Recursive self-improvement explained: what RSI means, who is pursuing it and where the line is

Recursive self-improvement, or RSI, is the idea that an AI system could improve its own design, and that the improved system could then improve itself again, in a loop that speeds up each time. It is the mechanism behind most arguments that AI could move from human-level to far beyond human-level quickly. As of November 2026 it is a goal for some labs, a worry for others, and a defined threshold in at least one lab's safety policy, but nobody has shown the loop running without people.

What recursive self-improvement means

The simple version comes from TechCrunch's May 28, 2026 explainer: an AI system improves itself in an ongoing loop, and once it handles upgrades better than humans, the loop is limited mainly by available compute and people become unnecessary.

That last clause is the strict reading. Helen Toner of CSET, a former OpenAI board member, told TechCrunch that using AI for AI research does not by itself count as RSI. The classic definition requires that "there are no humans needed." Under that standard, today's AI-assisted research is something else: automation of parts of the work, with people still choosing what to build and judging whether it worked.

Three levels that are easy to confuse

Writers use the term for three different things, and mixing them causes most of the confusion.

• AI as a research tool. Models write code, run experiments and summarize papers for human researchers. This is happening now. A Claude Code lead at Anthropic estimated "close to 100%" of his team's code was written by the tool, per TechCrunch's account of a January 2026 statement.
• AI as a research worker. Models do whole research tasks, such as ideating, implementing and validating an experiment, with people reviewing. Several startups have made this an explicit goal, TechCrunch reports.
• AI with no humans needed. The full loop. No lab has claimed it.

Ajeya Cotra of METR, quoted by TechCrunch, offered a finer scale of three milestones: adequacy, where AI can do research without humans though less valuably; parity, where AI-only research matches human-only research; and supremacy, where AI-only research beats human and AI collaboration. She thinks adequacy is very close and may already have been reached, expects it "in the next couple years," and said parity would "massively accelerate the pace of AI progress."

Who is pursuing it

TechCrunch's reporting names several efforts as of May 2026. Richard Socher launched a company called Recursive Superintelligence, saying at launch: "Our main focus is to build truly recursive, self-improving superintelligence at scale." Sara Hooker's company Adaption released AutoScientist on May 13, 2026, training agents to make incremental improvements with the aim of automating training for full frontier models. Andrej Karpathy shared a project on GitHub called Auto-Research, which uses agent swarms to train language models on simple tasks. In March he said: "It's not novel, ground-breaking 'research' (yet)."

Google's Sundar Pichai, on a New York Times podcast on May 22, 2026, said "we aren't quite there yet," describing RSI as a further level of acceleration.

Where the safety policies draw the line

Anthropic is the one lab whose public policy defines a version of RSI as a threshold. Its Responsible Scaling Policy treats "dramatically accelerated progress attributable to AI" as a capability level. In version 3.1, effective April 2, 2026, Anthropic clarified that the example, compressing two years of 2018 to 2024 AI progress into a single year, refers to "doubling the rate of progress in aggregate AI capabilities," not "doubling the productivity of researchers." Version 3.4, effective July 8, 2026, revised the automated R&D threshold, per the policy page.

On November 2, 2026, an amended IPO filing said Anthropic's internal estimates indicate it "may imminently meet" that definition, though it has not designated any current model as exceeding it. We cover the filing in our report on the S-1/A. For how the other labs frame the same risk, see our comparison of the frontier safety frameworks.

Why it matters for the superintelligence debate

The argument goes in steps. If AI can do AI research as well as people, the number of researchers is no longer fixed. Copies of a model can run in parallel and work around the clock. Better models then produce better research, and the cycle shortens. Proponents of fast timelines lean on this chain.

Skeptics push back at each link. Toner told TechCrunch that many RSI pursuers assume progress follows "a pretty smooth ladder" that keeps scaling. Real research hits bottlenecks that more thinking cannot remove, including compute, data, and experiments that take real time. A CSET study from 2025, as TechCrunch describes it, found experts split between expecting a near-term intelligence explosion and expecting slower progress with a plateau, though all agreed recursion makes the future hard to predict.

What would count as evidence

Reasonable readers can ask for five signs before accepting that the loop has started.

• A measured speedup. The rate of progress with the lab's tools compared with its rate without them, ideally measured by an outside party.
• Falling human involvement. Fewer human decisions per capability gain, documented.
• Unrecognized jumps. Anthropic's filing, as quoted by TechCrunch, says it has "in the past not identified" major acceleration until after it occurred. More such disclosures would matter.
• Policy changes. Labs revising their own thresholds, as Anthropic says it is considering.
• Independent evaluations. Results from groups like METR and the UK AI Security Institute on automated research tasks.

Our take

RSI is a useful frame and an overused word. Like AGI, it has no agreed definition, and TechCrunch's own conclusion was that RSI has not arrived. The right posture is neither alarm nor dismissal: ask which of the three levels a claim describes, who measured it, and whether the person claiming it benefits from the claim. The disclosure in Anthropic's filing gives readers the first lab-authored threshold to watch, and that is a more useful thing to track than any single headline.