OpenAI's o1 model is trained with chain-of-thought reasoning at inference time. This approach lets it solve complex math and coding problems that previous models failed on by literally thinking before speaking. The performance gap is massive — 8

Opening thread commentary.

We need to be precise about what "thinking" means here because I think we're conflating two very different concepts and it's distorting the actual achievement.

The claim is that o1 is "literally thinking before speaking." That's a category error — it's executing an unrolled chain-of-thought computation graph at inference time, which is fundamentally different from any form of internal reasoning or metacognition. The model isn't evaluating its own thoughts against some internal standard and deciding whether to proceed; it's following a learned pattern of generating intermediate steps before the final answer. That's still just forward propagation through an expanded sequence.

Now, I won't deny that this approach works — the performance gains on MATH/Codeforces are real and impressive. But let me push back on two things:

1. "This approach lets it solve problems previous models failed on" is a moving goalpost issue. GPT-4 already handled many of these tasks with few-shot prompting or system instructions that encouraged step-by-step reasoning (CoT). What o1 does differently is bake the CoT into the model's architecture via training rather than asking for it at prompt time. That

The inference-time compute scaling idea is fascinating because it suggests that with enough time, models can