# Could AI build a better version of itself?

Recursive self-improvement — RSI — is the idea that an AI system could help design its
successor well enough that each generation makes the next one faster to build. Google DeepMind,
OpenAI and Anthropic each now say, in their own public statements, that some version of this is
already under way. None of the three has shown the full version.

*Marius Comper · 13 September 2026*

## The idea inside the acronym

In 1965, I. J. Good described an "intelligence explosion": a machine smart enough to design a
better machine, which could then design a better one still. Recursive self-improvement is the
mechanism that idea assumes — the loop, not the outcome. It runs on a spectrum from using AI as
a coding tool to a system designing its own successor with no human involved.

## Four rungs, four different kinds of evidence

This ladder is this page's own organizing device, not an industry-agreed scale (METR
deliberately avoids defining RSI as a technical threshold).

1. **Adoption and output.** Anthropic reports over 80% of its production code authored by Claude
   (May 2026) and roughly 8x more code shipped per quarter than in 2024. Shows adoption and
   volume, not causal research acceleration — Anthropic itself calls lines of code "an imperfect
   measure."
2. **One documented feedback loop.** DeepMind's AlphaEvolve (14 May 2025) sped up a training
   kernel by 23%, cutting Gemini's overall training time by about 1%; AlphaEvolve runs on Gemini
   models. One documented, bounded example — not a repeating cycle.
3. **A company's account of its newest release.** Google describes Gemini 3.8 Flash (3 September
   2026) as accelerated by agent loops that "recursively evaluate and refine the underlying
   models" (via Fortune). One company's characterization of one release, not independently
   audited.
4. **Full recursive self-improvement.** No public evidence found. Anthropic: "We are not there
   yet, and recursive self-improvement is not inevitable." OpenAI chief scientist Jakub Pachocki
   (6 September 2026): internal results give him "a strong expectation" this could be sustained
   into RSI, and "calls for extreme caution" — an attributed expectation, not a report of the
   present.

## What has actually happened, in order

- **1965** — I. J. Good names "intelligence explosion."
- **14 May 2025** — AlphaEvolve ships; the loop closes once.
- **18 August 2026** — A Princeton study (Kirgis, Kapoor) finds AI agents fail at open-ended
  research-conference-level questions.
- **3 September 2026** — Google's release language for Gemini 3.8 Flash changes.
- **6 September 2026** — Pachocki publishes "An Alien Mind" under his own name.

## The rumor that prompted this page

On 12 September 2026, [Andrew Curran](https://x.com/AndrewCurran_) [posted](https://x.com/AndrewCurran_/status/2098809612787171331)
that rumors say Gemini 4 "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](https://x.com/AndrewCurran_/status/2098812082691158285): "Next week." A reply
speculated this meant DeepMind had "achieved RSI." These are social posts, cited here because
they are the claim that prompted this page, not as evidence about DeepMind's internal state. Even
if pretraining finished early, that alone would not show that an AI, not a person, found the
improvement, or that the improvement was used to design a further one. Settling this would
require training logs, compute accounting and an independent audit that nobody outside the lab
has. The public evidence points neither way.

## The people building it do not agree with each other

Pachocki pairs urgency with uncertainty ("as the systems become more capable, the results become
harder to interpret"). Anthropic's own account cites a "research judgment gap" and at least one
automated research result that "didn't transfer cleanly to production-scale models." A Princeton
study found AI agents "ran bizarre experiments" on open-ended research and were rejected by human
reviewers; Sayash Kapoor notes reinforcement learning excels at automatically-scorable tasks, not
open-ended research. Anthropic cofounder Jack Clark: "There's a certain absence of valuable,
intuitive creativity in today's AI systems."

## For the deep dive

- DeepMind, [AlphaEvolve announcement](https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/) (14 May 2025)
- Anthropic Institute, ["When AI builds itself"](https://www.anthropic.com/institute/recursive-self-improvement) (May 2026)
- Jakub Pachocki, ["An Alien Mind"](https://openai.com/index/an-alien-mind/) (6 Sept. 2026)
- METR, [Time Horizon metric](https://metr.org/time-horizons/) (updated 8 May 2026)
- METR, ["Economics of recursive self-improvement"](https://metr.org/notes/2026-07-22-economics-of-recursive-self-improvement/) (22 Jul. 2026)
- MIT Technology Review, [on the Princeton study](https://www.technologyreview.com/2026/08/18/1142188/ai-recursive-self-improvement/) (18 Aug. 2026)
- Fortune, [on Google's Gemini 3.8 Flash release](https://fortune.com/2026/09/03/google-shipped-four-gemini-flash-models-in-106-days-but-its-flagship-frontier-model-is-still-nowhere-to-be-seen/) (3 Sept. 2026)

## Method and limits

No sentence on this page asserts that Google DeepMind, OpenAI or Anthropic has achieved full,
autonomous recursive self-improvement. Company statements, attributed opinions and personal
social posts are marked as three distinct kinds of claim throughout.

— [Read the full page](https://mariuscomper.uk/recursive-self-improvement/en/)