Inference Answers Questions. It Doesn't Have Them.

Key Takeaways

• In both its machine-learning and logical senses, inference takes the question as given: the premises arrive from outside, already posed.
• The visible “thinking” a reasoning model emits is not reliably the cause of its answer; measured faithfulness runs well under half, and drops further on harder questions.
• Of the five ways inference differs from thinking, three are being actively engineered shut: initiation, boundary, and persistence.
• The two that remain are stakes and question-formation, and the second is the one that matters. Inference answers questions; it does not have them.
• You cannot introspect whether you out-thought the machine. Experienced developers in a randomised trial were 19% slower with AI while believing they were 20% faster.

You can now buy thinking by the token.

That is not a figure of speech. In Anthropic’s API there is a parameter called budget_tokens, and it “sets a target for how many tokens Claude can use for its internal reasoning process.” The minimum is 1,024. If you want to know what the thinking cost you, the response hands you a field, usage.output_tokens_details.thinking_tokens, which reports “how many of the billed output tokens were internal reasoning.”

Thinking has a line item now. It has a budget, a minimum, and a unit price.

Something happened to the word on its way into that pricing page, and I want to work out exactly what. Not to complain about the marketing, since the terminology is reasonable enough as engineering shorthand. But because the substitution runs in both directions. If what these systems do counts as thinking, then thinking is a thing that starts on request, runs to a stop token, and bills out at a few dollars per million. And if I have quietly started to believe that about thinking, I should notice, because I do it for a living.

So: is thinking the same as inference, or are they genuinely different things?

They are not the same, and the difference is precise: inference is the part of thinking that produces an answer once the question is fixed. It is a proper subset. Everything upstream of the question, including whether it was the right question to ask, sits outside the forward pass by definition.

What Is Inference, Exactly?

Inference has two precise definitions, and they agree on the thing that matters.

The machine-learning one is mechanical. Inference is the forward pass: a trained network with frozen weights takes an input, propagates it through, and emits a distribution over next tokens. It is deliberately distinguished from training, which is the phase where the system changes. Inference is the phase where it doesn’t. The episode begins when input arrives and ends at a stop token, and when it ends, the model is bit-for-bit what it was before.

The older definition, from logic, is about warrant rather than mechanism. To infer is to derive a conclusion from premises you already accept, whether deductively, inductively, or abductively. Aristotle’s syllogism is inference. So is a doctor reasoning from symptoms to a diagnosis.

These look like different animals. They share one property, and the whole essay turns on it:

The forward pass does not select its prompt. The syllogism does not choose its premises. Inference is a transformation applied to something handed to it. It is, quite precisely, the part that happens after someone has decided what to ask.

That is not a criticism. It is a boundary, and boundaries are useful. It just means that if thinking is only inference, then everything upstream of the question, all of it, is either not thinking or not happening.

Where Do Thinking and Inference Actually Diverge?

Five places, and they are worth separating, because they turn out to have very different futures.

Initiation. Inference requires a prompt. Nothing happens in a model between requests; there is no idling, no drift, no returning to yesterday’s problem. You, by contrast, did not choose to start worrying about the thing you are worrying about. Some of your most useful thinking arrives unbidden, in the shower, about a question nobody asked you.

Boundary. Inference terminates. There is a stop token, and after it the episode is closed. Thinking has no such marker. You can be finished with a problem and have it come back three days later with an objection attached.

Persistence. The weights are frozen. Whatever the model worked out during that forward pass leaves no trace in the model; the next request starts from the same parameters as the last. When you work something out, you are different afterward, in a way you cannot fully undo.

Stakes. Nothing is at risk for the system that infers. It does not get the thing wrong and have to live in the consequences. Whatever pressure is on the output comes from us, standing outside, checking.

Question-formation. The question arrives from outside. This is the one I keep circling back to, because it is the one that does not look like an engineering problem.

That is the taxonomy. The interesting question is not whether these gaps exist, since they plainly do, but whether they are permanent. Most of them aren’t.

Is the Visible “Thinking” Even the Thinking?

Before assessing the gaps, an inconvenient result: the reasoning a model shows you is not reliably the reasoning that produced its answer.

