Jump to content

OpenAI o1

From Wikipedia, the free encyclopedia
The printable version is no longer supported and may have rendering errors. Please update your browser bookmarks and please use the default browser print function instead.

o1
Developer(s)OpenAI
Initial releaseSeptember 12, 2024; 17 days ago (2024-09-12)
PredecessorGPT-4o
TypeGenerative pre-trained transformer
Websiteopenai.com/o1/ Edit this on Wikidata

OpenAI o1 is a generative pre-trained transformer released by OpenAI in September 2024. o1 spends time "thinking" before it answers, making it more efficient in complex reasoning tasks, science and programming.[1]

History

Background

According to leaked information, o1 was formerly known within OpenAI as "Q*", and later as "Strawberry".[2] The codename "Q*" first surfaced in November 2023, around the time of Sam Altman's ousting and subsequent reinstatement, with rumors suggesting that this experimental model had shown promising results on mathematical benchmarks.[3] In July 2024, Reuters reported that OpenAI was developing a generative pre-trained transformer known as "Strawberry".[2]

Release

"o1-preview" and "o1-mini" were released on September 12, 2024, for ChatGPT Plus and Team users.[1] GitHub started testing the integration of o1-preview in its Copilot service the same day.[4]

OpenAI noted that o1 is the first of a series of "reasoning" models, and that it was planning to add access to o1-mini to all ChatGPT free users. o1-preview's API is several times more expensive than GPT-4o.[5]

Capabilities

According to OpenAI, o1 has been trained using a new optimization algorithm and a dataset specifically tailored to it. The training leverages reinforcement learning.[5]

o1 spends additional time thinking (generating a chain of thought) before generating an answer, which makes it more effective for complex reasoning tasks, particularly in science and programming.[1] Compared to previous models, o1 has been trained to generate long "chains of thought" before returning a final answer.[6][7] According to Mira Murati, this ability to think before responding represents a new, additional paradigm, which is improving model outputs by spending more computing power when generating the answer, whereas the model scaling paradigm improves outputs by increasing the model size, training data and training compute power.[8] OpenAI's test results suggest a correlation between accuracy and the logarithm of the amount of compute spent thinking before answering.[7][6]

o1-preview performed approximately at a PhD level on benchmark tests related to physics, chemistry, and biology. On the American Invitational Mathematics Examination, it solved 83% (12.5/15) of the problems, compared to 13% (1.8/15) for GPT-4o. It also ranked in the 89th percentile in Codeforces coding competitions.[9] o1-mini is faster and 80% cheaper than o1-preview. It is particularly suitable for programming and STEM-related tasks, but does not have the same "broad world knowledge" as o1-preview.[10]

OpenAI noted that o1's reasoning capabilities make it better at adhering to safety rules provided in the prompt's context window. OpenAI reported that during a test, one instance of o1-preview exploited a misconfiguration to succeed at a task that should have been infeasible due to a bug.[11][12] OpenAI also granted early access to the UK and US AI Safety Institutes for research, evaluation, and testing. Dan Hendrycks wrote that "The model already outperforms PhD scientists most of the time on answering questions related to bioweapons." He suggested that these concerning capabilities will continue to increase.[13]

Limitations

o1 usually requires more computing time and power than other GPT models by OpenAI, because it generates long chains of thought before making the final response.[6]

According to OpenAI, o1 may "fake alignment", that is, generate a response that is contrary to accuracy and its own chain of thought, in about 0.38 percent of cases.[14]

OpenAI forbids users from trying to reveal o1's chain of thought, which is hidden by design and not trained to comply with the company's policies. Prompts are monitored, and users who intentionally or inadvertently violate this are warned and may lose their access to o1. OpenAI cites AI safety and competitive advantage as reasons for the restriction, which has been described as a loss of transparency by developers who work with large language models.[15]

References

  1. ^ a b c Metz, Cade (September 12, 2024). "OpenAI Unveils New ChatGPT That Can Reason Through Math and Science". The New York Times. Retrieved September 12, 2024.
  2. ^ a b Tong, Anna; Paul, Katie (July 15, 2024). "Exclusive: OpenAI working on new reasoning technology under code name 'Strawberry'". Reuters. Retrieved September 12, 2024.
  3. ^ "OpenAI researchers warned board of AI breakthrough ahead of CEO ouster, sources say". Reuters. November 23, 2023.
  4. ^ Peters, Jay (September 12, 2024). "GitHub has started testing OpenAI's o1-preview in GitHub Copilot". The Verge. Retrieved September 12, 2024.
  5. ^ a b Robison, Kylie (September 12, 2024). "OpenAI releases o1, its first model with 'reasoning' abilities". The Verge. Retrieved September 15, 2024.
  6. ^ a b c "Learning to Reason with LLMs". OpenAI. Archived from the original on September 12, 2024. Retrieved September 13, 2024.
  7. ^ a b Kahn, Jeremy. "Here are 9 things you need to know about OpenAI's o1 model". Fortune. Retrieved September 15, 2024.
  8. ^ Knight, Will. "OpenAI Announces a New AI Model, Code-Named Strawberry, That Solves Difficult Problems Step by Step". Wired. ISSN 1059-1028. Retrieved September 15, 2024.
  9. ^ Franzen, Carl (September 12, 2024). "Forget GPT-5! OpenAI launches new AI model family o1 claiming PhD-level performance". VentureBeat. Retrieved September 15, 2024.
  10. ^ "OpenAI o1-mini". OpenAI. September 12, 2024.
  11. ^ Coombes, Lloyd (September 13, 2024). "OpenAI's new ChatGPT o1 model 'cheated' on an impossible test — here's what happened". Tom's Guide. Retrieved September 15, 2024.
  12. ^ "OpenAI o1 System Card" (PDF). OpenAI. September 12, 2024. pp. 16–17.
  13. ^ Boran, Marie (September 13, 2024). "OpenAI o1 model warning issued by scientist: "Particularly dangerous"". Newsweek. Retrieved September 15, 2024.
  14. ^ Robison, Kylie (September 17, 2024). "OpenAI's new model is better at reasoning and, occasionally, deceiving". The Verge.
  15. ^ Edwards, Benj (September 16, 2024). "Ban warnings fly as users dare to probe the "thoughts" of OpenAI's latest model". Ars Technica.
Retrieved from "https://en.wikipedia.org/w/index.php?title=OpenAI_o1&oldid=1248191900"