In a head-to-head comparison, o3-pro was far less reliable and secure, and reasoned excessively compared to GPT-4o.

MIT researchers developed SEAL, a framework that lets language models continuously learn new knowledge and tasks.