16 Sep OpenAI o1-Preview
OpenAI has introduced a new series of reasoning models, branded as the o1-preview, which is designed to tackle harder problems with improved reasoning capabilities. This new model series, made available starting on September 12, 2024, marks a significant advancement in AI’s ability to solve complex tasks in fields such as science, coding, and mathematics. Unlike previous models, the o1 series has been developed to take more time to think before responding, allowing it to reason more effectively through difficult problems.
The initial release of this model series is now accessible via ChatGPT and through OpenAI’s API as a preview. It has been stated that regular updates and improvements will be expected. Additionally, evaluations for the next iteration of this model, which is currently under development, are being included alongside the release.
These models have been trained with a focus on spending more time refining their thought process, trying various strategies, and recognizing potential mistakes. The next model update, as noted in tests, has demonstrated performance comparable to PhD students in challenging benchmark tasks in physics, chemistry, and biology. Further tests in math and coding show that in a qualifying exam for the International Mathematics Olympiad, GPT-4o only solved 13% of the problems, whereas the reasoning model scored 83%. Coding abilities were also evaluated, placing the model in the 89th percentile in Codeforces competitions.
As an early version, it has been mentioned that certain features common in ChatGPT, such as browsing the web for information or uploading files and images, are not yet available. However, for complex reasoning tasks, this model represents a leap forward, offering capabilities that are expected to surpass existing models like GPT-4o.
In terms of safety, the o1-preview model includes a new approach that leverages its reasoning abilities to better align with safety guidelines. It has been noted that the model’s capacity to adhere to safety rules, especially in resisting attempts at jailbreaking, has improved significantly. In tests, while GPT-4o scored 22 out of 100 in a jailbreaking test, the o1-preview model achieved a score of 84.
OpenAI has also emphasized its strengthened internal safety governance and increased collaboration with federal agencies. Partnerships with AI Safety Institutes in the U.S. and U.K. have been formalized, and early access to research versions of the model has been granted to these institutes. This collaboration is part of a broader effort to establish protocols for testing and evaluating AI safety before and after the public release of future models.
These new reasoning capabilities are expected to be particularly useful for researchers and developers working in complex fields such as healthcare, physics, and coding. For instance, the o1 model could assist healthcare researchers in annotating cell sequencing data, physicists in generating complex mathematical formulas for quantum optics, and developers in creating multi-step workflows across various industries.
No Comments