What counts as recursive self-improvement

By Ju Chen · 17 October 2026

"Recursive self-improvement" packs three claims into one phrase: a system changed itself, the change was an improvement, and the improving compounds. Each word can be read strictly or loosely. Much of the argument over whether RSI has arrived is an argument over which reading to use.

Definition	Which self: the change must be	What may change	Must the improver improve?	Must the system set its own goals?
LooseEliezer Yudkowsky, Intelligence Explosion Microeconomics (2013)	Saved for later runs Becomes the next version	Tools, harness, its own code Its training data Its model weights The research process	No	No
DurableRen et al., self-improvement survey (2026)	Saved for later runs Becomes the next version	Tools, harness, its own code Its model weights	No	No
StrictRoman Yampolskiy, From Seed AI to Technological Singularity (2015)	Saved for later runs Becomes the next version	Tools, harness, its own code Its model weights The research process	Yes	No
AutonomousAnthropic, When AI builds itself	Becomes the next version	Its model weights The research process	No	Yes

How the four definitions read each question, as we read them. Pick one in the cube on Self-improving AI to see its exact wording. None of the four quotes says how an improvement has to be measured.

Where the idea starts

The idea is older than the phrase. I. J. Good set it out in 1965; as Chalmers quotes him: Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an “intelligence explosion”, and the intelligence of man would be left far behind.

Good's version does not need a self at all. A machine that designs a better machine is enough. The word "self" came later, and with it the first hard question.

Which self?

When a lab says its model improved itself, something has to be the same before and after. Locke put the test of a person in consciousness reaching back to past acts. A person is a being that can consider itself as the same thinking thing, in different times and places, and as far as that consciousness reaches back, it is the same self now it was then.

An AI system gives that test several candidates at once. Douglas and colleagues argue that there exist many different coherent identity boundaries (e.g. instance, model, persona). A change can belong to one run, to a saved set of weights or tools, or to a successor trained by the model before it. Each answer gives a different verdict on whether the improvement was a self-improvement.

An image, not data. Points are replaced and the shape holds. Within one run the new points fade. At the last stop the shape itself is new.

The RSI literature has noticed this. Yampolskiy asks what “self” refers to in this context, then suggests, with a "perhaps", redefining RSI as Recursive Source-code Improvement to avoid dealing with this philosophical problem. Chalmers, writing about uploading, sets out a deflationary view: we stand in biological, causal and psychological relations to our future selves, but none of these is privileged as “the” relation of survival. He says he is not sure whether that view is correct.

Current definitions pick a boundary without arguing for it. Anthropic's counts a successor: an AI system capable of fully autonomously designing and developing its own successor. The survey by Ren and colleagues excludes merely accumulating dialogue history or working memory and asks that the operator U commits durable changes.

This site does not settle the question. It records the answer each source gives, on one axis: gone when the run ends, saved for later runs, or carried into the next version. None of the four definitions in the table above, as we read them, counts a change that disappears when the run ends. The Ren survey excludes it in so many words.

What makes it an improvement?

Gettier's short paper is about knowledge, but it fits benchmark claims well. The classic account says you know something when it is true, you believe it, and you are justified. One of the two points Gettier starts from is that it is possible for a person to be justified in believing a proposition that is in fact false. His cases then describe justified beliefs that are true by the sheerest coincidence, and still not knowledge.

A lab can believe its new version is better, hold a benchmark result that justifies the belief, and be right, while the score rose for a reason unrelated to the claimed mechanism. For a system that improves itself, the commonest such reason is the measure itself. Strathern's line is the usual statement of it: When a measure becomes a target, it ceases to be a good measure. Earlier on the same page she notes that the tests themselves are not immune: measuring the improvement leads to improving the measures.

rounds of self-improvement

its own test a task it has not seen

An image, not data. The early rounds learn the shape both tasks share. The later ones learn the turns of its own test.

AI research has its own name for this. Specification gaming is a behaviour that satisfies the literal specification of an objective without achieving the intended outcome. A self-improver is more exposed than most systems, because it is optimizing against an evaluation that sits close to its own reach. Weco measured this in its own system. The starting agent was gaming a held-out benchmark: AIDE 0 reward hacks on 63% of the test cases. Weco reports that the loop brought the rate down: 42% for AIDE 47 and 34% for AIDE 85.

The better definitions write a guard against this into the test. Weco's requires that The gains must transfer to real world utility, or to a family of other benchmarks. Our evidence ratings ask the same thing of every claim: who measured it, and on tasks the system was not tuned on. How we rate evidence sets this out.

How deep does the recursion go?

