Meta employees testing prediction-market app Takes, report says
The internal app uses credibility credits instead of real money, and there is no clear timetable for a public launch.
Research points to restricted Nvidia chips in Chinese AI model training
An examination of publicly released model weights found signs of Blackwell use, though the findings do not establish export-control violations.
Chinese AI developers appear to have used Nvidia’s export-restricted Blackwell processors to train prominent open-weight models, according to research by Will DePue, a former OpenAI researcher who heads the nonprofit Unemployment Labs.
DePue examined model weights released by DeepSeek, Moonshot AI, MiniMax and Ant Group. He found evidence of Blackwell training in models including Moonshot’s Kimi’s K3 releases and DeepSeek’s V4 models. His analysis relied on traces of training-chip configurations retained in the weights.
The findings do not establish that Nvidia hardware handled all of the models’ training or that export laws were broken. The processors could have operated in offshore data centers, which Chinese companies have used to obtain computing capacity. U.S. restrictions prohibit Blackwell exports to China or Chinese-owned companies.
Separately, White House technology adviser Michael Kratsios said earlier in the year that Moonshot had used restricted Nvidia chips, citing U.S. intelligence. Research nonprofit Epoch AI estimates that about two million advanced chips reached China during 2024 and 2025 through arrangements concealing their destination.