Miles Turpin and colleagues demonstrated this in 2023 in a paper whose title says it, Language Models Don’t Always Say What They Think. Their method was elegantly cheap. Reorder the multiple-choice options in a few-shot prompt so the correct answer is always (A), then ask a model a fresh question. The models picked up the pattern and let it move their answers, with accuracy dropping by as much as 36% across thirteen BIG-Bench Hard tasks. The written explanations never mentioned it. The models produced fluent, plausible justifications for conclusions they had reached for a reason they did not state.

Anthropic ran the harder version of this experiment on reasoning models in 2025 and published the results. They fed models a hint, confirmed the hint changed the answer, and then checked whether the chain of thought admitted it. Averaged across hint types, Claude 3.7 Sonnet mentioned the hint 25% of the time; DeepSeek R1, 39%. Two further findings sharpen it. Faithfulness “was on average lower when the question being asked was more difficult,” which means it degrades exactly where you would most want to trust it. And training did not fix it: outcome-based reinforcement learning improved faithfulness sharply at first, then flattened, stalling at 28% on one evaluation and 20% on another.

So the thing sold as thinking is, at best, a partial and unreliable transcript of the process it accompanies. That is also why grading the trace is a poor substitute for grading the work, which is the failure mode behind judging model output with another model.

Now the part that stops this from being a victory lap. Human introspection fails the same test. In 1977 Richard Nisbett and Timothy Wilson published Telling More Than We Can Know, reviewing evidence that “there may be little or no direct introspective access to higher order cognitive processes.” Asked why we chose what we chose, we do not read out the process. We construct a plausible account from implicit theories about what ought to have caused it, and we believe the account. That paper has been cited more than thirteen thousand times, largely because nobody has managed to make it go away.

Anthropic’s own concept-injection work lands in the same uncomfortable middle. Models can sometimes notice a concept injected into their activations and name it. Claude Opus 4.1 managed it in roughly 20% of trials under optimal conditions, which is real, functional introspective access. It is also, as the authors are careful to say, “highly unreliable; failures of introspection remain the norm,” and models “often provide additional details about their purported experiences whose accuracy we cannot verify, and which may be embellished or confabulated.”

Unfaithful self-reports, then, are not the distinguishing feature. Both parties confabulate. If we want a real difference, we have to look somewhere the symmetry breaks.

Which of These Gaps Are Actually Closing?

Three of the five, at varying speeds. That is the honest answer, and it is the reason I don’t think this question stays philosophical for much longer.

Initiation is closing for unremarkable reasons. Put a model on a scheduler, or inside an agent loop with a standing objective, and something asks it a question every few seconds without a human present. The gap was never deep; it was a property of the request/response API, not of the model.

Boundary is closing fast, and this is the one to watch. METR measures a metric called the 50% task-completion time horizon: the length of task, in human-expert time, that a model finishes half the time. In 2019 that was around four seconds. By early 2026 frontier agents were completing coding tasks that take humans over fourteen hours. The doubling period ran at about seven months over 2019–2025 and has since compressed to roughly four. “Ends at a stop token” is technically still true and decreasingly meaningful.

Persistence is closing, with an asterisk. Retrieval, scratchpads, and persistent memory files mean a system can carry conclusions across episodes. But this is a bolt-on, and worth being precise about: the model is not changed by what it concluded. It is handed a note it wrote earlier. Functionally that narrows the gap a great deal. Structurally it is a different thing from having been altered by understanding something. If you have ever failed to un-learn something you wish you could, you know that alteration is not optional for us.

Stakes are not closing, and notably, nobody is trying. Nothing about scaling, distillation, or reinforcement learning puts anything at risk for the model. It is not a capability problem, so capability work sails straight past it.

And question-formation is not closing, which requires more argument, because it sounds like exactly the sort of thing that should yield to scale.

Can You Out-Think an Inference?

Start with where you lose, because it is most of the board. On recall you lost years ago. On breadth you were never close. On speed it is not a contest. And on any well-posed problem, meaning one where the goal is stated, the constraints are known, and success is checkable, the trend line above should end the argument. Fourteen hours of expert work, at 50% reliability, doubling every four months. Betting your professional value on out-answering that is a bad trade.

