YOUR MODEL LEFT FINGERPRINTS

The real news

The story, straight

Apparently, chatbots need gloves now.

An AI model learns by adjusting numbers called weights. Before learning begins, software gives those numbers starting values.

The chip's processing cores divide that work, leaving an arrangement that depends on the hardware. Some starting values survive training, like pencil marks beneath an edited drawing.

Unemployment Labs reports Blackwell-compatible patterns in 6 Chinese models from DeepSeek, Moonshot AI, MiniMax and Ant Group. The Register covered the findings.

The patterns point to hardware used to create the starting numbers. That supports, but does not prove, Blackwell also performed the training.

Lil'meme is dusting a model file for fingerprints. Apparently deleting your browser history was not the advanced privacy strategy we thought.

Nvidia Blackwell chips require a US export licence for China; none has been reported for this use. But these files cannot establish where the chips were, who owned them or whether any law was broken.

Renting Blackwell systems outside China remains a possible explanation.

The useful takeaway: published model files can reveal clues about their beginnings. They are not shipping receipts, and starting hardware is not necessarily training hardware.

A specially configured AMD system remains an alternative for some matches.

Read the report: https://unemploymentlabs.com/checkpoint-forensics

Repost this before the chatbot demands fingerprint-resistant gloves.

Thanks for supporting Lil'meme. Your support is appreciated.