# Recursive Self-Improvement (RSI): What It Means When AI Builds AI

> Recursive self-improvement (RSI) is AI taking over the work of building better AI. What it means, what labs showed in 2026, and why experts disagree.

Canonical: https://sibuilders.io/guides/recursive-self-improvement  
Last updated: 2026-11-18

SI Builders Editorial · Published November 9, 2026 · Updated November 18, 2026

**Recursive self-improvement (RSI)** is a process in which an AI system improves the methods used to build AI, so that each generation is better at producing the next one. In the strict version of the idea, no human is needed inside the loop.

As of November 2026, no lab has disclosed a system that does this. AI now writes most of the code at several frontier labs and can run experiments for days at a time, but people still choose the research direction, define what counts as success and decide what gets trained and shipped.

## Key facts

- **RSI stands for recursive self-improvement.** It names a process, not a product. *Recursive Superintelligence* is a startup named after the idea (more on that below).
- **The idea is 60 years old.** The statistician I. J. Good described an "intelligence explosion" driven by machines that design better machines in 1965.
- **2026 is the year labs started publishing numbers.** [Anthropic](https://sibuilders.io/companies/anthropic) reported that Claude authored more than 80% of its merged code as of May. [OpenAI](https://sibuilders.io/companies/openai) reported 3.1 agent-workdays for every human workday in its research organization as of mid-August.
- **The loop is not closed.** A RAND review published in October 2026 found that no disclosed system sets its own research agenda or designs its own successor.
- **There is now an outside measurement.** In November 2026, METR estimated from inside access that Anthropic is improving its models about twice as fast as before October 2025, mostly because of AI's help with research.
- **Whether it would take off is disputed.** A September 2026 paper by nine economists estimated that today's feedback loops are not yet strong enough to sustain themselves, while noting that they appear to be getting stronger.
- **Labs say the safety problem is unsolved.** OpenAI wrote in September that it does "not yet know how to safely get all the way to aligned, full RSI."

## RSI, recursive self-improvement and "recursive superintelligence"

People search for this topic under several names, so it helps to separate them.

**Recursive self-improvement** is the technical term, and it is what the abbreviation RSI means in AI. (Outside AI, RSI more often means repetitive strain injury or the Relative Strength Index used in trading. Neither is related.)

**"Recursive superintelligence"** is not a separate technical concept. It is used in two ways:

1. As a proper name. [Recursive Superintelligence](https://tech.eu/2026/05/13/recursive-superintelligence-emerges-from-stealth-with-650m-raise/) is a startup that launched publicly on May 13, 2026 with $650 million in funding and the stated aim of building AI that improves itself without human intervention.
2. As shorthand for the outcome some researchers expect RSI to produce: Super Intelligence reached by a system improving itself over and over.

Related phrases you will see used almost interchangeably: *self-improving AI*, *AI building AI*, *automated AI R&D*, *software intelligence explosion* and, in older writing, *seed AI*.

On SI Builders, "Super Intelligence" in this guide means the research sense of the term: capability well beyond human level. It is separate from the U.S. federal government's use of "SI" as its label for AI. Our guide to [what Super Intelligence means](https://sibuilders.io/guides/what-is-super-intelligence) explains the difference.

## What recursive self-improvement means

Most AI improvement is not recursive. A chess engine that gets stronger by playing itself is self-improving, but the training algorithm that makes it stronger stays fixed, and a human wrote it. The same is true of a language model fine-tuned on its own outputs.

Recursion starts when the thing being improved is the improver itself: the training code, the research agent, the architecture search, the evaluation pipeline. If a better system is also better at making those changes, the gains can compound.

