A forensic report suggests that six open-weight large language models from Chinese labs may have been trained on NVIDIA Blackwell GPUs, which require a U.S. export license for China. The report, published by Will DePue at Unemployment Labs, identifies a "fingerprint" in the models' weights that aligns with the specific random number generator seeds and SM layouts characteristic of Blackwell chips. While the report carefully states this "supports, without proving" Blackwell's involvement, the evidence presented includes matches on held-out weights and verification on physical hardware, with some models showing strong statistical significance (up to 1,385 sigma) for the Blackwell fingerprint. AI
IMPACT Suggests potential circumvention of US export controls on advanced AI hardware, impacting global AI development and supply chains.
RANK_REASON Forensic analysis of LLM training fingerprints published by a researcher. [lever_c_demoted from research: ic=1 ai=1.0]
- DeepSeek
- GB200
- Hopper
- Kimi K3
- MiniMax-M3
- NVIDIA
- NVIDIA Blackwell B200
- NVIDIA H100
- Peter Liu
- PyTorch
- Unemployment Labs
- Will Depue
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →