🚀 OpenAI Releases AI Models Tailored for Science and Coding
posted 13 Sept 2024
OpenAI launched two new AI models equipped with reasoning capabilities: o1 and o1-mini, designed specifically for scientific tasks and programming. These models are distinguished by their extended response generation time, which allows for the application of various strategies and early detection of errors.
According to the developers, the main model achieved an 83% score at the qualifying stage of the International Mathematical Olympiad (IMO), which is 70% higher than GPT-4's results. However, both new models currently lack standard features such as internet access or the ability to upload files.
Model Comparison in Scientific Tests. Source: OpenAI's official website
The reasoning capabilities not only improve the quality of answers to complex scientific questions but also enhance the models' defenses against user attempts to bypass security measures (jailbreaks). In a specialized test, the o1 model scored 84 points, vastly outperforming GPT-4, which scored only 22.
The o1-mini model is more efficient in programming than in scientific research and is 80% cheaper than the main version. Paid ChatGPT users can now experiment with these models, which come with preset weekly limits: 30 messages for o1-preview and 50 for o1-mini.
More breaking news
