Atlas record: https://www.aitimeline.dev/en/events/evt-0308/
Announcement · September 12, 2024
o1 and compute at answering time
OpenAI showed a model trained to think before answering: it generates a long hidden chain of reasoning, and the longer it thinks the more accurate the answer.
Why it matters
A second axis of growth appeared: quality began to depend on time spent thinking, not only on model size and training volume.
The model was trained by reinforcement on problems with a checkable answer: mathematics, code, the sciences. Reward went to the correct result, and the model developed habits that look human, checking itself, backtracking, trying another approach. On the qualifying exam for the American mathematics olympiad the score jumped from a few per cent to over eighty. OpenAI hid the chain of reasoning and showed only a summary, which drew criticism. The practical consequence: the cost of an answer stopped being fixed, because accuracy is now paid for in seconds.
Event record
- Event date
- September 12, 2024
- Timeline date
- Event date
- Verification
- Sources gathered automatically · September 18, 2026
- Lines
- ID
- evt-0308
Announcement of 12 September 2024.
Related events
Earlier
- Related Compute at answering time against a bigger model 2024links to this record
The paper came out five weeks before o1 was shown, and measured the same dependency independently on open models.
- Related OpenAI's Preparedness Framework 2023links to this record
The o1 system card reports the evaluation of o1-preview and o1-mini under this framework: medium risk overall.
- Related InstructGPT and learning from human ratings 2022
Reinforcement learning applied not to rater preference but to the correctness of a solution.
- Related Chain-of-thought prompting 2022links to this record
o1 uses a chain of thought when solving a problem and learns to hone it through reinforcement learning; OpenAI does not cite this paper, so the record claims kinship, not descent.
- Related Scaling laws 2020
Quality grows with compute spent at answering time, not only at training time.
Later