OpenAI Engineer Proposes AI Self-Modification for Agentic AI

An OpenAI engineer, Weng Jiayi, has proposed a new idea for Agentic AI. He suggests that AI can become more powerful by independently modifying its own code. This differs from the usual method of simply adding more data and computing power to larger models. Weng Jiayi conducted an experiment where an AI, named Codex, repeatedly wrote and modified its own strategy code. Codex ran the code, reviewed logs and replays, identified failures, and then made corrections. Through this process, Codex developed a Python strategy that achieved a perfect score in the game Atari Breakout. It also performed well in robot control simulations. What's key here is that the AI's learning was written into code and logs, becoming a system that can be read and modified. This suggests that the future of Agentic AI might involve models continuously evolving an engineering system, rather than just training larger models. This could change how AI improves and operates.