OpenAI’s new AI model can solve multi-stage tasks

The OpenAI company presented a new o1 model, which significantly surpasses previous versions in terms of speed and efficiency in solving complex tasks. One of the main features of this model is an improved “thinking” algorithm that allows it to solve problems that require multi-step analysis faster than a human can do.

The OpenAI o1 model is particularly good at programming and solving problems that require a sequential approach. Together with it, a lighter version was released – o1-mini, which is aimed at a wider range of users due to its affordable price. Currently, o1-preview and o1-mini are available to ChatGPT Plus and Team subscribers, with Enterprise and Edu users getting access next week. In the future, it is planned that free ChatGPT users will also be able to try o1-mini.

Special attention was paid to a new approach to model training. OpenAI principal researcher Jerry Tworek explained that o1 was trained using reinforcement learning, which allowed the system to better process complex queries step by step, simulating the human thought process. This new approach was evident in o1’s ability to solve complex puzzles. For example, at the presentation, the model solved the problem about the age of the prince and princess in 30 seconds, demonstrating her unique analytical ability.

The o1 interface has also been refined to improve the user experience. The model now shows the stages of her “thinking” using phrases like “I’m curious” or “Let’s see.” This creates the illusion of dialogue similar to human thinking, although, according to the developers, this is only a visual representation of the model’s working process, and not an attempt to imitate the human mind.

OpenAI Chief Scientist Bob McGrew noted that o1 significantly outperforms GPT-4o in solving complex mathematical problems. For example, in the qualifying exam of the International Mathematical Olympiad, the model correctly solved 83% of the problems, while the previous version only solved 13%.

Although the o1 model has significant improvements in reasoning and algorithms, it is still inferior to GPT-4o in actual knowledge of the world and cannot handle web pages or process files and images. However, OpenAI sees o1 as an important step towards creating more advanced and intelligent systems capable of solving complex tasks at a human-like level.