Could AI build a better version of itself?

The idea inside the acronym

In 1965, the mathematician I. J. Good, who had worked with Alan Turing on wartime codebreaking, described what he called an "intelligence explosion": a machine smart enough to design a better machine, which could then design a better one still. Good was not writing fiction. He was pointing out a possible consequence of building machines that reason well enough to improve their own design.

Recursive self-improvement is the mechanism that intelligence explosion assumes — the loop, not the outcome. It is not one threshold that a system crosses on a given afternoon. It runs on a spectrum: at one end, engineers use an AI tool the way they use a compiler, to write code faster; at the other end, a system designs, trains and deploys its own successor with no person deciding what happens next. Almost everything interesting is happening between those two ends, and the honest answer to "has anyone achieved RSI" depends entirely on which point on that spectrum the question means.

Four rungs, four different kinds of evidence

The ladder below is this page's own way of sorting the evidence, not an industry-agreed scale. METR, a research organization that measures AI capability and whose work this page draws on, deliberately avoids defining RSI as a technical threshold at all, and focuses instead on how strong the feedback is and whether it is strong enough to be self-sustaining. Keep that in mind reading what follows: each rung shows what a specific, cited claim supports — no more. The rungs are ordered by how close a claim comes to full recursive self-improvement, not by how rigorously it has been checked: rung 3 rests on a single company's own account of itself, while rung 2's narrower claim comes with an independently checkable number.

1. Rung 1 · Adoption and output

   A company reports its engineers rely on AI to write code, at scale.

   Anthropic reports that more than 80% of its production code was authored by Claude as of May 2026, and that engineers were shipping roughly eight times as much code per quarter as in 2024.

   What it shows: adoption and volume. What it doesn't show: that more code means better or faster research — Anthropic's own account calls lines of code "an imperfect measure," since it counts quantity, not quality.

2. Rung 2 · One documented feedback loop

   An AI system found a real improvement that fed back into its own training.

   DeepMind's AlphaEvolve, announced in May 2025, discovered a way to split a large matrix-multiplication operation into smaller pieces, speeding up that operation inside Gemini's training pipeline by 23% — DeepMind reports this cut Gemini's overall training time by about 1%. AlphaEvolve itself runs on Gemini models, so the improvement flowed back into training the system that produced it.

   What it shows: one real, bounded, published example of an AI-found improvement entering an AI's own development. What it doesn't show: a repeating cycle — there is no public record of a later, AlphaEvolve-improved Gemini being used to improve AlphaEvolve again.

3. Rung 3 · A company's account of one release

   Google's Gemini 3.8 Flash release notes describe agent loops that recursively refine underlying models.

   Google's release notes for Gemini 3.8 Flash, on 3 September 2026, describe the model as "further accelerated" by "long-running AI-agent loops that recursively evaluate and refine the underlying models" — reported language that goes further than Google's own description of a two-agent loop building a game in May 2026, or a three-agent loop training a robotics model in August 2026. A DeepMind researcher, Shunyu Yao, posted that the release represented "one small step for model, one giant leap for RSI."

   What it shows: one company's account of its 3 September release, reported by a business outlet and not independently audited. What it doesn't show: that the method recurs across generations, is routine, or describes how Gemini 4 Argon was trained.

4. Rung 4 · Full recursive self-improvement

   Feedback strong and autonomous enough to keep accelerating on its own.

   This is the version people mean when they say "RSI" as a finished thing: recursion (output feeding back into development), autonomy (how little human direction is left) and acceleration (the rate of progress itself increasing) all present together, and self-sustaining. No public evidence for this rung was found for any lab. Anthropic's own account from May 2026 is direct about it: "We are not there yet, and recursive self-improvement is not inevitable." OpenAI's chief scientist, Jakub Pachocki, wrote under his own name in September 2026 that internal results give him "a strong expectation that this speed of progress could be sustained into recursive self-improvement" — an expectation about where things are heading, based on results the company has not published, not a claim that it has already happened.

   What it shows: no lab has published evidence for this rung; Anthropic now says a form of RSI dynamics has started appearing. What it doesn't show: anything about labs not covered in this page's sources.

What has happened, in order

1. 1965

   The concept gets a name

   I. J. Good describes an "intelligence explosion" — the outcome a self-improving machine could produce, not yet a claim about any real system.

2. 14 May 2025

   AlphaEvolve ships, and the loop closes once

   DeepMind's evolutionary coding agent optimizes a kernel used in training Gemini — the models AlphaEvolve itself runs on. A documented, bounded case of AI work feeding back into AI development.

3. 18 August 2026

   A Princeton study finds AI agents still fail at open-ended research

   Researchers give AI agents unpublished, research-grade questions. Both resulting papers are rejected by the human authors of the originals — a data point for how far autonomous research judgment still has to go. Discussed in full below.

4. 3 September 2026

   Google's language changes

   Gemini 3.8 Flash ships with release language describing agent loops that "recursively evaluate and refine the underlying models" — more direct than Google's two earlier, narrower descriptions of self-improvement loops.

