The term 'recursion' has joined the ranks of AI buzzwords, with two startups and many more labs vying to develop Recursive Self-Improvement (RSI) – an AI system that can continuously upgrade itself. Once self-improvement cycles surpass human capabilities, humans may no longer be necessary.

Richard Socher's Recursive Superintelligence aims for full automation in ideation and validation processes. Meanwhile, Alex Karpathy is using agent swarms to incrementally improve models, while Adaption launched AutoScientist to train agents on frontier tasks, potentially advancing towards RSI.

Doris Xin’s self-trained machine learning agent won a Kaggle competition, highlighting the challenge of reliability in RSI systems. However, significant weaknesses persist, especially in self-direction and complex task management.

The AI industry remains far from achieving meaningful recursive systems, with experts predicting both imminent explosions and slower progress. Helen Toner at Georgetown’s Center for Security and Emerging Technology noted that current use of AI tools to do research falls short of the classic RSI definition, which requires no human intervention.