For years, the way to improve a model was to invest more compute in training it (scaling laws). Test-time compute is the other tap: spending more computation each time the model answers. Since each generated token is a full pass through the network, writing more before answering amounts to thinking more. Chain-of-thought prompting already worked this way: it lets the model distribute more computation over the reasoning process (Alammar & Grootendorst, 2024, p. 185). Another variant is self-consistency: generate several answers by sampling and keep the one that comes up most often (Alammar & Grootendorst, 2024, p. 188).
The 2024 turn was to train models to exploit this route. Introducing o1, OpenAI reported that its performance consistently improves with more reinforcement learning (train-time compute) and with more time spent thinking (test-time compute) (OpenAI, 2024). DeepSeek showed something similar in the open: R1-Zero, trained with reinforcement learning alone on problems with verifiable answers, learned on its own to write ever longer reasoning, from hundreds to thousands of tokens, and even had an "aha moment" in which it started using the word wait to revisit what it had been doing (Guo et al., 2025). See Reasoning model (o1, R1).
The idea has practical consequences: answering costs more, and the cost shifts from the lab that trains to the user who asks. And it has a tempting philosophical reading: Alammar & Grootendorst (2024) associate direct generation with fast, intuitive thinking ("System 1") and explicit reasoning with slow, deliberate thinking ("System 2") (2024, p. 185). But the analogy has a limit. That the intermediate text improves the answer does not prove it is the reasoning that produces the answer; and a model that "thinks more" does not know more, it only explores more. Is deliberation a matter of the amount of computation?