5. 6 September 2026

   OpenAI's chief scientist writes under his own name

   Jakub Pachocki publishes "An Alien Mind," saying internal results give him a "strong expectation" that current progress could sustain into recursive self-improvement, and that this moment "calls for extreme caution."

6. 12 September 2026

   Anthropic's CEO writes under his own name

   Dario Amodei publishes "We Must Pace the Frontier," saying that since roughly summer 2026, AI has been advancing drastically faster precisely because AI increasingly builds the next generation of AI — a dynamic he explicitly names recursive self-improvement, "starting to happen across the industry, including at Anthropic." That goes further than May's "not there yet," but it remains a public position about the industry's direction, not evidence of a fully autonomous, self-sustaining loop (rung 4).

7. 30 September 2026

   Google announces Gemini 4 Argon

   The official announcement Gemini 4 Argon describes a phased rollout to trusted cyber defenders and agent workflows for optimizing algorithms, saving memory, and migrating code. It does not confirm that AI accelerated Argon’s own pretraining or that an RSI loop has been closed.

The rumor that prompted this page

On 12 September 2026, the AI commentator Andrew Curran posted that rumors had been "circulating for three days that Gemini 4 has finished pretraining early, in part due to new discoveries GDM made during the training run. Nothing confirmed, and impossible to know what's true." Asked how soon, he replied: "Next week." A reply to a separate question on the same thread — what's the most optimistic thing you've heard — answered: "That they successfully closed the loop and have achieved RSI."

This claim cannot be checked from outside. Even if Gemini 4's pretraining finished ahead of schedule, that alone would not show that an AI system, rather than a person, found the improvement that saved the time. It would not show that the improvement was then used to design a further one. And it would not show that any of this could continue with a growing share of AI direction rather than human direction. Settling "RSI" — at rung 4, not rung 1 — needs access nobody outside the lab has: training logs, a record of who or what proposed each change, compute accounting, an independent technical audit. Without that access, the public evidence points neither way.

On 30 September 2026, Google officially announced Gemini 4 Argon, with an initial phased rollout to trusted cyber defenders. The announcement describes agents optimizing quantum algorithms, finding memory savings, and migrating code. It does not say Argon accelerated its own pretraining through AI discoveries, or that Google has closed a recursive self-improvement loop. The public “next week” prediction did not materialize: the announcement came 18 days after the 12 September post. Gemini 4’s existence is confirmed; the pretraining and RSI claims remain unconfirmed.

The people building it do not agree with each other

The disagreement over how close RSI is runs through the people doing the work, not just outside commentary. Pachocki's own essay pairs urgency with an admission of uncertainty: progress, he writes, is "grown more than designed," and "as the systems become more capable, the results become harder to interpret." Anthropic's public account of its own systems carries a parallel set of caveats: a "research judgment gap" between Claude and human researchers, and at least one automated research result that "didn't transfer cleanly to production-scale models."

On 12 September, Anthropic CEO Dario Amodei wrote that recursive self-improvement "is starting to happen across the industry, including at Anthropic." That is a public position about the industry's direction, not a technical report of a fully autonomous, self-sustaining loop; it does not by itself move the evidence to rung 4.

The clearest counter-argument comes from a study, not an opinion. Researchers led by Peter Kirgis and Sayash Kapoor at Princeton set AI agents loose on unpublished, research-conference-level questions. Both resulting papers were rejected by the human researchers who had posed the original problems: the agents, the study found, "ran bizarre experiments," struggled to write intelligibly about their own work, committed to weak approaches too quickly, and managed their own resources and feedback poorly. Kapoor's explanation points at a limit in the current method, not simply a lack of scale: reinforcement learning is good at narrow tasks a computer can score automatically, and open-ended research is not one of those tasks, because "success can't be checked automatically." Jack Clark, an Anthropic cofounder, made a similar point about the same underlying problem: "There's a certain absence of valuable, intuitive creativity in today's AI systems."

In a podcast published on 11 September 2026, John Schulman (chief scientist at Thinking Machines), Beren Millidge (chief technology officer at Zyphra) and Charlie O'Neill (head of model training at Baseten) discussed how close AI is to recursive self-improvement. One concern raised: further progress may need increasingly difficult work to build training environments realistic enough to teach a model its next skill.

Put together, the honest picture is not "RSI is here" or "RSI is far away." It is that the people closest to the work are watching the same evidence and reaching different conclusions about how fast the remaining gap will close — which is itself a reason to read the next headline about it carefully.

For the deep dive

Every claim on this page traces to one of the sources below. Where a claim rests on one company's own account, an individual's post, or a news report rather than an independent audit, the table says so, since that shapes how much weight it can bear.

Method and limits

This page reports what named, dated sources say, and marks plainly where a claim is a company's own account of its own systems, an individual's attributed opinion, or a personal social-media post — those are three different kinds of claim with different evidentiary weight. No sentence here asserts that Google DeepMind, OpenAI or Anthropic has achieved full, autonomous recursive self-improvement, because none of the sources above support that claim.

This is a dated account of a fast-moving subject. Revisit it if a lab publishes stronger evidence at rung 3 or 4, if the September 2026 rumor is confirmed or credibly denied, or if a major new independent study changes the balance of evidence described above.