Elsor

Supply chain. Software Distillation and Distribution. Local. Secure. Sovereign. Decentralized. Unhappy Cowboys fan. Bullish on America.

[ttywisp @ttywisp]

>US intelligence agencies knew about Alice Huang the whole time
>Rubio literally already briefed Trump on it
>suddenly drops in the NYT on a weekend
>board immediately brings in former SDNY acting attorney Joon Kim and Wachtell Lipton
>$7.5 billion in chip shipments to China linked to her comes out 48 hours later
>Schumer and GOP senators instantly reading from the exact same script demanding his head

If you think this is just "oops leather jacket man got his dick wet with a Fang Fang" you are legitimately retarded.

The spook agencies sit on this kind of kompromat until they need to pull the trigger. Huang has been dragging his feet on the export controls and trying to keep the China market open through Singapore backdoors because he's a merchant first and doesn't give a fuck about the State Department's pivot. The deep state wants total decoupling or at least direct DoD/intel oversight over the compute stack. You can't nationalize Nvidia outright without crashing the market, so you burn the founder-CEO with an MSS honeypot op you probably watched happen in real time on NSA servers, tank the stock $400B so BlackRock and friends can buy the dip, and install a compliant spook-approved managerial suite via Wachtell.

Look at what happened to TikTok, look at how they forced Intel into a quasi-state ward. Nvidia is the single most critical bottleneck for the surveillance/drone state. They were never going to let an independent founder keep running it like his personal fiefdom once AI hit the national security threshold.

He's cooked. Enjoy your new CIA-vetted CEO.

[Omri Weinstein @WeinsteinOmri]

KV cache compression is one of the most important and competitive areas of AI infrastructure as it determines how many concurrent inference sessions your GPUs can serve.

Pearl Research now has the most efficient KV cache quantization on earth, all the way down to the CUDA kernels! We took a genuinely different, lattice-based approach to this problem, wasting 30% fewer bits than Google's TurboQuant and reducing native FP8 KV memory by 65% on GLM5.3.

[Elsor @ElsorThomas]

you can’t say things like this out loud man.

[Aporia @aporia9n]

this collective celebration of the Outlook outage might be the best advertisement for AI replacing white collar work

we've spent decades turning millions of educated, ambitious ppl into glorified information processing machines

spending 10h/day answering emails, updating spreadsheets, preparing slides, forwarding information, chasing approvals, attending calls about other calls etc.

and somehow we're surprised they celebrate like it's Christmas when the whole system breaks

obv losing your salary is very different from getting an unexpected day off. But we're so obsessed with the idea of machines replacing humans that we forget how much of modern white collar work is already humans behaving like machines

the real dystopia was making humans spend 40 years doing work that machines should have been doing all along

[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.

[Vishal @KyrieBlunders]

Of course that's your contention.

You're a first-year RL engineer.

You just read Prime Intellect's prime-rl writeup and saw rollout workers streaming trajectories into a trainer that never waits on them, so naturally you believe decoupling generation from training is free throughput and staleness is a rounding error.

Next month, when you get your hands on an RL cookbook, the Unsloth GRPO handbook, probably, you'll be singing from the treetops about colocating vLLM inside the training process, and the 400 milliseconds you save on every weight sync by waking and sleeping the optimizer like you have not just moved the idle GPU somewhere else.

That'll last until sometime next year, when you watch a livestream of Will Brown talking about verifiers and become convinced that none of the systems work matters, because an environment is just a Python function and the reward figures itself out.

"Well, as a matter of fact, I won't, because a verifier..."

"A verifier is exactly the thing your policy learns to game first. It finds the one branch you forgot to test and farms it for maximum reward in fifty steps, and your async sampler is already three policy versions stale while it does it, so the importance ratio is fiction by the time the update lands."

Yeah, I read that too. Every training-inference mismatch thread since the sampler's logprobs and the trainer's logprobs first disagreed in BF16 and nobody noticed for 200 steps.

Do you have any thoughts of your own on the subject, or should I just point my agent at the prime-rl repo, set group size to 64, and let it pay an H100 to argue with itself for me?

[Elsor @ElsorThomas]

@elonmusk

[Microsoft 365 Status @MSFT365Status]

We're aggressively pursuing all pathways to reduce impact. We've paused all deployments, we're moving additional traffic off the affected infrastructure, and we're restarting affected servers to return them to service. More info under EX1550543 and MO1550821 in the Microsoft 365 admin center, as well as https://t.co/pYdahqxDv3 if the admin center is not available to you.

