Author. Coder. CTO. θηριομάχης. Building: https://xitter.cf/t.co/otXT4Wy6WR. Writing: https://xitter.cf/t.co/dBPBtyCIHw.
Austin, TX
Joined July 2008
- Tweets59.1K
- Following1.6K
- Followers34.9K
- Likes86.8K
We're always under attack. The real question right now is, to what degree are the present problems related to failures of defense.
1
12
998
Jon Stokes retweeted
The FAA Telecommunications Infrastructure (FTI) is a nationwide, secure managed network managed that connects FAA facilities and supports daily air traffic operations across the U.S. National Airspace System.
Why is it down? Some historical reasons below although the most popular ones among controllers messaging me are the new AI system is at fault or it’s a cyber attack.
Jon Stokes retweeted
“All airspace is currently being sterilized and centers are preparing for ATC Zero operations. Airborne’s should NOT be majorly effected— i.e, at least for now, this is not a “9/11 type of event” where airbornes are being forced to land at the closest airport available. All airborne aircraft are being allowed to continue on their filed routes of flight.”
Everything awesome in tech used to be "three to five years away" (it was a running joke at Ars back in the day), so total ASI singularity being perpetually 2yrs away does actually represent a speedup... of a sort
Seb's Law: The widespread, transformative societal impact of AI is always approximately two years away, regardless of the current year.
2
33
1,389
Jon Stokes retweeted
Very good post from @lugaricano and @alexolegimas!
"Exponential increases in capabilities requiring exponential increases in resources is at the heart of the original conceptualization of scaling laws. But the scaling laws are distinct from an intelligence explosion, which requires increasing returns to resources, i.e., fewer inputs to generate the same level of output."
siliconcontinent.com/p/are-w…
Jon Stokes retweeted
For context, all these Chinese AI training runs combined are order 1% of China's in-country AI compute, or equivalent to several weeks of Huawei's chip output. You could train all these models in parallel on maybe a 30MW Blackwell cluster. Call it 100MWs using Huawei.
Chinese AI companies are incredibly compute efficient, because they have to be.
I just don't see how export controls make much of a difference, to be honest. In the short term they're a small hassle for Chinese startups. In the long term they accelerate China's own AI hardware.
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.
It's gonna be cold enough this weekend in TX for me to wear my leather jacket, and I am gonna do it in honor of a true G who flew too close to the sun.
No this is not how resources are allocated in the world. It's how prices are set. Resources are sometimes allocated via market mechanisms (downstream of pricing), sometimes via government policy (downstream of voting), sometimes via violence, etc. What are doing here ppl?
resources are allocated in the world according to supply and demand, and implicitly we have always understood that to be *human* demand. I think the best way to understand gradual disempowerment is the fear that resources will be allocated more and more according to machine demand rather than human. I’m not saying I think this outcome is my prediction, but this seems like the best way to understand the fear people have. It may well require deliberate effort to avoid this outcome.
Probably related to recent news:
Look I think I've figured this whole thing out. Follow along as I try to steelman, and tell me where I'm wrong.
OpenAI guys on the TL believe that if they can't sell metered inference tokens at a sufficient markup, then they will not have enough of a business to fund the next big training run.
They are surely correct about this.
They believe that if people release powerful open models, this will probably fatally impact their ability to sell inference tokens at enough of a markup to fund the next big training run.
They are surely correct about this, too.
They also think that if they cannot fund the next big training run (again, by selling inference tokens at a markup), then NOBODY will be able to fund the next big training run because it means there's no money in it.
This last bit seems to me & many others to be not just wrong, but totally bananas in a "guy, have seen the actual software industry and how it works in real life?!" kind of way.
There are a lot of ways to monetize software out there in the world. Insofar as inference can add new capabilities to software, there will be lots of ways to monetize it.
In other words, if you're telling me, "we can't have a business selling inference if X or Y thing keeps happening," then my only response is, "ok well that sucks for you... sounds like that's a terrible business."
But if you're telling me that "selling metered inference tokens is a terrible business" is tantamount to "nobody will fund big training runs that are upstream of more effective & economically valuable inference tokens", then I think you are extremely wrong and should get out more and learn about other parts of the software ecosystem.
