OpenAI launches its ‘o1-preview’ AI series
For advanced reasoning
OpenAI has launched its new reasoning model series, dubbed ‘OpenAI o1-preview’. This new series is set to redefine AI’s capabilities in tackling complex problems across various fields, including science, coding, and mathematics.
The OpenAI o1-preview models are designed to emphasise extended reasoning time, mimicking human thought processes. Unlike previous models, these new systems take more time to deliberate over problems, allowing for refined problem-solving strategies and improved accuracy.
The training for these models has focused on enhancing their ability to evaluate and correct their thinking, resulting in notable advancements in performance.
Early tests reveal that the o1-preview models are comparable to PhD students on challenging physics, chemistry, and biology benchmark tasks. For instance, while the previous GPT-4o model correctly solved only 13% of problems in an International Mathematics Olympiad (IMO) qualifier, the o1-preview achieved an impressive 83%.
The models also performed exceptionally well in coding, reaching the 89th percentile in Codeforces competitions.
AI safety
In tandem with the advancement in reasoning capabilities, OpenAI has introduced a new approach to AI safety. The o1-preview models have been trained to adhere to safety and alignment guidelines more effectively by reasoning through these rules in context.
This approach has significantly improved safety test performance. The o1-preview model scored 84 out of 100 in the most challenging jailbreaking tests, a substantial increase from the GPT-4o’s score of 22.
OpenAI’s commitment to safety is further demonstrated through enhanced internal governance, rigorous testing, and collaboration with federal safety institutes. The organisation has formalised agreements with AI Safety Institutes in the US and UK to ensure thorough research, evaluation, and testing of these models before their public release.
Alongside the release of the o1-preview on September 12, 2024, OpenAI also unveiled the ‘OpenAI o1-mini’ model. Designed for efficiency, the o1-mini is adept at generating and debugging complex code. At 80% cheaper than the o1-preview, this smaller model offers a cost-effective solution for developers needing high-level reasoning without broader world knowledge.
Since September 12, ChatGPT Plus and Team users can access the o1-preview and o1-mini models through ChatGPT. Users will find these models in the model selector, with weekly message limits of 30 for o1-preview and 50 for o1-mini. ChatGPT Enterprise and Edu users will gain access to both models next week. Developers qualifying for API usage tier 5 can start experimenting with the new models today, although current API features are limited.
The introduction of the OpenAI o1 series marks the beginning of a new era in AI reasoning capabilities. OpenAI plans to continue evolving and expanding its GPT series alongside the o1 models, with future updates expected to include enhanced features such as browsing, file and image uploading, and more.
Featured image: The OpenAI o1-preview models are designed to emphasise extended reasoning time, mimicking human thought processes. Credit: OpenAI
