- What happened
- The Information reports DeepSeek and Moonshot fed answers from OpenAI's GPT 6 Astra into their own training, benefiting Moonshot's latest Kimi model, released last week, and a DeepSeek update. OpenAI had already tied a July copying campaign to people linked to Moonshot. OpenAI, DeepSeek, and Moonshot did not comment.
- Austin's take
- Austin explains distillation as querying a model at huge scale through paid accounts to train a cheaper copy, and says it isn't new. He calls Anthropic's earlier push for government action on distillation hypocritical, since frontier labs trained on the open internet without permission, but praises Anthropic for later saying it can block nearly all distillation itself. He expects OpenAI to get there too.
- Why it matters
- Distillation lets rivals catch up for a fraction of the cost, which affects US labs' lead and fuels the debate over AI intellectual property and terms of service.
- What to do
- Treat this as a developing report. If you build on AI APIs, know that distillation violates most providers' terms and accounts get cut off for it.
- Not confirmed yet
- The report rests on unnamed sources, and none of the companies commented.
- Sources
From the Thursday, October 29 show
Report: China's top AI labs trained on OpenAI's outputs
AI-generated from the transcript