What Is AI Reasoning Model's Chain of Thought Really Thinking?
Intro: I Thought I Understood It — Until I Actually Watched It Late last year, SFD Lab started using reasoning models at scale for business logic. The first …
Intro: I Thought I Understood It — Until I Actually Watched It
Late last year, SFD Lab started using reasoning models at scale for business logic. The first time I opened DeepSeek R1's "thinking" output and saw a long stream of self-talk, I was genuinely confused — what is this thing doing?
It wasn't answering the question. It was thinking through the question. And the way it thought was completely different from what I expected.
What Is Chain of Thought (CoT)?
Chain of Thought, or CoT, literally means "a chain of reasoning steps." It was first proposed by Google in a 2022 paper — they discovered that if you ask a language model to write out its reasoning steps before outputting an answer, accuracy improves dramatically.
This looks unremarkable on the surface, but the underlying logic is counterintuitive: the model doesn't "know the answer and then write steps" — it "figures out the answer by writing the steps."
From Prompt-Based CoT to Native Reasoning
Early CoT was crude — add "please think step by step" to the prompt. The real breakthrough came when OpenAI released o1 with reasoning baked into training.
DeepSeek R1 took a different approach: it exposes the thinking process directly inside <think></think> tags.
Several Counterintuitive Conclusions from Our Lab
Conclusion 1: More Reasoning Tokens ≠ Better Answers
We found that when reasoning was between 800–2000 tokens, answer accuracy was highest. After 3000 tokens, the model started "overthinking" — repeatedly second-guessing intermediate conclusions and ultimately giving contradictory answers. We call this "overthinking syndrome."
Conclusion 2: The Chain of Thought Lies
We found that R1's thinking content and the final answer sometimes contradict each other. The thinking output doesn't fully reflect the model's true reasoning process — it's more like a post-hoc explanation. Don't be too reverent about thinking output.
Conclusion 3: Reasoning Models' Edge Is Greater on Ambiguous Tasks
- Clear tasks (code completion, format conversion): minimal gap, reasoning models even slower
- Ambiguous tasks (business decision analysis, multi-factor tradeoffs): reasoning models vastly better
- Creative tasks (writing, brainstorming): reasoning models can actually be worse
The Cost Problem
o1-pro costs over 60x more than standard GPT-4o. Our practice at SFD Lab is tiered dispatch: routine tasks use standard models, complex judgments and final verification use reasoning models. This keeps reasoning model usage below 15% of total calls.
Summary
Chain of Thought isn't magic — it's trading more compute for better reasoning. Understanding its limits is what lets you capitalize on the reasoning model wave.