That is such a canard, IMO. FWIW, Anthropic and OpenAI encrypt "thinking" token outputs in their models, while Chinese labs don't. If anything, it's more likely

That is such a canard, IMO. FWIW, Anthropic and OpenAI encrypt "thinking" token outputs in their models, while Chinese labs don't. If anything, it's more likely that everyone is using open-weight models in their synthetic training data generation pipelines. It's way easier to distill from logits than it is to distill from hard tokens. https://x.com/EricSimons/status/2099252922098061714 https://x.com/EricSimons/status/2099252922098061714