OpenAI, creator of ChatGPT, announced o3 and o3 mini inference AI models to tackle complex challenges.
According to CEO Sam Altman, OpenAI plans to release o3 mini by the end of January, followed by a full o3 model with performance enhancements to attract new investment and users.
In a livestream on Friday, Altman explained that this is the beginning of the next phase of AI, in which “these models will be able to handle increasingly complex tasks that require significant inference.”
OpenAI’s next leap forward in solving complex challenges
Earlier this year, OpenAI introduced the o1 AI model, which is designed to spend more processing time solving complex queries. However, the new model has been proven to be 20% more effective than o1. Ofir Press, a postdoctoral fellow at Princeton University who contributed to the development of SWE-Bench, noted the significant improvement and expressed surprise at the marked increase and uncertainty about how it was achieved. .
Moreover, these models are also great at tackling difficult problems in fields such as science, coding, and mathematics. The company has announced that its upcoming o3 and o3 mini models, which are currently undergoing internal safety testing, will exceed the performance of the o1 model.
The o3 model recorded 96.7% accuracy in the AIME 2024 math competition, with only one mistake. They also achieved 87.7% on GPQA Diamond Scientific Reasoning, which was higher than 70% of typical PhD-level experts.
o3’s notable achievement was that it solved 25.2% of the problems on EpochAI’s Frontier Math benchmark, a significant improvement from the 2% accuracy of the previous model. It also scored 87.5% on the ARC-AGI benchmark, outperforming humans in conceptual reasoning.
“OpenAI o3 ranks 2727th on Codeforces, equating to #175 in the world for competitive human programmers,” X’s post said, documenting his superhuman achievements in AI and technology.
Additionally, o3-mini is a streamlined version of o3, designed to streamline your coding efforts. It provides strong performance with low computational cost and tunable inference settings (low, medium, high), allowing flexibility across a variety of tasks.
The company also introduced a new safety method called deliberative adjustment, which uses the model’s reasoning skills to better identify and manage unsafe prompts. This represents a major advance in AI safety, increasing the accuracy of rejecting harmful requests while avoiding rejecting too many valid requests.
Researchers invited to test the o3 model
OpenAI is inviting outside researchers to apply for early access to its o3 model, and the application process will close on January 10, Reuters reported. The company sparked an AI arms race with the launch of ChatGPT in November 2022, and its growing success, along with new product releases, led to it securing $6.6 billion in funding last October.
Meanwhile, Google is also conducting similar research. Google researcher Noam Shazeer revealed on X that the company has developed its own inference model, Gemini 2.0 Flash Thinking.
As WIRED writes, the competition between OpenAI and Google continues to intensify as both companies strive to improve their AI capabilities. While OpenAI strives to attract more investment and grow its business, Google aims to maintain its dominance in AI research, with both companies looking to build intelligence rather than simply scale up models. We are focusing on improving.
