source: techcrunch ai: rsi is the new agi — and it’s just as hard to pin down
level: business
recursive self-improvement, or rsi, refers to an ai system that can continuously upgrade itself. once ai can manage the upgrade cycle better than humans, the process becomes a closed loop limited only by compute power. richard socher's recursive superintelligence and alex karpathy's auto-research project are among the efforts chasing this goal. karpathy's work has mostly made minor improvements on a gpt-2 scale model, but it has inspired others. adaption's autoscientist and disarray's kaggle-winning agent also aim to automate frontier training and machine learning tasks.
there is evidence that the industry is not close to meaningful recursive systems. google ceo sundar pichai said progress is being made but the kind of acceleration people describe is not here yet. anthropic's claude code writes nearly all of its team's code, but engineers still direct it. a survey tied to anthropic's mythos preview found some engineers believe it could substitute for a mid-level programmer, but it struggles with self-direction, understanding priorities, and verification. these weaknesses are central to rsi.
experts are split on how far away rsi is. georgetown's center for security and emerging technology found some expect an imminent superintelligence explosion while others foresee slower progress. helen toner distinguishes between using ai for research and true rsi where no humans are needed. ayeja cotra's milestones include adequacy, parity, and supremacy. cotra believes ai is close to adequacy, where it can produce some work on its own, but parity is less clear. toner notes that handing off the entire research process will require overcoming significant engineering and alignment challenges.
why it matters: understanding rsi helps gauge how close ai is to automating its own development, which could dramatically accelerate progress and reshape the field.
source: techcrunch ai: rsi is the new agi — and it’s just as hard to pin down