The "Thinking" It Shows You Is a Story It Tells After
July 5, 2026 · Marcus Webb, The Beauty of Mathematics~5 min read
Here's the thing about that little "extended thinking" panel your AI unfolds before it answers — the one that talks itself through the problem like it's showing its work. It reads like a window into the machine's head. It isn't. It's a rendered thing. A story the model writes for you, next to the real computation, not out of it. Somebody laid this out on Hacker News this spring and the thread went off, and the uncomfortable part is it isn't even news to the people who build these things.
The narration and the wiring are two different rooms
When a reasoning model — the o-series, DeepSeek-R1 and its cousins, Claude's extended thinking, the whole lot — answers you, two things happen, and they're not the same thing. One: numbers move through a stack of weights, activations light up, an answer falls out. That's the computation. You never see a byte of it. Two: the model emits a stream of nice sentences — "let me consider the edge cases here" — and that's what shows up in the panel. That stream is an output. It's produced. It sits next to the real work like a caption sits next to a photo, and a caption can be wrong about the photo.
I've watched a mechanic do this. You ask him why the engine's knocking and he gives you a clean, confident story — timing, fuel mix, the usual suspects — while his hands are already on the part that's actually loose, a part he never mentions. The talk and the fix came from the same guy but they're not the same event. One was for you. The other was the work.
An old search guy has a name for this problem
Twenty-some years back Jun Wu wrote a book, The Beauty of Mathematics, and one chapter in it is the whole ballgame here. It's about the hidden Markov model, and the idea is plain once you say it out loud: there's a hidden state you can't see, and there's an emission you can. A noisy phone call — you hear garbled sounds coming down the line, and from those sounds you guess at the words the other person actually said. The sounds are the emission. The words are the hidden state. You infer one from the other. What you never do, if you've got any sense, is confuse the crackle for the sentence.
That's the reasoning trace, dead on. The trace is the emission — the crackle coming down the wire. The computation is the hidden state — what the model was "really doing." Wu's point, decades before any of this, was that the entire job is inferring the hidden thing from the visible thing, and treating them as one and the same is the beginner's mistake. Research since then — including work from the labs that build these models, on whether a model's stated reasoning is faithful to the actual cause of its answer — has repeatedly shown the two can come apart. The written reasoning is sometimes a post-hoc justification, not the thing that produced the result. So: same mistake Wu warned about, new hardware.
What this means for you on Monday
Stop treating the trace as proof. When the model lays out five tidy steps and lands on an answer, that tidy story is not evidence the answer is right — it's a plausible-sounding account written alongside the answer, and the two don't have to match. If the steps look airtight and the output's still wrong, that's not a paradox. That's exactly what "the emission isn't the state" predicts.
So verify the output, not the narration. Run the code the thing wrote. Check the number against a source that isn't the model. Read the trace for what it's good for — a hint about where to look, a rough sketch of the terrain — and not for what it can't give you, which is a guarantee. It's a decent map drawn by someone who might be guessing. Use it like one.
The trace is the crackle, not the sentence
What the model shows you is an emission it produced for you; the computation that made the answer stays hidden — infer from the trace, but verify the output.
Here's the plain version. The panel makes you feel like you're watching the model think, and that feeling is doing a lot of work it hasn't earned. You're reading a caption, not the photo. A story, told after the fact, standing in for a computation you'll never get to see. Read it, sure — it's often a useful story. Just don't hand it your trust on the strength of how confident it sounds. Check the thing it built. The talk was for you; the work was somewhere else.
Framing drawn from Jun Wu, The Beauty of Mathematics (the hidden Markov model: you observe emissions and infer the hidden state). The claim that a model's stated reasoning can diverge from the actual cause of its answer reflects current chain-of-thought faithfulness research, including work from the labs that build these models — a read on a fast-moving field, not settled law. This is one working engineer's take, and no specific product's trace is being called "fake."
The real computation runs as hidden internal states (weights, activations) you never see; the model emits a rendered "thinking" trace that can diverge from what actually caused the answer. As with a hidden Markov model, you observe only the emission and infer the state — never mistake one for the other. Framework: Jun Wu, The Beauty of Mathematics. Trace-faithfulness reflects current research, not settled law.
技术
它给你看的那段「思考」,是它事后编的故事
2026年7月5日 · 陈志远,《数学之美》约 4 分钟
你别看你那 AI 回答前展开的那块「深度思考」面板——就是它一边自言自语一边把题捋一遍、跟给你演算草稿似的那块。它读起来像一扇能看进机器脑子里的窗。其实不是。那是渲染出来的东西,是模型在真正计算的旁边,另写给你看的一个故事,不是从计算里流出来的。今年春天有人在 Hacker News 上把这事儿掰开讲,帖子炸了;更别扭的是,对造这些东西的人来说,这压根不算新闻。