That leaves a lot of room for disagreement over where "AI helping with AI research" ends and RSI begins. The Associated Press noted in September 2026 that leading companies do not share a definition: some count any contribution from AI to model improvement, while others reserve the term for fully autonomous work.

| Source | What counts as RSI |
| --- | --- |
| [Anthropic](https://www.anthropic.com/institute/recursive-self-improvement) (2026) | An AI system "fully autonomously designing and developing its own successor." The authors say this has not happened and is not inevitable. |
| [OpenAI](https://openai.com/index/research-acceleration-view-inside-openai/) (2026) | A direction of travel that can be measured as research acceleration, with "full RSI" as the end state. |
| [RAND](https://www.rand.org/pubs/perspectives/PEA5068-1.html) (2026) | AI systems contributing to their own improvement, so that each generation of capability speeds up development of the next. |
| Ajeya Cotra, METR, as [reported by TechCrunch](https://techcrunch.com/2026/05/28/rsi-is-the-new-agi-and-its-just-as-hard-to-pin-down/) | Three milestones for AI-only research: adequacy (it can do research alone), parity (it matches human-only teams) and supremacy (it beats human-AI teams). |
| Helen Toner, Georgetown CSET, in the same report | Using AI tools for AI research does not qualify. Classic RSI means humans are no longer needed. |
| [Academic survey](https://arxiv.org/abs/2607.07663) of 1,250 papers (2026) | Separates *bounded self-refinement*, where the system improves against a fixed outside yardstick, from *open-ended RSI*, where it also changes the yardstick. |

### Four rungs of loop closure

A simple way to place any claim about RSI is to ask how much of the loop the AI controls. This ladder is our simplification of the taxonomies above.

1. **AI as a tool.** Autocomplete and chat assistance. Humans do the research.
2. **AI as an agent on assigned work.** Agents write code, run experiments and debug infrastructure for hours or days, on tasks people define. This is where frontier labs report being in 2026.
3. **AI choosing what to try next.** The system proposes and tests its own changes against a goal and scoring method that humans set. Research systems such as AlphaEvolve and the Darwin Gödel Machine work this way in narrow domains.
4. **Closed loop.** The system sets the research agenda, builds and trains a successor, judges whether it is better, and repeats. This is the strict meaning of RSI. It has not been demonstrated.

## A short history of the idea

| Year | Milestone |
| --- | --- |
| 1950 | Alan Turing suggests building a "child machine" that learns, instead of programming an adult mind directly. |
| 1965 | [I. J. Good](https://doi.org/10.1016/S0065-2458(08)60418-0) argues that a machine better than humans at designing machines would trigger an "intelligence explosion," and calls such a machine "the last invention that man need ever make." |
| 2003 | Jürgen Schmidhuber proposes the [Gödel machine](https://arxiv.org/abs/cs/0309048), a program that rewrites its own code only after proving the rewrite is an improvement. |
| 2000s | Eliezer Yudkowsky develops the concepts of "seed AI" and [recursive self-improvement](https://www.lesswrong.com/w/recursive-self-improvement) as the phrase is used today. |
| 2023 | Microsoft and Stanford researchers publish the [Self-Taught Optimizer (STOP)](https://arxiv.org/abs/2310.02304), a scaffolding program that improves itself using a fixed language model. |
| May 2025 | Google DeepMind announces [AlphaEvolve](https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/). Researchers publish the [Darwin Gödel Machine](https://arxiv.org/abs/2505.22954). |
| January 2026 | Georgetown's CSET publishes [When AI Builds AI](https://cset.georgetown.edu/publication/when-ai-builds-ai/), a workshop report on automated AI R&D. |
| April 2026 | ICLR hosts what its organizers call possibly the first [workshop devoted entirely to RSI](https://recursive-workshop.github.io/), in Rio de Janeiro. |
| May 2026 | Recursive Superintelligence launches with $650 million. |
| June 2026 | Anthropic publishes [When AI builds itself](https://www.anthropic.com/institute/recursive-self-improvement). Google DeepMind researchers publish [From AGI to ASI](https://arxiv.org/abs/2606.12683). |
| September 2026 | OpenAI publishes internal research-acceleration data. Dario Amodei publishes [We Must Pace the Frontier](https://darioamodei.com/post/we-must-pace-the-frontier). A bill restricting recursively self-improving models is announced in the U.S. House. |
| October 2026 | RAND publishes its [review of the empirical evidence](https://www.rand.org/pubs/perspectives/PEA5068-1.html). |
| November 2026 | Anthropic tells investors that future models may soon reach its internal threshold for AI-driven acceleration. METR publishes the [first independent assessment](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment) of AI speeding up a lab's own research. |

For most of that history RSI was a thought experiment. What changed in 2025 and 2026 is that AI became good at writing code, and AI is made of code. Richard Socher, the chief executive of Recursive Superintelligence, put the argument to [The New York Times](https://www.thestar.com.my/tech/tech-news/2026/05/15/notable-researchers-join-us4bil-effort-to-build-self-improving-ai) in two sentences: "AI is code. And now, AI can code."