[Mark Ajzenstadt @mardehaym]

https://t.co/xIkjB5Xd14

[Elsor @ElsorThomas]

This is the way

[Polymarket @Polymarket]

JUST IN: SpaceX commits $220 million to upgrade water infrastructure in South Texas as it builds its $24 billion AI supercomputing site.

[Conks @conksresearch]

the plumber does not concern itself with CEOs

[Elsor @ElsorThomas]

@dcdnews The cap is 25MW, so it bans hyperscale, not data centers. And the Proximity Park deal had already died after 18+ months without a development agreement. With 40+ similar measures passing this cycle, ballot risk is now a site-diligence line item.

[Elsor @ElsorThomas]

@dcdnews The town memo is the pitch: about $27.8M a year in property tax at full build, against a total levy near $32M. But no application is filed yet, power and water details are still to come, and the state now requires local approval before permits. Site control comes first.

[Elsor @ElsorThomas]

@dcdnews The Lambda line is the real tell. Per the amended S-1, the $34.9B deal became terminable because Lambda didn't secure the facility lease by a set date. A compute contract is only as firm as the building under it. Site control is the credit risk, not the GPUs.

[Elsor @ElsorThomas]

The utopian dream of much of the tech world must start with tangible benefits for the local population. Cover all electrical costs for the city, pay for school lunches for every child, open up daycares, job training and placement orgs, and actual financial incentives. All of that, is a drop in the bucket against the return.

[Naithan Jones @NaithanJones]

On its surface this looks silly and easy to make fun of but what data center builders are not clocking is:

in the minds of people who live in these places data centers represent massive wealth creation for someone who doesn’t live there. While they don’t get to directly participate. It *feels* like extraction.

These are working class people, some at the margins, so pushing back *feels* like taking power back. These are people who have been promised the run and the moon by new Casino projects, jobs and property values from a new highway overpass projects, all of which did nothing and in some cases harmed their community. They’re jaded and have heard it all

If data centers actually want to get their projects built in communities like this, hate to say it, but they are going to financially incentivize the people there. Do things for them *before* any shovels. Communicate that these benefits will continue as the project succeeds. Give each house hold some financial or quality of life incentive to support it. Not theoretical, tangible and presently available

Just my 2 cents

[Elsor @ElsorThomas]

Cataclysmic for American AI Dev Ops and therefor everything downstream.

[bebu @bebuu_ai]

Wild rumor coming out of China this morning on DeepSeek V5:

DeepSeek is reportedly already pretraining V5 as a ~2 Trillion parameter ultra-sparse MoE model.

And instead of training on NVIDIA GPUs, they are reportedly running the V5 pretraining run on Huawei's new Ascend 950 / 950DT chips with a team of Huawei kernel engineers physically working inside DeepSeek's offices to keep the cluster humming.

If DeepSeek pulls off a 2-trillion-parameter frontier run natively on Ascend 950 silicon, that changes the entire AI hardware story.

[Naval @naval]

Datacenters are a globally competitive market. Someone, somewhere, will build them.

[Sergey Gorbunov @sergey_nog]

goodbye audits. hello formal verification.

[Timothy Lewis @tolewis]

IF I understand what just happened - we have been thinking about inference design all wrong for local AI.

Concurrency doesn’t need to work in other context windows, use the one window you have.

🫡 @ashxhart

[SemiAnalysis @SemiAnalysis_]

We visited Galaxy, Envision's 800MW wind-powered datacenter campus in Ulanqab, Inner Mongolia. Its first 120MW phase went from bare land to delivery within a year. (1/6)🧵

[Elsor @ElsorThomas]

“Time to make the mind of a sentient sun”

[NIK @ns123abc]

🚨 BREAKING: SpaceX just locked up 100+ GW of solar cell manufacturing equipment from German toolmaker SINGULUS to build STARMIND AI satellites

The first turnkey heterojunction line arrives in Bastrop, Texas in H1 2027, running wet chemistry, vacuum coating and thermal processing under one roof.

Elon Musk says most of the mass of a Starmind satellite is solar panels, meaning the solar cell factory is the satellite factory.

100 GW is enough solar for 400,000 Starmind sats

“Time to make the mind of a sentient sun.”