Where the advantage survives is where the problem is ill-posed: ambiguous, underspecified, missing context, or quietly built on a false premise. And there is a measurable signal here, which is that current systems struggle not with answering such questions but with noticing them. Recent work on detecting ill-posed questions from a model’s internal states reports baseline classification accuracy of 67.4% on AmbigQA and 57.6% on the nine-way CLAMBER benchmark. That is, before adding specialised topological machinery, a frontier system’s ability to recognise that a question cannot be answered as asked hovers not far above chance on the harder split.

Sit with what that means. The failure is not at the answer. It is one step earlier, at the point where somebody should have said wait, this question is broken. Which is exactly the step inference is defined to skip. It received the premises. Its job starts after.

That is the seam. Not intelligence, not depth, not creativity in any romantic sense. Just this: someone has to notice the question is wrong, and inference structurally cannot be that someone.

And now the part that should worry you, because it is about your ability to tell whether you are doing that job.

In 2025 METR ran a randomised controlled trial with sixteen experienced open-source developers, on 246 real issues in mature repositories of 22k+ stars and a million-plus lines that they had worked in for years. Before starting, they expected AI to speed them up by 24%. Measured, they were 19% slower with AI available. Afterward, having lived through it, they still believed it had sped them up by 20%.

Not a small misestimate. A sign error, held confidently, by experts, about their own recent work.

I want to be fair to that study, because METR themselves have been. They have since revised the experiment design and say plainly that developers “are more sped up from AI tools now” in early 2026 “compared to our estimates from early 2025.” They flag real selection problems: a growing share of developers won’t accept doing half their work without AI, and 30–50% declined to submit tasks they didn’t want to attempt unaided. The 19% is not a durable fact about AI tools, and I am not going to use it as one.

The durable finding is the other one. The gap between what those developers experienced and what the clock recorded was not fixed by expertise, and was not visible from inside. Recall Nisbett and Wilson: we do not read out our own processes, we narrate them plausibly. Fluency feels like progress. An answer arriving quickly feels like a question well handled.

So the answer to “can you out-think an inference” has an uncomfortable second half. Sometimes, yes, in the specific place where the question is still open. And you will not be able to tell from the inside whether you did. The same unreliable-transcript problem that makes chain-of-thought a poor window into a model makes your sense of having thought a poor window into you.

So: Same Thing, or Different?

Different, but not in the flattering way the question usually invites.

Inference is a proper subset of thinking. Not a lesser one; a bounded one. It is the part that produces an answer once the question is fixed, and at that job it is now extraordinary and getting better on a four-month doubling. The parts it excludes are not excluded because machines are inadequate. They are excluded by the definition: a transformation applied to given premises cannot also be the act of choosing them.

Three of the five differences are architectural, and architecture yields. Initiation fell to scheduling. Boundary is falling to long-horizon agents. Persistence is falling, functionally, to memory. Anyone claiming those as permanent marks of the human mind is going to be embarrassed on a schedule.

Inference answers questions. It doesn’t have them.

What This Changes About How You Work

The risk was never that machines would start thinking. It is that we would quietly stop, not through laziness, but through a sensible-looking optimisation.

Because if you believe thinking is inference, the rational move is to get better at inference: faster retrieval, faster answers, shorter gaps between question and output. And you will lose, on a four-month doubling, at the one thing you optimised for. Meanwhile the capacity that was actually yours, sitting with a badly formed question long enough to notice it is badly formed, atrophies from disuse, and atrophies invisibly, because it never announces itself. Nothing tells you the question was wrong. That is the whole problem with wrong questions.

The mechanism here is not a fading appetite. Longitudinal work says curiosity itself does not decline with age, and a 2025 experiment found that AI answers carrying sources and an explanation cut follow-up questions from 71.4% to 12.7%. What atrophies is the room to keep asking, not the will to.

This is the same friction I keep coming back to. Not friction as virtue-signalling difficulty, but as the mechanism by which a question gets examined before it gets answered. Handing over the answer-production is fine, genuinely fine and often correct. Handing over the question-formation, without noticing you did, is the move that costs you something you cannot audit.

The systems infer. Beautifully, cheaply, at fourteen hours a go and climbing.

Whether anyone was thinking is a question about who decided what to ask.

More essays at Call to Think · About this project