Ask me what two plus two is and I will say four.
I do not derive it. I do not picture two stones, add another two stones, count the pile and arrive at four. The answer is just there. To use the machine-learning analogy loosely, it is accessible directly from my weights.
Now ask me for 237 multiplied by 48. I do not have 11,376 sitting in memory beside my phone number. I have to turn the question into questions whose answers I do have:
237 × 48
= 237 × (50 - 2)
= 11,850 - 474
= 11,376
I know how to multiply by 50. I know how to double 237. I know how to subtract. The useful thing is choosing a sequence of unimpressive steps that ends at an answer I could not retrieve directly.
This is how our brains work. Or, more precisely, this is how deliberate thinking feels from inside my brain. I do not answer a genuinely difficult question whole. I create smaller questions, answer those, carry the answers forward and eventually make the original question easy enough to finish.
Brains are not transformers and neither stores facts as neat rows in a database. The weights are an analogy, not a neuroscience claim. But the analogy explains why chain of thought mattered so much.
The standard explanation of an LLM is true. It predicts the next token, appends it and predicts another. Pretraining teaches the associations; instruction tuning and reinforcement learning make them more useful to humans.
Everyone watched some version of the "How ChatGPT works" explainer: an enormous statistical machine repeatedly guessing what word comes next. Once you understand that mechanism, calling it autocomplete with a datacentre feels devastating.
Except it mistakes the training objective for the capability ceiling.
In direct-answer mode, the model has to jump from the prompt towards a conclusion without leaving itself useful intermediate state. Two plus two becomes four. A problem requiring several dependent transformations is much more brittle.
Chain of thought changes the computation without changing the weights. The model writes an intermediate step into the context. That step becomes part of the next input. It writes another, and the next prediction is now conditioned on a problem the model has already partially transformed.
The text can be working memory, not merely an explanation displayed after the real work. Each generated step gives the next prediction a different, often easier, problem to solve.
Prediction is not the opposite of reasoning when predictions can create the state used by future predictions.
The phrase "stochastic parrot" came from a serious 2021 paper about the costs, data, bias and misleading fluency of very large language models. Most of those criticisms survived chain of thought perfectly well.
But the phrase escaped the paper and became a complete theory of capability: these systems only remix language, therefore they cannot reason.
The first half describes a mechanism. The second half does not follow from it.
A next-token predictor with no scratch space is mostly trapped in direct association. Give it a notebook and it can compose those associations over time. The operation did not become magical. The system became recursive.
The original chain-of-thought prompting paper, submitted in January 2022, showed that sufficiently large models performed much better on arithmetic, commonsense and symbolic reasoning tasks when examples included intermediate steps.
Then Zero-shot-CoT showed something stranger. The researchers added essentially one instruction: "Let's think step by step." On MultiArith, the same large InstructGPT model went from 17.7% accuracy to 78.7%. On GSM8K, it went from 10.4% to 40.7%.
Same weights. A different route through them.
The effect was not universal. Smaller models often did not benefit, and easy questions can get worse when a model overthinks them. Chain of thought mattered on the problems that needed a chain, once the model was capable enough to build one.
I always thought it was weird that you could ask GPT-3.5-turbo to decompose a problem into smaller chunks and suddenly get a better result. Same model. Same chat window. I had simply made it take the long way around.
I was prompting my own shitty chain of thought into the model before there was a polished reasoning mode for it. Sometimes I made it spend an entire pre-emptive turn just thinking: break down the task, list the constraints, identify what it did not know and decide the order of attack. Then I asked for the final answer in the next turn. The results were noticeably better.
The labs had already published the papers, so they clearly knew. What I find strange is how long it took them to make this the centre of the product. For years the obvious trick was sitting there: stop forcing the model to answer immediately and give it a turn to work first.
More reasoning tokens meant more inference-time compute. Models could explore several paths, check intermediate results and spend longer on hard problems. Tool calls took the same idea outside the chat window.
A tool-using model can call Python to calculate 237 multiplied by 48 exactly. It can search the web for information published five minutes ago, query a database for private state, inspect a repository, ask a compiler whether code is valid and run a test to see whether a fix worked.
Python gives it deterministic computation. Search gives it fresh information. Compilers and tests give it feedback that does not care how plausible the answer sounds.
None of those results needed to be embedded in the weights. The model's job changes from know the answer to know how to obtain the answer: choose a tool, construct the call, read the result and decide what to do next.
This is the bridge from chatbots to agents. An agent turns "fix this bug" into inspect the error, find the caller, form a hypothesis, edit the source, run the reproduction and inspect the result. Each action changes the context for the next one.
Chain of thought made the notebook part of the computation. Tools gave the notebook hands.
A written chain of thought is not guaranteed to be a faithful transcript of why the model produced an answer.
Models can guess first and manufacture a clean derivation afterwards. They can follow a biased cue without mentioning it. They can make one wrong intermediate step, build six elegant steps on top and land confidently in the wrong place. Research on unfaithful chain-of-thought explanations demonstrated exactly this.
So chain of thought is not a window into the model's soul. It is generated text, and generated text can lie.
That does not erase its computational role. A notebook can contain useful working, mistakes or a retrospective excuse. Usually it contains some mixture. The answer is verification, not blind trust.
Tools have the same limitation. A model can call the wrong tool or misread perfectly good output. The point is not that reasoning and tools make it infallible. The point is that it is no longer limited to remixing whatever survived pretraining.
Chain of thought did not prove that language models understand the world like humans. It did not solve hallucination, make their explanations trustworthy or invalidate the serious criticisms in the stochastic-parrots paper.
It did destroy the comforting idea that next-token prediction implied only shallow, one-step imitation.
The crucial capability was not storing every answer. It was learning enough small transformations that a missing answer could be assembled from available ones, with language acting as temporary memory and tools supplying what the weights could not.
Two plus two is simply there. A difficult formula is not. I walk from the difficult formula to smaller things that are there, carry each answer forward and keep going until the original question has nowhere left to hide.
A stochastic parrot with no notebook predicts. A stochastic parrot with a notebook can decompose, record, reconsider, verify and continue. It is still stochastic. It is still a parrot. But those words no longer tell you what the whole system can do.
Saying "it only predicts the next token" is now about as explanatory as saying "your brain only fires neurons."
The breakthrough was not escaping next-token prediction. It was realising that the next token could be a clock cycle.