AI researchers Beren Millidge, John Schulman, and Charlie O'Neill appeared on a podcast to discuss the potential for "Recursive Self-Improvement (RSI)," a process in which AI conducts research autonomously.
NewsBeren MillidgeJohn SchulmanCharlie O'Neill
AI Researchers Discuss the Timing of Recursive Self-Improvement and AI Job Displacement
This article is a translation. Read the Japanese original
The discussion demonstrated that current training methods using Reinforcement Learning (RL) play a crucial role in improving model capabilities. Schulman noted that RL based on specific reward settings has the potential to exponentially enhance model performance. On the other hand, challenges such as "entropy collapse," where a model falls into repeating existing patterns, and the difficulty of reward design (alignment) in reinforcement learning were also highlighted.
Furthermore, predictions were exchanged regarding when AI might replace white-collar jobs. One researcher predicted that, if not browser-based, AI could function as a remote worker equivalent to a human within about a year. However, opinions were also expressed that further technical challenges remain for automating tasks involving the physical world or long-term decision-making.
Sources
- AI researchers debate how close we are to recursive self-improvement （Hacker News Frontpage、2026-09-11）