LLMs like OpenAI o1/o3 and DeepSeek that appear to 'think' via a thinking/reasoning box in the UI. Hasiński's debunking: there is no thinking — these models simply have an extra role (e.g. 'reason') in the chat template where they generate tokens that aren't shown to the user. What some UIs show is a summary produced by a separate model, not the raw reasoning. DeepSeek exposes the actual reasoning tokens. Mechanically identical to normal generation, just with a hidden role.