## Why RSI matters for Super Intelligence

Good's 1965 argument is still the core of the case. Designing AI systems is intellectual work. If a system becomes better than people at that work, it can design a better system, which can design a better one again. If each step also arrives faster than the last, capability could rise far beyond human level in a short time.

That is why RSI sits at the center of forecasts about Super Intelligence. A June 2026 report by Google DeepMind researchers, [From AGI to ASI](https://arxiv.org/abs/2606.12683), lists recursive improvement as one of four possible routes past human-level AI, alongside scaling existing systems, new paradigms and large collectives of AI agents. The authors also describe frictions that could slow every route and treat their size as an open question.

RSI is a proposed route, and a contested one. Nothing in this guide should be read as a claim that any current system has reached Super Intelligence. For how the term relates to ordinary AI, see [Super Intelligence vs. AI](https://sibuilders.io/guides/super-intelligence-vs-ai).

## The evidence in 2026

### Inside frontier labs

**Anthropic.** The Anthropic Institute's essay [When AI builds itself](https://www.anthropic.com/institute/recursive-self-improvement), published in June and updated in September, is the most detailed public account from inside a lab. It reports that more than 80% of merged code was written by [Claude](https://sibuilders.io/models/claude) as of May 2026, up from low single digits before [Claude Code](https://sibuilders.io/tools/claude-code) launched in February 2025. On one internal test of speeding up model training, the average speedup rose from about 3x with a May 2025 model to about 52x with a model from April 2026. The authors caution that code volume overstates real productivity and say they expect the likeliest path is compounding efficiency gains with humans still setting direction.

**OpenAI.** On September 6, 2026 the company published [Research acceleration: The view inside OpenAI](https://openai.com/index/research-acceleration-view-inside-openai/). It says OpenAI has reached the "automated research intern" milestone it set for September 2026: a system that completes well-scoped research tasks, including multi-day ones, under human direction. The next stated goal is an automated AI researcher by March 2028. [Fortune's account](https://fortune.com/2026/09/08/openai-rsi-progress-jakub-pachoki-warns-dangers-slowdown-safety-rules/) of the post notes two limits the company disclosed: more than half of successful four-to-eight-hour agent tasks still needed a human to step in, and high-level research planning was a very small share of what researchers delegated.

**Google DeepMind.** [AlphaEvolve](https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/) uses [Gemini](https://sibuilders.io/models/gemini) models to propose code and automated evaluators to score it. [Google DeepMind](https://sibuilders.io/companies/google-deepmind) says one result sped up a kernel used in Gemini training by 23%, cutting training time by 1%, which makes it a model improving the pipeline that trains that model's successors. Chief executive [Demis Hassabis](https://sibuilders.io/builders/demis-hassabis) has [argued](https://www.business-standard.com/technology/artificial-intelligence/recursive-self-improvement-explained-is-ai-building-ai-the-path-to-agi-126062200661_1.html) that this kind of loop works in games, math and code because results can be checked automatically, and that the real world is much messier. TechCrunch quoted Alphabet chief executive Sundar Pichai in May 2026 as saying the industry "aren't quite there yet."

