OpenAI Engineer's Agentic AI: Code Self-Modification

An OpenAI engineer, Weng Jiayi, has proposed a new approach for Agentic AI. Historically, AI improvement relied on more data and computing power for larger models. However, this often makes it hard to explain why AI succeeds or fails. Weng Jiayi's experiment shows another path. In environments with clear goals and feedback, AI can become more powerful by "independently modifying code," not just through training. He had AI repeatedly write strategy code, run tests, and fix failures. Through this process, the AI "developed" a strategy that achieved a perfect score in Atari Breakout and performed well in robot control simulations. The core idea is that experience can be written into code, tests, and logs, creating an evolving software system. This suggests that Agentic AI could move beyond just training larger models, to models that help maintain their own engineering systems. This could change how AI learns and evolves.