RSI becomes AI's new buzzword: Startups chase recursive self-improvement

Recursive self-improvement (RSI) has become the new AGI—a three-letter buzzword for AI that continuously upgrades itself. Startups like Recursive Superintelligence, founded by AI researcher Richard Socher, explicitly target RSI: "Our main focus is to build truly recursive, self-improving superintelligence at scale," Socher told TechCrunch. Alex Karpathy, a legendary figure from Tesla and OpenAI, is working on Auto-Research, using agent swarms to train LLMs on simple tasks. So far, it only makes minor improvements on a GPT-2-scale model, but Karpathy has been transparent about milestones via GitHub. Adaption (founded by Sara Hooker from Cohere and Google) launched AutoScientist to automate frontier training with incremental agent-based improvements. Meanwhile, Disarray founder Doris Xin's self-trained ML agent won 28 medals in a Kaggle competition, suggesting reliability may be the main hurdle.

There's evidence the industry is closer to RSI than many think. Anthropic's lead programmer for Claude Code admitted that "close to 100%" of their team's code was written by the tool—effectively AI writing itself to improve itself. A recent survey tied to Anthropic's Mythos preview found that 5 out of 18 engineers believed that version could substitute for an L4 engineer (mid-level programmer). However, weaknesses include managing week-long ambiguous tasks, understanding org priorities, and verification. Google CEO Sundar Pichai acknowledged progress but noted, "We aren't quite there yet." The race is on, but full recursive automation remains a continuum.