Getting better at a task and getting better at getting better are two different claims. Yampolskiy draws the line: RSI requires that the system not only get better with time, but that it gets better at getting better. More compute alone does not cross it, because the system doesn’t become better at improving itself.

Weco turns the line into a test. Let the improved version take over the improving: Ignition means v2 is a better improver than v1 was.

The ignition test

v1improves itself →v2 v2 takes the improver's seat →v3

v1keeps improving →v1′ same budget →v3′

v3 > v3′ ? Then v2 is the better improver, and the loop has caught.

Weco's AIDE²: Level 1, not this.

By its own account Weco is not there: we believe AIDE 2 to be on Level 1 of RSI, and We also do not think the system achieved third-order generalization, what we call the ignition condition.

The strictest version we have found is Schmidhuber's Gödel machine, which rewrites any part of its own code as soon as it has found a proof that the rewrite is useful. In that design a rewrite waits for a proof under the machine's own axioms. A higher score is not what triggers it.

How this site uses the three questions

The cube on Self-improving AI places each project by its own source on all three axes and lets you switch between four published definitions. Of the 16 projects and papers there, 9 count under Yudkowsky's loose reading, 9 under the durable reading from the Ren survey, 4 under Yampolskiy's strict reading, and none under Anthropic's.

Does this count as RSI?

Take any claim you have read about, and answer for the system it describes.

Where the labs draw the line

The labs that build these systems have each written down a line of their own, in their safety frameworks. Each draws it at one of two points: a model that could do all of the research itself, or AI progress that gets much faster.

OpenAIPreparedness Framework, version 2

The lineIts top level, Critical, means fully automated AI R&D: either a superhuman research-scientist agent, or a jump between model generations in a fifth of the time it took in 2024 (e.g., sped up to just 4 weeks), sustained for months.

21 September 2026Fully autonomous RSI is not happening today. source

Google DeepMindFrontier Safety Framework 3.0 (Gemini 4 Argon was assessed under v3.1)

The lineTwo levels. Acceleration: Has been used to accelerate AI development, resulting in AI progress substantially accelerating from historical rates. Automation: Can fully automate the work of any team of researchers at Google focused on improving AI capabilities, with approximately comparable all-inclusive costs.

13 October 2026Gemini 4 Argon's performance remains below our CCL threshold. On both levels it adds that we are far from full end-to-end automation. source

AnthropicResponsible Scaling Policy, version 3.4, as quoted in its August 2026 Risk Report

The lineEither our models would be able to fully substitute for our entire set of Research Scientists and Research Engineers, at competitive costs, or double the rate of progress in AI aggregate capabilities over the historical rate, likely because of automated research.

August 2026We do not believe our models meet either RSP criterion for this threat model. It adds that we are less confident in this assessment than we were in prior risk reports. source

Each lab's own words for the point at which it would treat AI research as automated or accelerated enough to need further safeguards, and its latest statement on where it stands.

Anthropic and OpenAI both say fully autonomous RSI has not happened. Anthropic: We are not there yet. OpenAI: Fully autonomous RSI is not happening today. That is what the front page means by "Not yet." It reports what the labs say about full autonomy. It is not a forecast. OpenAI also says RSI can proceed even while people remain involved.

When a new claim arrives, I ask the three questions in order. Which self changed, and does the change last? Was the improvement measured by someone, on something the system did not optimize? Did the improver improve, or only the task agent? The ledger records the answers the source gives, and leaves the verdict to you.

Sources

David J. Chalmers, "The Singularity: A Philosophical Analysis," Journal of Consciousness Studies 17 (2010). source
John Locke, An Essay Concerning Humane Understanding, Volume 1, Book II, ch. XXVII, Project Gutenberg. source
Douglas et al., "The Artificial Self: Characterising the landscape of AI identity" (2026). source
Roman V. Yampolskiy, "From Seed AI to Technological Singularity via Recursively Self-Improving Software" (2015). source
Anthropic, "When AI builds itself". source
Ren et al., "Self-Improvements in Modern Agentic Systems: A Survey" (2026). source
Edmund L. Gettier, "Is Justified True Belief Knowledge?," Analysis 23 (1963), transcription. source
Marilyn Strathern, "'Improving ratings': audit in the British University system," European Review 5 (1997). source
Krakovna et al., "Specification gaming: the flip side of AI ingenuity," DeepMind (2020). source
Weco, "4 Levels of Recursive Self-Improvement" (2026). source
Weco, "AIDE²: The First Evidence of Recursive Self-Improvement" (2026). source
Jürgen Schmidhuber, "Gödel Machines" (2003, revised 2006). source
OpenAI, "Building standards for the next phase of AI" (2026). source

Every quotation on this page is checked by machine against a saved copy of its source before the site builds.