RSI is the new AGI, and just as hard to pin down

A new term is crowding out AGI in AI pitch decks: RSI, or recursive self-improvement. TechCrunch reports that it has become the field's favorite way to describe transformative AI, and like AGI before it, almost nobody agrees on what it means. The clean version is a system that upgrades itself in a closed loop, limited only by the compute it can reach, with humans no longer needed in the cycle.

A wave of startups is selling exactly that promise. Richard Socher launched a company called Recursive Superintelligence. Sara Hooker's Adaption is building AutoScientist to automate frontier model training. Doris Xin's Disarray made an ML agent that won 28 Kaggle medals, and Xin argues that with enough compute "we are already there." Even Andrej Karpathy, now at Anthropic, built an Auto-Research system using agent swarms, though he was blunt that it is "not novel, ground-breaking 'research' (yet)." The skeptics are louder than the marketing suggests. Helen Toner of CSET points out that using AI for as much research as possible is not the same as the classic loop with no humans in it, and Sundar Pichai says the industry is not quite there. The full piece is on TechCrunch.

Why it matters

When you see RSI in a pitch or a launch, treat it as a spectrum from "we use AI in our research" to "no humans in the loop," and ask which end the speaker means before you read anything into the claim.