o1

OpenAI's reasoning-model series, launched September 2024 and introduced in the primary page "Learning to reason with LLMs." o1 is trained to spend extra inference-time compute generating an internal chain of thought before answering, which OpenAI reported measurably improved performance on hard math, coding, and science problems relative to GPT-4o.

Matters to this vault as the reference point for the industry's first widely-discussed hidden-reasoning product: the model whose raw chain-of-thought OpenAI chose not to show users, citing competitive advantage among its reasons (claim-openai-hid-o1-raw-chain-of-thought-partly-for-competitive-advantage), showing a model-generated summary instead (claim-openai-o1-shows-model-generated-summary-of-chain-of-thought-not-raw-trace); the benchmark s1 was built to match on 1,000 distilled traces and 26 minutes of fine-tuning, cited as "o1-preview-level" performance (claim-s1-distilled-reasoning-from-1000-traces-in-26-minutes); and the case study for whether a hidden reasoning trace can function as a durable competitive moat (claim-a-chain-of-thought-trace-is-codified-so-it-cannot-form-a-tacit-moat).

References

• claim-openai-hid-o1-raw-chain-of-thought-partly-for-competitive-advantage
• claim-openai-o1-shows-model-generated-summary-of-chain-of-thought-not-raw-trace
• claim-s1-distilled-reasoning-from-1000-traces-in-26-minutes
• claim-a-chain-of-thought-trace-is-codified-so-it-cannot-form-a-tacit-moat