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?
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
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link