Workplace automation is huge and will be even bigger in the future as models get better. You can sell workplace automation very profitably in lots of different packages (depending on the workplace and the type of automation). Like, I'm sorry that you really really want to be in the metered inference token business and not the workplace automation business, but them's the breaks. The market wants what the market wants. We all need to live in reality and not beg for Uncle Sam to save us all from open source -- because that was already tried and it didn't work.
Jon Stokes retweeted
Following his ouster as CEO of Nvidia Jensen Huang is reportedly considered for a high executive position at Mistral. Some experts, however, warn that for a high-profile position in France, two affairs might not be enough.
I buy some version of this. The reality is that the "stop AI" agenda & Anthropic's desire to have a global monopoly on inference are almost perfectly aligned. Anthropic wants to control all of AI, & a single point of control is necessary for "pacing" or pause.
My favourite, totally speculative theory, is that the Haung chip rat lines operated with the blessing of the White House. Alice was the fixer, but the affair itself was a cover. The rationale for this was a quiet quid pro quo for access to rare earths and other essentials, while maintaining the optics of being tough on China.
Moreover, the economic circularity of the arrangement suited both sides. Without it, the main narrative for growth collapses both in the West and in China.
The fact the whole thing was blown open is connected to the Dems coming back into power, and wanting to shut down the AI race. Their agenda is to lock ai down and take it fully in house, citing safety concerns. But everyone knows that is unlikely to happen in practice. A lockdown means asymmetric tech advantage for the US military industrial complex. That is concerning for everyone on the planet as it destabilies the MAD balance that has prevented nuclear annihilation so far.
I think on that basis it probably served the interests of the world for some of these models to be "distilled" by China, and released into the wild. Without deepseek and Kimi the asymmetry would be far worse.
The only thing standing in the way of Chinese, French or even British labs catching up with American labs is access to chips. Blow up Jensen and the rat lines, and the power grab is complete.
Even if it's not the truth, I think it's a good story. And it makes even more sense when you frame it against the attempted ousting of Sam Altman in 2023, and the appointment of Larry Summers to OpenAI, PLUS, the latest revelations about Dario Amodei's wife.
If you are someone on my feed who comments on AI but who has been at such pains to be taken seriously by the cool kids that you will not speak out against the doommongering, then congrats you're part of the problem.
1
15
630
Jon Stokes retweeted
so:
- no succession plan.
- 60 reports which means no obvious person with complete understanding of the co operates. or at least it’ll be insanely messy. potential for a lot of infighting.
- a weak board suddenly tasked to manage a crazy crises w/ arguably the most important co of this era.
- possible talent exodus since founder is leaving.
- a ceo search that could take a while with no obvious outsider candidate at the top.
- potentially more revelations.
- jensen will have zero involvement going forward. no chairman role.
- downstream impact on ai ecosystem from funding, support etc.
this is not good. what am i missing?
Jon Stokes retweeted
The real pressure on NVDA isn’t just the Jensen drama or macro noise, it’s Anthropic quietly driving the policy fight.
Their incentives are structurally opposed.
For Anthropic to win and justify a sky-high valuation, they need high-margin closed-source frontier models to dominate, with open-weight models heavily restricted or de facto banned (or at least sandboxed to the point they’re not competitive).
Wide availability of strong open models collapses pricing power, enables local/self-hosted use, and turns inference into a commodity that anyone with GPUs can serve.
For NVIDIA to keep compounding, the opposite is required: open models everywhere, running on as much hardware as possible. More open weights = more developers, more fine-tunes, more agents, more total tokens, more GPUs sold.
Jensen’s public push for open weights and the Open Secure AI Alliance wasn’t charity, it was core to the business model.
When one side needs the open ecosystem throttled and the other needs it to explode, the “safety” and “national security” arguments start looking a lot like competitive strategy.
Show me the incentive and I’ll show you the outcome.
Jon Stokes retweeted
This thread is astonishing. The most puzzling aspect yo me is not even the incompetence and paralysis of the autorities but the resignation and fatalism. It feels like reading some chronicle by a medieval peasant ...
