- What happened
- Former OpenAI researcher Will DePue says six Chinese AI models, including DeepSeek's, carry a numerical fingerprint matching Nvidia's Blackwell chip, which is barred from export to China. Some of a model's random starting numbers survive training and carry a pattern from the chip that generated them. DePue says he reproduced a DeepSeek model's untouched starting numbers bit for bit on Blackwell, and his method agreed with 24 cases where developers named their hardware. He says MiniMax's April model didn't match Blackwell but its June model did. Ant Group posted Blackwell test logs dated March 31 on its own GitHub. The fingerprint can't show where a chip was located or prove anyone broke export law, and DePue says labs can erase many of these signatures. The Chip Security Act would track chip locations and require diversion reports, and the Remote Access Security Act would oversee chips rented to China through overseas data centers. Nvidia opposes the main chip bill in the defense package, and Congress hasn't agreed on the final defense bill.
- Austin's take
- Austin backs chip tracking legislation in direction: track shipments abroad, know your customer, and require diversion reports, rather than relying on random numbers to work border patrol. He adds that the US shouldn't police its own citizens the same way.
- Why it matters
- It's a rare technical clue about whether export controls are working. Publishing the method could also help labs hide future use.
- What to do
- Treat the fingerprint as evidence, not a verdict, and watch what happens to the Chip Security Act in the defense bill.
- Not confirmed yet
- DePue says his match supports Blackwell training without proving it. It can't show chip location or legal violations, and Chinese labs could be using chips abroad.
From the Friday, November 13 show
Researcher says Chinese AI models carry banned Nvidia Blackwell chip fingerprint
AI-generated from the transcript