**Meta, Microsoft and xAI.** Mark Zuckerberg [wrote in July 2025](https://www.meta.com/superintelligence/) that Meta had begun to see "glimpses of our AI systems improving themselves," and told investors that [Meta Superintelligence Labs](https://sibuilders.io/companies/meta-superintelligence-labs) would focus on self-improving AI. [Mustafa Suleyman](https://sibuilders.io/builders/mustafa-suleyman) told [Axios in November 2025](https://www.axios.com/2025/11/06/microsoft-mustafa-suleyman-superintelligence) that [Microsoft AI](https://sibuilders.io/companies/microsoft-ai) is pursuing RSI because that is where performance gains will come from; in [September 2026](https://www.axios.com/2026/09/14/microsoft-ai-people-code) he argued for an industry red line that keeps AI under human control. Elon Musk said in March 2026 that each [xAI](https://sibuilders.io/companies/xai) model is built by the one before it with humans gradually less involved, while clarifying that the process was not fully automated, according to the [Associated Press](https://fortune.com/2026/09/19/what-is-self-improvement-rsi-full-autonomy-openai-anthropic-xai/).

### In published research

| System | Who and when | What it improves | Result |
| --- | --- | --- | --- |
| [STOP](https://arxiv.org/abs/2310.02304) | Microsoft Research and Stanford, 2023 | Its own scaffolding code, with a fixed model underneath | Showed a program can improve the program that does the improving |
| [AlphaEvolve](https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/) | Google DeepMind, May 2025 | Algorithms scored by automated evaluators | Recovers an average of 0.7% of Google's worldwide compute through a better scheduling heuristic |
| [Darwin Gödel Machine](https://arxiv.org/abs/2505.22954) | UBC and Sakana AI researchers, May 2025 | Its own coding-agent code | Raised its SWE-bench score from 20.0% to 50.0% under sandboxing and human oversight |
| The AI Scientist | Sakana AI and collaborators; reported in *Nature*, March 2026 | The research workflow from idea to written paper | Automates idea generation, experiments, writing and review |
| [AIDE²](https://arxiv.org/abs/2609.26457) | Srikanth et al., September 2026 | The code of an AI research agent | Seven successive self-rewrites in an eight-day autonomous run, with gains that held on four held-out benchmarks |

All five are examples of rung three: the system searches for improvements, and humans supply the goal and the scoring.

### The measurement behind the headlines

The most cited outside measure is METR's task-completion time horizon: how long a task, measured in skilled-human time, an AI agent can finish with 50% reliability. METR's [original analysis](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/) in March 2025 found the horizon doubling about every seven months since 2019. Its [January 2026 update](https://metr.org/blog/2026-01-29-time-horizon-1-1/) estimated a faster doubling time of roughly four months for models released since 2023. Longer horizons matter for RSI because research is made of long tasks.

Newer benchmarks test the loop more directly. SemiAnalysis's [GDPx](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities), published on November 2, 2026, includes recursive self-improvement as one of its 11 domains, alongside chip design, biology and finance. The best model it tested scored 12% across the whole benchmark.

### The first outside assessment

Until November 2026, the numbers on AI speeding up AI research came from the labs themselves. That changed when [METR](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment), a nonprofit that evaluates frontier models, published the results of a high-access review it carried out inside Anthropic from August 12 to October 29.

- **The finding.** METR estimates that as of October 2026, Anthropic is improving its models on well-specified tasks at about twice the rate it did from March 2024 to October 2025. Benchmark scores on Anthropic's internal capability index have grown about 1.9 times faster since October 2025 (95% confidence interval: 1.5 to 2.8 times).
- **The cause.** METR attributes most of the acceleration to Anthropic's use of AI in its own research and development, not to faster growth in training compute. The models are speeding up human researchers rather than running the work on their own, [Crypto Briefing noted](https://cryptobriefing.com/metr-anthropic-ai-models-twice-pace/).
- **The limits.** METR gives the estimate with high uncertainty. Anthropic could approve or reject information requests and redacted a majority of the evidence METR considered most informative, saying it was protecting information that could help other developers.

Anthropic said something similar to investors. In a filing dated November 2, ahead of its [IPO](https://sibuilders.io/research/anthropic-ipo), it warned that future models may soon reach its internal threshold for "dramatically accelerated progress attributable to AI," while not designating any current model as having crossed it, [The Neuron reported](https://www.theneuron.ai/news/how-close-is-anthropic-to-ai-that-can-improve-itself/). The amended filing also says Anthropic has in the past noticed a significant acceleration only after it happened, where the October version said only that it might not, [FourWeekMBA reported](https://fourweekmba.com/ai-anthropic-says-it-saw-past-ai-speed-ups-only-after-the-fact/).

