AI researchers are pursuing recursive self-improvement (RSI) as the next frontier, with multiple labs building systems that can upgrade themselves, though experts disagree on how close the industry is to achieving it.

Richard Socher launched Recursive Superintelligence with RSI as an explicit goal, stating 'the entire process of ideation, implementation, and validation of research ideas would be automatic.' Alex Karpathy is pursuing a similar vision through Auto-Research, using agent swarms (AI systems that work together) to train language models on simple tasks, while Adaption released AutoScientist to automate frontier training.

RSI refers to an AI system that can continuously upgrade itself; once AI systems manage the upgrade cycle better than humans, the process becomes a closed loop limited only by available compute power. METR researcher Ayeja Cotra distinguishes milestones: 'adequacy' (the system performs research without humans), 'parity' (AI-only systems match human-only systems), and 'supremacy' (AI-only systems outperform human-AI collaboration).

Google CEO Sundar Pichai stated in a recent podcast interview that 'we aren't quite there yet' regarding RSI, though Helen Toner from Georgetown's Center for Security and Emerging Technology noted that simply using AI tools for research differs from the classical RSI definition, which requires that 'there are no humans needed.' Researchers fundamentally disagree on whether recursive systems are imminent or will plateau.