sourced story
12 September 2024Primary source · 2 sourcesWell documented

OpenAI's o1 introduces visible chain-of-thought reasoning

A model trained to think longer before answering jumps from 13 percent to 83 percent on a math test

On the timeline · around 12 September 2024 · Recognition and ReasoningTransformers and the Generative AI WaveRecognition and ReasoningOpenAI's o1 introduces visible chain-of-thought reasoning20232025

Quick facts

Institution
OpenAI
Method
Reinforcement learning on chain-of-thought reasoning
Benchmark
83% on IMO qualifying exam vs 13% for GPT-4o

What happened

OpenAI released o1-preview and o1-mini, describing them as a new series of models trained with reinforcement learning to work through a chain of thought before producing a final answer, rather than generating a response immediately as prior GPT models did. Through this training process, the model learned to break difficult problems into smaller steps, try alternative approaches when one path stalled, and recognize and correct its own errors, spending more computation at answer time on harder problems. On a qualifying exam for the International Mathematical Olympiad, OpenAI reported that its existing GPT-4o model correctly solved only 13 percent of problems, while the new reasoning model scored 83 percent. OpenAI deliberately kept the model's internal chain of thought hidden from users, citing both safety monitoring and competitive concerns, and reset its model-naming scheme to '1' to mark what it described as a new level of capability distinct from the GPT series.

Why it matters

o1 established test-time compute, letting a model think longer on harder problems, as a second scaling dimension alongside the training-time scaling that had driven progress from GPT-2 through GPT-4, shifting where AI labs began directing their heaviest computational investment.

How we know

OpenAI's own announcement and companion technical post document the training method and the reported math-olympiad scores directly; Georgetown's Center for Security and Emerging Technology, an independent policy research institute, corroborates the same reasoning-based training approach and its significance for future scaling strategy.

Sources

See something wrong? . Corrections with a source get fixed fastest.

Part of a timelineArtificial Intelligence30 events · From a wartime theory of neurons to machines that write, paint, and fold proteinsView all →