There’s a new three-letter acronym taking over AI roadmaps, and it’s starting to sound exactly like the last one. According to TechCrunch AI, “recursion” is the word of the moment, with two startups naming themselves after it and a growing pack of labs writing recursive self-improvement (RSI) into their pitch decks. The promise is simple and a little terrifying: an AI that upgrades itself, over and over, until humans are no longer part of the loop.
If that sounds like AGI in a new costume, that’s because it kind of is. RSI has become shorthand for an AI takeoff that nobody can quite define or schedule. What stands out here is how fast the term went from theory to a recruiting magnet.
Who’s chasing it
The names involved are not lightweight. TechCrunch AI reports a clear cluster forming:
- Richard Socher launched a startup literally called Recursive Superintelligence, with RSI as the stated goal: automating “ideation, implementation, and validation of research ideas.”
- Andrej Karpathy, ex-Tesla and OpenAI, is running agent swarms in a public project called Auto-Research. He’s now doing pre-training at Anthropic, which gives him a much bigger sandbox.
- Sara Hooker’s startup Adaption shipped AutoScientist to automate frontier model training.
- Doris Xin’s self-trained ML agent took 28 medals in a Kaggle competition, beating human-trained agents.
Xin’s take cuts through the mysticism. “Given infinite compute and infinite time horizon, we are already there,” she told TechCrunch AI, framing RSI as “meat-and-potatoes engineering” rather than some creative leap. The hard part, she says, is reliability.
The reality check
Here’s where it gets honest. Google CEO Sundar Pichai basically admitted the industry isn’t close. “It’s a continuum, and we are all definitely making progress,” he said, “but… we aren’t quite there yet.”
And yet the continuum is crowded with self-improving systems. One Anthropic lead said “close to 100%” of his team’s Claude Code work was written by the tool. In an internal survey tied to the Mythos preview, five of 18 Anthropic engineers thought the model could soon stand in for a midlevel (L4) programmer.
The catch is the list of weaknesses. Per the report, Claude still struggles with self-managing week-long ambiguous tasks, understanding org priorities, taste, verification, and epistemics. That’s the whole problem. Every weakness is about self-direction, which is exactly what real RSI requires.
Why the definition matters
Helen Toner, who runs Georgetown’s CSET and used to sit on OpenAI’s board, draws the hard line. Using AI to do AI research isn’t RSI. “They’re just using AI for as much as they can,” she told TechCrunch AI. “The classic definition of RSI is really that there are no humans needed.”
She points to METR’s Ajeya Cotra, who breaks the path into stages: “adequacy” (the system keeps doing research after humans leave), “parity” (AI-only matches human-only), and “supremacy” (AI-only beats human-plus-AI). We’re not at the first rung. A CSET expert panel last year split hard, some expecting an imminent explosion, others a slow plateau. The one thing they agreed on: recursion makes the future genuinely hard to predict.
The future cast
Look 1 to 3 years out and the pattern is predictable. RSI will be the headline labs use to raise money and recruit, the same way AGI was. Expect more startups named after it, more benchmark wins like Xin’s Kaggle run, and more internal surveys hinting that mid-level engineering roles are on the clock. But the gap between “AI helps with research” and “AI needs no humans” stays wide, because the missing pieces are judgment and self-direction, not raw compute.
What to do about it:
- Don’t buy the acronym, watch the milestones. Track Cotra’s adequacy-parity-supremacy framing instead of the marketing. It’s a cleaner ruler.
- Audit your own “100% AI-written” claims. Heavy tool use is not autonomy. Know which one you actually have.
- Plan for L4 substitution, not extinction. The near-term squeeze lands on routine mid-level engineering work, where verification and supervision still matter.
- Bet on reliability, not magic. If Xin is right that this is engineering, the winners are whoever makes these agents trustworthy first.
RSI is real as a research direction and oversold as a near-term event. Worth watching closely, worth treating skeptically. Full reporting is at the original source.