Test-time compute — the new AI scaling law

What if the same model got smarter — just by thinking longer? That's exactly what o1 did.

For years, smarter AI meant one knob — bigger model. More parameters, more data, more GPUs. Then OpenAI shipped o1: same brain, but allowed to think longer before answering. And hard-math scores doubled — with no new training.

What's actually happening? A normal LLM (large language model) answers in one shot — left to right, no pause. A reasoning model writes a chain of thought first — a scratchpad where the model talks to itself. It tries one path, hits a dead end, tries another. The longer it scratches, the more paths it explores. That's test-time compute — thinking at answer time, not at training time. In plain English: you pay for thought, not a bigger brain.

On AIME — a hard math benchmark — the gap is wild. GPT-4o, answering fast, scored about 13%. o1, allowed to think, hit around 83%. Same family of models — just more compute at answer time. The catch: every thinking token costs money. Longer answers, bigger bill.

The new rule: make the model think, not just bigger.

Music: Markvard - Time [NCS Release] (NoCopyrightSounds)
https://ncs.io

#ai #reasoning #o1 #scaling #llm #shorts #programming

Видео Test-time compute — the new AI scaling law канала ProCode