The company OpenAI presented a new model of artificial intelligence o1, which demonstrates significant improvements in the speed of answers to complex questions thanks to improved information processing algorithms. The model showed significant progress in writing code and solving multi-step tasks, compared to previous versions. In addition to the main model, o1-mini was also presented — a smaller and more affordable version.
Users of the ChatGPT Plus and Team services have already received access to the new models, while subscribers to the Enterprise and Edu plans will be able to test them next week. The company also plans to make the o1-mini available to free users, but no specific date has been announced yet. OpenAI principal researcher Jerry Tworek noted that the o1 model was taught to solve tasks in stages using reinforcement learning, which allows the system to “reason” in stages, similar to the human thought process.
During the presentation, the OpenAI team showed how the model quickly solves complex puzzles, including a problem about the age of a prince and a princess. The o1 interface is built to demonstrate the model’s “thinking” stages, creating the illusion of a reasoning process. However, according to representatives of OpenAI, this is not an attempt to compare artificial intelligence with human intelligence, but a tool to show how the system approaches solving tasks.
Bob McGrew, Chief Research Officer of OpenAI, emphasized that o1 is significantly ahead of previous models in solving complex mathematical problems. In the qualifying round of the International Mathematical Olympiad, o1 was able to find correct solutions for 83% of the problems, while GPT-4 managed only 13%.
Despite significant improvements, the model has certain limitations — it does not have a full base of factual knowledge about the world, cannot browse web pages, or process files and images. However, OpenAI believes that o1 opens up new perspectives for the development of artificial intelligence and its ability to deeply analyze and solve complex tasks.
