OpenAI has introduced a new model, “ChatGPT o1”. This model has the ability to make judgments and solve logical problems by imitating the human thinking process. It breaks down complex tasks into simpler steps, analyzes different approaches, and corrects errors.
Oku.Az-according to information provided by “3DNews”.
It is reported that the “o1” model has shown results that can be compared with the results of experts in tests and competitions. In the field of programming, “o1” ranked 49th in the 2024 International Olympiad in Informatics (IOI), surpassing 89% of participants on the Codeforces platform. In mathematics, “o1” ranked among the top 500 students in the United States in the selection stage of the American Mathematical Olympiad (AIME).
In the natural sciences, the model outperformed PhDs and M.D.s on the GPQA Diamond test, which assesses knowledge in chemistry, physics and biology.
The developers note that this does not mean that the model is superior to any candidate in science, but it demonstrates its ability to solve problems at the level of highly qualified specialists.
The “o1” model also outperformed previous models in intelligence and problem-solving tests, including MMMU and MMLU. According to “OpenAI”, “o1” significantly outperforms “GPT-4o” in tasks related to logical thinking. In AIME tests, “o1” solved an average of 83% of the problems and “GPT-4o” solved 13%.
But “o1” runs slower than “GPT-4o” and its service is more expensive. It is behind “GPT-4o” in terms of encyclopedic knowledge, it cannot process web pages, files and images.
The success of the model is explained by a new training algorithm: “chain of thought”. This allows the model to recognize and correct errors, break down complex steps into simpler ones, and improve its results by trying different approaches.
OpenAI has released a preliminary version of the o1-preview model. It is available for use in ChatGPT and for developers via API. The cost of using o1-preview is $15 for 1 million input tokens and $60 for 1 million output tokens. GPT-4o costs $5 and $15, respectively.