Stephen

[Stephen @stephenwalker]

“We are choked with news, and starved of history.” — Will Durant

https://t.co/04gpLaKfGQ

[Stephen @stephenwalker]

when jensen pulls off his second act

he proves the moat is the great man, talent, and network of relationships

everything else is imagined

[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.

[Stephen @stephenwalker]

@flyosity This is cool, but I want a monochrome one please :) remember when iTunes was blue? That one was cool too

[Vercel @vercel]

Introducing the UI SDK.

[Stephen @stephenwalker]

@Greglaar @vercel @grok Use vercel to host your web apps/sites

[Stephen @stephenwalker]

@awoooouwuooooo @ns123abc They werent

[Stephen @stephenwalker]

Pong

[Richard Davey @photonstorm]

https://t.co/lw7LWv7LVM

[Stephen @stephenwalker]

it's even worse and won't come as a surprise: his primary mode of "writing" is borrowing from others, continuing to present in non-academic works -- he's neither an original thinker or researcher

in 2020, he lifts phrasing and key arguments from a book of the same title, and never mentions the author as as source for this argumentation

https://t.co/mZNLmvTQgp

https://t.co/041wXbJPAX

in 2022, he uses this blogspot's attribution and phrasing of a quote that doesn't exist in the source material, he's reusing the third party content instead of going directly to the source himself

https://t.co/vgvSPumGxL

https://t.co/AKXrlKBQDH

https://t.co/E8UAWaEBnF

in 2026, his infographics on palimpsest manuscripts borrows figures and bibliography from got questions, including getting the book title wrong, and he reuses HMML Palimpsest Project's phrasing

https://t.co/fFkswpIdHC

https://t.co/hCQwlNCcDa

https://t.co/DiGmB4fbuA

[Stephen @stephenwalker]

@patrick_oshag @eladgil Never underestimate the value of vanity

[Stephen @stephenwalker]

@JosephRRusso @davidu Yes, however, it’s important not to be dismissive to understand their motivations and the playbook he ran to get their attention

[Stephen @stephenwalker]

I use Codex for 99% of email with an EA automation that ends up trashing, moving to spam, and unsubscribing from most mail

It took a long time to get the logic to what I want and teach the model how to properly validate cold outreach, but can’t imagine any traditional client competing on that level of intelligence… just use Mail app and point your assistant at it

[Stephen @stephenwalker]

@lulumeservey The 9s are the nuclear gambit?

[Theo Bearman @theobearman]

Anthropic has confirmed in an updated S-1 today that it has determined that the Autonomy 2 capability threshold has been met, based on the “dramatic acceleration of overall AI R&D progress” condition, given that it applies where Anthropic "observes or expects" accelerated progress. Anthropic further states that it does not make this determination based on current models.

https://t.co/HdbiIM0nGk

[SemiAnalysis @SemiAnalysis_]

Interestingly, Fable 5.2 scores about the same as Sonnet 5.5 on RSI. Anthropic explicitly limited advancements in AI research during training, and it's cool to see GDPx accurately capture both the SWE improvements and plateau in RSI abilities. (4/7)

[Stephen @stephenwalker]

@Idaho4570 @KatieMiller @elonmusk @grok ookkyyoouuaarreerriigghhtt

[Stephen @stephenwalker]

Thankfully Codex will manage this.

[Stephen @stephenwalker]

Claude needs approval to work on artifacts that you asked it to make... this is the stuff that keeps assistants from going mainstream

[Stephen @stephenwalker]

@theo wdyt about using opus and calling fable for review? fable seems valuable for an intelligence level opus is not capable of, but clearly not as valuable for all work.

[Stephen @stephenwalker]

@jsngr @chenzhangreal @mainframe Ai reply

[cody @codycollier]

88x31.gif