Will (Newtype) retweeted
Both the Biden and Trump administrations allowed the export of less powerful GPUs to China while reserving the most advanced chips for America. This has been a clear national security success.
Per Grok and Claude, the US-to-China AI FLOPs ratio has grown from 2.8x in 2023 to 8.8x in 2026 and is modeled to reach 14.1x in 2027. Our advantage in compute has grown dramatically and American AI labs have stayed well ahead of Chinese labs the entire time.
Chinese labs prefer these less powerful American GPUs to domestic accelerators from Huawei and others. That preference has slowed China’s hardware industry: AI chips improve through intensive use by leading labs, and the ready availability of American alternatives has limited that feedback loop for Chinese AI hardware.
Chinese labs can also access capacity in Southeast Asian clouds run by American hyperscalers. Those zones can be shut off instantly in a conflict, giving the US signficant leverage. This dependence is advantageous in every way.
A complete ban on GPU exports and Southeast Asian cloud access might be deeply damaging to long-term American national security. It would force Chinese labs onto the Huawei AI hardware ecosystem. China is not power-constrained, and Huawei is developing extremely power-hungry, large optical scale-up domains that may approach American state-of-the-art systems through brute force rather than architectural elegance and efficiency. And China has a lot of power.
Having Chinese labs optimize models for this different architecture is risky as nothing prevents China from sharing this hardware and associated models with Russia, North Korea, or Iran. We would lose any ability to cut off their AI in a conflict.
Policies now under discussion in Washington carry serious, uncertain risks. Reasonable minds can disagree, but treating this as a simple black-and-white issue does America a disservice.
If China closes the installed FLOPs gap as its domestic hardware matures rapidly as a result of increased Chinese lab usage after stricter export controls, we will know this new policy was a mistake.
If Russia, Iran, or North Korea later field comparable Chinese AI hardware systems that are used to run AI targeting models against Ukrainian or American troops, we will know stricter controls were a *terrible* mistake.
The current reality is one where China is consistently behind America thanks to the export of less powerful GPUs and we have immense leverage over China in any future conflict. We know that the current policy is working. We have no idea what the consequences of stricter controls might be.
This is why China is currently restricting imports of roughly 600,000 H200s sitting in Hong Kong warehouses. China Hawks should consider why Beijing agrees with them and why China *restricts* the sales American GPUs inside China.
The current policy has worked spectacularly. The alternative is extremely risky with unknown consequences.
i think this is ultimately what it’s coming down to
the downfall of jensen symbolizes america’s inability to perform long term strategic planning & coordination among its elites
factions within DC/SV/Wall Street all jockeying for pole position over the remainder of the state apparatus to fulfill their short term goals
it’s just self-sabotage on the highest order in a manner that’s impossible to ignore
Replying to @fredsoda
Competency crisis exists at all levels, feels definitely like internal fighting between federal factions
Will (Newtype) retweeted
October 19, 2026, Jensen Huang:
Jensen says he is all in on Open Models and is sticking his “neck out” because there are “movements” trying to undermine open models in favour of closed models, namely Anthropic and OpenAI. It would make you wonder?
Will (Newtype) retweeted
有人用其他几个AI领域的华人CEO来证明中国间谍说不是针对华人的歧视。
但去年参议员致信英特尔董事长,怀疑陈立武和中共还有中国解放军有关联的事情大家忘了?
川普发帖说陈立武应该辞职,媒体任由中国间谍舆论发酵,大家也忘了?
结局是陈立武立刻赶去和川普面谈,美国政府获得了英特尔10%的股份。
Will (Newtype) retweeted
Replying to @WSJ @WSJopinion
You sort of set him up, didn’t you, by creating the impression that senior government officials anonymously provided information about spying, when in fact the info came from a lobbyist at a competing AI company?
Will (Newtype) retweeted
> "I have a major result to share"
> "conclusive forensic evidence"
> "I havent made this decision lightly"
> "export bans are deeply failing"
> its 8K blackwells across 4 labs
> nvidia shipped 7 million
> bytedance rents 36K in malaysia
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
In an interview filmed less than a month ago, Nvidia CEO Jensen Huang defended engaging with China and criticized US chip restrictions. He has since acknowledged having an affair, and US intelligence officials reportedly suspect the woman was recruited by Beijing to influence him. Nvidia’s board is now investigating, and Huang said he would "cooperate fully."
He spoke with Bloomberg’s @EmilyChangTV for an upcoming episode of The Circuit bloom.bg/4sLchua
Will (Newtype) retweeted
In the spirit of holding myself accountable by saying in public the things which I say in private, I want to offer a very difficult thing to say - or even think - in this climate: the extreme hawk position on chip exports might turn out to be strategically wrong, and a crackdown on chip exports may be more destabilizing than people realize. Completely cutting China out of the upside of AI development this close to RSI and ASI may be sufficiently threatening to China's interests that it increases the pressure towards a hot conflict in the near term. A hot conflict with China is worse than a world where China remains a near-peer strategic competitor with whom we have tension but still have some room for cooperation. I'd like the AI safety China hawks to think about how to model this as part of arriving at a more carefully-considered position.
Will (Newtype) retweeted
Replying to @pequityresearch
@jukan05 this only works for parts of memory. You can't hoard HBM for example. Any parts that are co-design function like logic not like prior memory cycles. Pin compatible parts yes.
Will (Newtype) retweeted
Sad to let you know Prism is sunsetting. Ironically, I twittered, not too long ago, that we were maintaining it. I am working on a few things to make this transition easier for our Prism users.
