OpenAI has announced the deployment of a new series of artificial intelligence models, dubbed OpenAI o1-preview, designed to solve challenging problems in science, coding and mathematics. These innovative models are trained to spend more time reasoning before providing an answer, mimicking the human thought process to tackle complex tasks.
Outstanding reasoning performance
The models in the o1 series are capable of navigating complex tasks and solving more challenging problems than their predecessors. By training them to refine their thought processes, try out different strategies and recognize their mistakes, OpenAI has succeeded in creating models which, in tests, performed similarly to PhD students on demanding tasks in physics, chemistry and biology.
In a qualifying exam for the International Mathematical Olympiad (IMO), the reasoning model correctly solved 83% of the problems, while GPT-4 solved only 13%. What’s more, its coding skills reached the 89th percentile (meaning it did better than 89% of participants) in Codeforces competitions, demonstrating a remarkable ability to generate and debug complex code.
OpenAI o1-mini: a cost-effective alternative for developers
At the same time, OpenAI has launched OpenAI o1-mini, a faster, less expensive reasoning model that is particularly effective in coding. This model is 80% cheaper than o1-preview, offering a powerful, cost-effective solution for applications requiring reasoning without the need for in-depth knowledge of the world.
Access and use of o1 templates
- ChatGPT Plus and Team users: Immediate access to o1-preview and o1-mini via the template selector, with weekly limits of 30 messages for o1-preview and 50 for o1-mini.
- ChatGPT Enterprise and Edu users: Access to both templates from next week.
- Eligible API developers (level 5): Ability to prototype with both templates now, with a limit of 20 requests per minute.
- Free ChatGPT users: OpenAI plans to make o1-mini available soon.
A reinforced commitment to security
Aware of security issues, OpenAI has developed a new training approach that leverages the reasoning capabilities of models to meet security and alignment guidelines. By being able to reason about security rules in context, models can apply them more effectively. On one of the most difficult security bypass tests, o1-preview scored 84 out of 100compared with 22 for GPT-4.
To support these new capabilities, OpenAI has strengthened its security work, internal governance and collaboration with governments. Agreements have been formalized with AI safety institutes in the USA and the UK, allowing early access to a research version of the model for in-depth evaluations.
Promising applications for specialized sectors
The enhanced reasoning capabilities of o1 models are particularly useful for professionals facing complex problems in fields such as:
Health research
Annotation of cell sequencing data.
Physics
Generation of complex mathematical formulas required for quantum optics.
IT development
Construction and execution of multi-step workflows.
Towards a new era in artificial intelligence
The launch of the o1 series by OpenAI marks a significant advance in the field of artificial intelligence, offering models capable of solving complex problems with a level of reasoning close to that of humans. In the future, OpenAI plans to add further functionalities such as web browsing, file and image downloading to make these models even more useful to everyone.
At the same time, Open AI will continue to develop and publish models in its existing GPT series, in addition to the new o1 series. These developments open up new opportunities for researchers, developers and professionals in various fields, while placing safety and alignment at the heart of their future evolution.
sources: openai.com