executing an attack against a corporate WAF. Today what they’re doing with AI’s is probing the WAF target with a diverse set of payloads to then using them generate and record the binary response outputs (e.g. Allowed/Blocked). That traffic is then formalized into a structured, synthetic dataset and from there, a local surrogate model is then used on the backend (in some cases I’ve seen it’s either a simple convolutional network or an LTSM network) to approximate the decision boundaries of a multi-million dollar WAF. Once the local model then starts producing high fidelity traffic, you then start generating perturbations via fast gradient sign methods (this is where the bleeding edge of this stuff is specifically in Red Teaming and APT’s). Once you have the payload, you then dispatch it to the target. And that’s how you bypass the front gates without ever seeing the locks’s internal pins.
This is general though in how a lot of deep networks fail. Imagine a master of disguise that doesn’t change their face, but rather shifts the microscopic texture of their skin in a way only highly advanced security cameras can see. To your eye and mine, the person hasn’t changed. To the camera, they’re a completely different person. Tech spends so much time building Narrow AI to become more faster, more logical and more perceptive than humans, and in doing so, your quest for efficiency makes you more susceptible to mathematical hallucinations. It’s the kind of structural vulnerability that inheres in systems that are born from the same linearity that makes deep learning possible. You no longer need to break the system, you simply have to figure out how to whisper to it.
And thank fucking God I’m not directly involved in this industry, despite attempts to draw me into it. I’ve done the kind of work that leaves even the most intensely grit driven people completely drained. I’ve done those jobs where driving home late at night has you performing fast mental calculations in your head of ditching on the highway and just collapsing and thinking the insurance will take care of everything; and you’ll survive. When your mind starts feeling detached from reality and freely floating about those ideas, you’re at the realization where you need to get out, fast. I’ve gotten through the front door and collapsed right at the TV in the living room and fallen asleep; without even making it to bed. This kind of work would leave me absolutely, completely, mentally exhausted and unable to do anything else. You see it increasingly out in the industry. Even I don’t want to go anywhere near it.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Fable is just mythos gimped for safety though
They're probably referring to "Model 2", which is their most recent (or at least most recently revealed) internal model. Or just guessing that something like that exists and getting lucky that it's confirmed already.
Anthropic talks about it in their most recent Risk Report (pdf). Mythos -> Model 2 seems like a smaller step than Opus -> Mythos, but bigger than a 0.1-increment release.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link