OpenAI has announced its latest artificial intelligence model, o3, which is designed to tackle complex questions that require step-by-step logical reasoning. The new model scores higher on several measures than its predecessor, o1, including advanced math and science competency, and is three times better at answering difficult mathematical and logic problems.
According to OpenAI CEO Sam Altman, the company views this as the beginning of the next phase of AI development, where models can be used to perform increasingly complex tasks that require reasoning. The model’s new deliberative alignment technique involves training a model with safety specifications and having it reason about its own answer to ensure it doesn’t contravene any rules.
The o3 model is also available in a mini version, called o3-mini, which will be made available for testing by external researchers. OpenAI has not yet made the models publicly available but plans to invite outsiders to apply for testing opportunities.
The announcement comes just days after Google revealed its own reasoning model, Gemini 2.0 Flash Thinking, which achieved high scores on a test measuring agentic abilities. However, OpenAI’s o3 model outperforms Google’s in several areas, including complex coding-related skills and advanced math and science competency.
OpenAI’s o1 model was developed using a similar approach to the new o3 model, with training on step-by-step problem-solving that enables AI models to tackle basic math and logic puzzles. The company plans to use these models to develop so-called AI agents that can reliably solve complex problems on behalf of users.
The announcement highlights the growing competition between OpenAI and Google in the field of artificial intelligence research, as both companies seek to demonstrate their capabilities and attract investment.
Source: https://www.wired.com/story/openai-o3-reasoning-model-google-gemini