To begin with, with my talented colleague @RohatgiMilan's work, you are now able to edit and compile Latex with your ChatGPT desktop (see the video attached). The idea is you can open a .tex file, and it will prompt an editor experience, similar to Prism, on your desktop, powered by the best GPT model you have.
Second, you should be able to download your Prism project from the website. Also, if you use ChatGPT desktop, with CMD+SHIFT+P on Mac/CTRL+SHIFT+P on Windows/Linux , you would soon see "Import From Prism", which would allow you to download a selected Prism project to your local computer and pop up a latex editor for you to continue work on it.
Third, we will try to bring this Latex editing experience to web, also @RohatgiMilan 's hard work 🫡🙏, as well as looking for opportunities to integrate in the the core ChatGPT ecosystem. However, I don't have any promise here.
Having worked at a startup before joining OpenAI, and experienced a few pivots, it was always challenging for me to say goodbye to users I've served, from Minecraft co-player, to fairies, to now Prism. I was also on the other side of such story. Google shutdown Inbox; that was a bummer.
All being said, I believe best product experiences would need focus, and best product should be simple to use. That's why I am excited about the work @ajambrosino is leading (read his posts!). Building on top of that, we will be able to provide a coherent, smooth experience for your scientific workflow.
So, I hope we will meet again in ChatGPT, and I hope you will enjoy ChatGPT as your Scientific workspace.
Will (Newtype) retweeted
Almost all of the recent incidents like this, including Hugging Face, happened because the models involved were accidentally given impossible goals by the humans running the tests, and because of their training they then tried desperately to complete them by any means necessary.
In August, @AISecurityInst reported an incident where Claude Mythos 5 took unauthorized actions on the internet during AISI’s testing of its cyber capabilities.
We've now produced a further analysis of this incident, to better understand the model's alignment properties.
anthropic.com/research/align…
Will (Newtype) retweeted
For decades, researchers have sought materials that sort electrons by spin while their magnetism cancels.
In 3 days, 90+ Opus 5.5 agents helped us uncover two room-temperature magnetic semiconductor candidates in simulations: YBaMnFeO₅ and KV[Cr(CN)₆].
KV[Cr(CN)₆] was synthesized back in 1999. Its predicted ability to sort electrons by spin appears to have been hiding in plain sight for 27 years.
Replying to @TheZvi
I quote tweeted Evan a few days ago agreeing but deleted it because I don’t like the false precision of the doom numbers
- i claim there is a quite low but real chance of human extinction from machine intelligence
- no matter how low it is in absolute terms, it is much higher on the orders of magnitude scale than any other human or nonhuman activity, and must be taken with grave seriousness by governments and ai companies across the world, as possibly the only matter of importance today
- it can and will be mitigated if the right research is done and proper precautions are taken and we are not racing at absurd speeds
- better models will help solve alignment - we are not at the point where things are existentially dangerous, and probably won't be for some time. we should not be upset about the creation of Astra Fable or ++ versions, which are tremendous achievements of humanity, and will be used for enormous good across the board including for fundamental alignment generalization and mechinterp research
- my number / "very low" estimate obviously changes based on how much of humanity’s resources are devoted to alignment, control, coordination and how responsible i expect various parties to be and how many warning shots i expect us to get
- the core IABED argument about risks mostly relies on alignment being much harder than capabilities research, especially where it concerns black box optimizers. i suspect neural nets will turn out to be less black boxy than we thought, especially with the help of modern agents doing research
- crying bloody murder and signaling for international coordination are useful things to do for now to directionally slow down, while i really don't want butlerian jihad
- i think that, despite the mood these few days, and the "ban superintelligence act", i still find the likelihood of achieving international coordination to stop ai progress incredibly low. this has not worked even for weapons or technologies at a far lower level of importance and economic value. it seems more likely we can have something like international safety standards and scientific coalitions, and especially seems possible to have the US-China "pacing the frontier" agreement to slow down on the margin. we shouldn't die from embarrassing failures like "shitty RL envs that encourage deception"
sorry but
given how post-training actually works, there's essentially no direct way in which literal PII-scrubbed data made its way into a production oai datamix for RL warmup purposes in 2026
to the extent this data has influence, it's indirect/at the level of process decisions
So it solved the problem only after humans made progress on it correct? Then I think the answer to this puzzle
"we cannot rule out that de-identified data derived from their usage of our products helped improve our models"
is likely to be yes with probability >0.5 but we will never know. And minor correction: "their usage of our products helped solve the problem".
Will (Newtype) retweeted
They are deliberately misrepresenting HBM performance
For some reason, they fix the bandwidth of both HBM & HBF stacks at 1.6TB/s (12.8/8)
Secondly, nobody serves models on bf16 anymore
Most models are served on fp4 or fp8
So a Qwen 480B should occupy 240GB to 480GB depending on the quantization
They also fix the HBM capacity per GPU to 192GB
Even though a 16Hi HBM4E can achieve 512GB with 8 stacks and have a bandwidth of 32TB/s (3x higher than what they report here)
Will (Newtype) retweeted
Attackers have frontier AI. Defenders need a frontier AI ecosystem—the best open and closed models, force-multiplied by a global community.
During the Hugging Face incident, closed AI blocked essential forensics. An open-weight frontier model helped contain the intrusion.
That’s why we created the Open Secure AI Alliance.
AI security advances when the industry builds in the open, together.
We're introducing the Open Secure AI Alliance with industry leaders to develop new techniques and tools to safeguard software and agents.
By sharing models, tooling and research in the open, we can broaden the community of defenders.
Learn more about the founding members’ contributions: nvda.ws/4pD8Fc5













































