AI Foundationspredict · compress · act
act X

The Frontier

Newer models spend compute at answer time, not just training time. Sometimes the way to be right is to think longer.

41

Thinking Before Answering

before this →Watching the Loop

For years, the recipe was: train a huge model, then read off its first answer. The frontier shifted. Now a model can think before it answers — generating a long internal chain of reasoning, checking itself, and only then committing.

train-time computebigger model, more datapaid once, before deploymenttest-time computelonger reasoning, search, votingpaid per question, adaptivelya harder question can buy more thought
two ways to spend compute: train bigger, or think longer

Chain-of-thought

Chain-of-thought (CoT) is startlingly simple: ask the model to reason step by step before giving the answer, and accuracy rises on maths, logic, and multi-step problems. The tokens of reasoning give the model working memory — it can hold intermediate results that would otherwise have to fit in a single forward pass.

Reasoning models

A reasoning model is trained specifically to use a long internal thought process: reinforcement learning rewards correct final answers and lets the model discover productive reasoning patterns. At inference, it may generate thousands of hidden tokens, backtrack, try another approach, and verify. Test-time compute or inference scaling is the knob: how much thinking to buy. Some systems even sample many answers and take a majority vote, or search over reasoning paths.

The new scaling axis

capability per question, not per model

This reframes the economics. A smaller model that thinks for longer can beat a bigger model that answers immediately — and you only pay for the hard questions. Compute becomes a runtime dial instead of a fixed architecture.

the open questionDoes more thinking always help? Not always: models can reason themselves into convincing errors. Longer chains are harder to supervise, which is why faithfulness of reasoning is now an active safety problem.
DRAFTreason step by step, in the open
CHECKverify, backtrack, try again
ANSWERcommit — and show your work
introduces →chain-of-thoughtreasoning modeltest-time computeinference scaling
← previousWatching the Loopnext →Models of the World