chaiirmann (@chaiirmann) profile picture

currently training small language models

Joined April 2022
Photo by chaiirmann (@chaiirmann)Photo by chaiirmann (@chaiirmann)Photo by chaiirmann (@chaiirmann)Photo by chaiirmann (@chaiirmann)Photo by chaiirmann (@chaiirmann)Photo by chaiirmann (@chaiirmann)Photo by chaiirmann (@chaiirmann)Photo by chaiirmann (@chaiirmann)Photo by chaiirmann (@chaiirmann)Photo by chaiirmann (@chaiirmann)
chaiirmann retweeted
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.
1,816
6,410
2,191
41,159
6,648,639
Those who do great things must learn to be patient
20
What do you mean you’ve gone to prison 4 times? A thoroughly bred jailbird
“The hideout of danger0us people” Oh lord 💔💔💔
1
4
723
I just got an idea! Time to go back to my abandoned CV projects. Hey @roboflow
openai just released Decisions API, Jev-like "system 1" multi-modal model based on GPT6 Luna! Each egg gets cropped and sent to OpenAI (~270ms latency) and returns P(clean, dirty, cracked). One image crop consumes ~280 input tokens (~$0.03/1K imgs). @roboflow RF-DETR for detecting eggs
1
43
chaiirmann retweeted
Scan Qr code and apply for wl good luck !
861
999
339
1,622
107,991
chaiirmann retweeted
watched my people get pwnd by a simple link. so yh, let's save them.
5
5
1
59
4,164
“Make sure you don’t sleep,” one trafficker belted out. “If you sleep, you fall off. And if you fall off, you die!” Part 1 of my latest investigation for @fijnigeria. UNDERCOVER: Europe Through the Sahara Desert and the Mediterranean Sea (I) fij.ng/article/undercover-eu…
492
4,485
513
8,813
1,147,062
This quoted post is unavailable.
40
chaiirmann retweeted
We are announcing our most ambitious move today with the release of DAX-1, our flagship model, post-trained specifically for spreadsheet editing. DAX-1 is 97% cheaper to run than Anthropic’s Fable 5 and achieved 91.7% accuracy compared to Fable 5’s 58.3% on our benchmark. DAX-1 is also 5× faster. This marks a shift in how we build Decide. Rather than relying solely on general-purpose frontier models, we will train specialized models where it makes sense for our operations. This gives us more control over accuracy, speed, and cost, and allows us to build better products for the businesses and professionals who rely on Decide. DAX-1 is the first step in this direction
75
313
36
751
33,656
I have launched the actual test of this AI Glass that I built to help blind people and people with visual impairments I started testing it directly on the face of a person with a visual impairment to see how it would work in real life👓 And what amazed me the most is that after the blind person wore the glass, he said that for him, it was amazing Watch the video in English below ⤵️ This project of mine is getting close to where I want it to be And the world wants to see it in action
642
7,383
315
24,348
647,959
1/ Free-first multi-provider LLM router that never dies on rate limits. flippy automatically fails over across OpenRouter, freeinference, Groq, NVIDIA & Cloudflare — then runs agents, evals, and load tests with Loomweaver. Zero non-stdlib deps. Pure Python.
1
134
chaiirmann retweeted
this is how memecoin exposure should work. the amount of time wasted in dms trying to get a memecoin listed… this fixes all of it. bid, pay on-chain, verified automatically, live. that’s the whole flow
We built the memecoin exposure auction. One board. Bids from $5. Pay on-chain (SOL/ETH/Base/Monad) → verified automatically → your listing goes LIVE in seconds. No DMs. No admins. No waiting. verysketchy.lol
1
1
2
213
We built the memecoin exposure auction. One board. Bids from $5. Pay on-chain (SOL/ETH/Base/Monad) → verified automatically → your listing goes LIVE in seconds. No DMs. No admins. No waiting. verysketchy.lol
2
1
2
218
Your spot isn't safe. Anyone can outbid you at any second. Watching two projects fight over #1 in real time is half the fun. Clicks tracked per listing — see exactly what your rank buys you.
1
16
Nothing here is financial advice. It's called verysketchy.lol, not verysafe.lol. DYOR. Now take your spot → verysketchy.lol
6
Introducing verysketchy.lol. Outbid, but for crypto
Aaaaaaand we're live! outbid.lol No ads. No API keys. No revenue sharing. Just outbid your competitors to rank #1 and consider marketing done for today 🚀
1
79