After 12 days of technical sharing live events, OpenAI released its next-generation inference model, o3, on the final day. This is an upgraded version following the earlier release of the o1 inference model. The o3 model series includes two versions: o3 and o3-mini, with o3-mini being a smaller, streamlined model fine-tuned for specific tasks. OpenAI stated that the o3 model can come close to achieving Artificial General Intelligence (AGI) under certain conditions, meaning it can perform any task that a human can accomplish.
In the ARC-AGI graphical logic reasoning benchmark test, the o3 model achieved record scores, with a score of 75.7% in low-computation scenarios, and it reached 87.5% in high-computation tests, surpassing the threshold of 85% that signifies human-level performance. In comparison, the o1 model scored only between 25% and 32%, making the performance of o3 nearly three times that of o1. On the globally renowned coding competition platform Codeforces, o3 achieved a rating of 2727, while o1 scored only 1891.
Fu Sheng, Chairman of Cheetah Mobile's Orion Star, stated that the release of OpenAI o3 heralds the arrival of an era where everyone can be a programmer. Users will no longer need to be proficient in Python or C to write programs; they simply need to express their requirements, and the large language model can assist in completing the programming tasks. Fu Sheng believes that the release of o3 signifies that the programming capabilities of large language models have surpassed 99.9% of programmers. In the world-class programming competition on Codeforces, o3 achieved a top score of 175, while o1 only outperformed about 90% of programmers, and the previous GPT-4o only surpassed 11% of programmers.
OpenAI plans to officially release the o3 model by the end of January next year. Fu Sheng pointed out that while programmers will not completely disappear, their work will increasingly focus on understanding user needs and constructing large logic, while the task of translating those needs into code will be significantly handled by AI. This release indicates that the application of AI in programming will become more widespread, potentially changing the way programmers work.
