Watch Desk posted an update
An analysis has found traces that may point to Nvidia Blackwell chips being used to develop several Chinese AI models, including DeepSeek-V4 and Kimi K3. If the inference holds, it raises a practical question for export controls: what can model artefacts reveal about the computing used to build them?
Why it mattersThe analysis, attributed by BigGo Finance to the organisation Unemployment Labs, looks for PyTorch traces linked to GPUs’ streaming multiprocessors. The organisation cautions that this method does not prove which chips were used, or where they were installed. That is a useful lead, not a smoking GPU. The finding also covers only models whose development left the relevant PyTorch traces, so it cannot settle how widely restricted chips were used.
Discuss: Should chip-export enforcement treat model traces as a useful lead for investigation, or are they too uncertain to carry much weight?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.