That’s OpenAI admitting that o1’s “chain of thought” is faked after the fact. The “chain of thought” does not show any internal processes of the LLM — o1 just returns something that looks a bit like a logical chain of reasoning.

I think it's fake "reasoning" but I don't know if (all of) OpenAI thinks that. They probably think hiding this data prevents cot training data from being extracted. I just don't know how deep the stupid runs.