Artificial intelligence · 2024

Reasoning models (o1)

Models trained with RL to think at length before answering.

OpenAI's o1, previewed September 12, 2024, used large-scale reinforcement learning to produce long private chains of thought. It reached 83% on AIME 2024 math problems (with consensus sampling) against GPT-4o's 13%, and opened a second axis of scaling beyond pretraining.

See it in the living treeWikipedia