AI Labs Chase Recursive Self-Improvement as Experts Say Full Human-Free Systems Are Not Here Yet
Updated
Updated · TechCrunch · May 28
AI Labs Chase Recursive Self-Improvement as Experts Say Full Human-Free Systems Are Not Here Yet
3 articles · Updated · TechCrunch · May 28
Recursive self-improvement has become a new AI rallying cry, with startups and researchers launching projects aimed at systems that can upgrade their own research, coding and training loops.
Richard Socher’s Recursive Superintelligence, Alex Karpathy’s Auto-Research and Adaption’s AutoScientist all target incremental self-improvement, while Doris Xin’s self-trained agent won 28 Kaggle medals, showing parts of the idea already work.
Google CEO Sundar Pichai said AI is progressing along a continuum but not yet at the next-level acceleration implied by RSI, and Anthropic’s own evidence still shows weaknesses in self-direction, verification and handling ambiguous long tasks.
Anthropic engineers said 5 of 18 respondents thought Mythos could soon substitute for an L4 engineer with better harnessing, but experts such as Helen Toner argue true RSI requires removing humans from the loop entirely.
CSET and METR assessments suggest AI may be nearing an 'adequacy' stage within the next couple of years, yet parity and supremacy remain uncertain because compute limits, engineering hurdles and alignment problems could slow any takeoff.
As AI automates science, are we building a tool for discovery or our own obsolescence?
AI agents are already breaking security rules. How can we trust them to build the next generation?