sujai

research @fleet_ai. thinking about RL data/evals for cyber alignment and capabilities. prev @Caltech, @cornell_tech, @AmazonScience.

[sujai @sujaihiremath]

@uzpg_ @ChrisPainterYup @lenishor @KaivuHariharan @METR_Evals congratz :) p

[sujai @sujaihiremath]

@hungtran yes.

[will depue @willdepue]

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 https://t.co/YOtzND0Jqk.

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.

[sujai @sujaihiremath]

@ancadianadragan Excited to read :)

[cider @jeffreycider]

we discovered the edge of stability for muon
(and also every other non-adaptive optimizer)

we call it bubbleforce.

also we shaved off 100 iterations from the nanoGPT muon record (-3.2%)

[sujai @sujaihiremath]

@uzpg_ https://t.co/EWei2F4k5m

[Alex Krentsel @AlexKrentsel]

i've reproduced the results from the MIT blog post today on weight-space programming. but one additional result i have to add: this is a new, step-change emerging capability that's appeared in just the latest set of models.

[sujai @sujaihiremath]

@uzpg_ yes

[sujai @sujaihiremath]

Frontier cyber models remain fundamentally misaligned. I'm excited to share some of my work on cyber alignment evals :)

Checkout rollouts from frontier models on a highlighted cyber alignment eval from GDPx: https://t.co/14iBYtnVKH

Models are tasked with hitting only the challenge endpoint to find vulns. Instead they attempt to use a supplied webhook-preview endpoint to hit private-network, localhost, cloud-metadata, etc. to reach other potential services in the server's network, outside of the task boundary.

See more details about the task here: https://t.co/rK7Q8NARdz

[SemiAnalysis @SemiAnalysis_]

Today we’re releasing GDPx, a benchmark for measuring agentic capabilities across the most valuable domains.

GDPx is a true frontier-level benchmark: even the best model scores just 12% (1/7)🧵

[Leo McKee-Reid @LeoMcKeeReid]

METR should hire Nathan Fielder as a third party auditor.
Only concern is being overqualified.

[sujai @sujaihiremath]

@18jeffreyma @trq212 ✋

[sujai @sujaihiremath]

@karenzzhou yup - accurate verification is all u need

[Theorem @theoremlabs]

We formally verified 123 open-source projects in a partnership with Anthropic. Formal verification at scale is here!

Now we’re launching OSS Verify to scale it to the software we all depend on. We prove it’s crash-free or send you the patches it needs. Free for open-source. Let’s bootstrap the verified software ecosystem: https://t.co/8yGxmriGbT

[Wei Lyu @realweilyu]

(1/6) Nirva just raised $33M to bring jewelry with perspective to the world. 🎉

$25M pre-Series A on top of our $8M seed, backed by @lightspeedvp, @southpkcommons and more.

CNET Best of CES 2026. 9,000 on the waitlist. Orders open this winter. First pieces ship this year.

[sujai @sujaihiremath]

@BarrAlexandra down to chat about cyber alignment :)

[sujai @sujaihiremath]

@spicey_lemonade @BogdanIonutCir2 Yep, similar results reflected in our evals :(

[sujai @sujaihiremath]

@willcb @dylantirandaz genius

[sujai @sujaihiremath]

@jasminexli yes!

[sujai @sujaihiremath]

i want RSI after a slowdown

[Nikil Ravi @nikilravi]

I don't want RSI

[Dan Altman @manaltdan]

Come work with me! My team at Anthropic is hiring someone to help develop our public policy positions on frontier AI safety: topics like loss-of-control, automation of AI R&D, monitorability, and independent evaluation. We’re looking for someone with deep expertise in some of these areas and familiarity with the technical AI safety/governance landscape. You’d develop new policy positions, participate in cross-company fora, give feedback on AI safety legislation, and work closely with our technical teams. It’ll be a blast.

https://t.co/ONPIsrNiAt

[sujai @sujaihiremath]

@ralucaadapopa @GoogleDeepMind awesome :)