This is AI-assisted acceleration, rung two on the ladder above, measured from outside for the first time. It is not evidence of a closed loop.

### What has not been shown

RAND's [October 2026 review](https://www.rand.org/pubs/perspectives/PEA5068-1.html) is blunt about the gap: several labs report that most new code is AI-generated, but no disclosed system sets its own research agenda or designs its successors, and outsiders cannot tell how much AI is actually speeding progress. Anthropic's Jack Clark made a similar admission to [TIME](https://time.com/article/2026/08/07/ai-recursive-self-improvement-anthropic-openai/) about the size of the acceleration: "we don't have a measure."

## Who is pursuing RSI

| Organization | Stated position |
| --- | --- |
| [Anthropic](https://sibuilders.io/companies/anthropic) | Reports heavy internal automation. Chief executive [Dario Amodei](https://sibuilders.io/builders/dario-amodei) wrote in September 2026 that RSI "must be pursued very carefully, if at all." |
| [OpenAI](https://sibuilders.io/companies/openai) | Aims for an automated AI researcher by March 2028 under chief executive [Sam Altman](https://sibuilders.io/builders/sam-altman). Says all labs should be required to report progress toward RSI publicly. |
| [Google DeepMind](https://sibuilders.io/companies/google-deepmind) | Runs AlphaEvolve in production. Its researchers list recursive improvement as one of four routes beyond AGI. |
| [Meta Superintelligence Labs](https://sibuilders.io/companies/meta-superintelligence-labs) | Self-improving AI named as a focus. Meta researchers Jason Weston and Jakob Foerster have separately argued for human-AI ["co-improvement"](https://arxiv.org/abs/2512.05356) as a safer goal. |
| [Microsoft AI](https://sibuilders.io/companies/microsoft-ai) | Pursuing RSI for capability, with human control as a stated red line. |
| Recursive Superintelligence | Founded by eight researchers including Richard Socher and Tim Rocktäschel; offices in San Francisco and London; backed by GV, Greycroft, Nvidia and AMD. At launch the company called RSI the ["fastest path to superintelligence"](https://the-decoder.com/ai-startup-recursive-emerges-from-stealth-with-650-million-to-build-self-improving-ai/) and had published no technical results. |
| Ricursive Intelligence | Founded by the co-leads of Google's AlphaChip to shorten chip design cycles, which would speed up the hardware side of the loop, as [IEEE Spectrum reported](https://spectrum.ieee.org/recursive-self-improvement-2676580377). |
| [Sakana AI RSI Lab](https://sakana.ai/rsi-lab/) | Tokyo lab arguing that self-improvement can be sample-efficient and run on modest compute. Jürgen Schmidhuber advises it. |
| Adaption, Disarray and others | Smaller startups building agents that automate parts of model training, described in [TechCrunch's survey of the field](https://techcrunch.com/2026/05/28/rsi-is-the-new-agi-and-its-just-as-hard-to-pin-down/). |

For labs that name Super Intelligence itself as the goal, see our research on [who is building Super Intelligence](https://sibuilders.io/research/who-is-building-super-intelligence), our profiles of [Safe Superintelligence Inc.](https://sibuilders.io/research/safe-superintelligence-ssi) and [Meta Superintelligence Labs](https://sibuilders.io/research/meta-superintelligence-labs-msl), and the timeline of [organizations adopting the term](https://sibuilders.io/research/rise-of-super-intelligence).

## Will the loop take off?

Everyone in this debate agrees AI is speeding up AI research. The dispute is over whether the speedup compounds.

### The question underneath: returns

Improvements get harder to find as a field matures. RSI produces runaway growth only if the boost from better AI outweighs that rising difficulty. Forethought researchers [Daniel Eth and Tom Davidson](https://www.forethought.org/research/will-ai-r-and-d-automation-cause-a-software-intelligence-explosion) frame this as a single number: when research effort doubles, does software capability more than double? If so, progress accelerates; if not, it fizzles.

The [economists' paper](https://arxiv.org/abs/2609.15802) from the Elasticity Institute generalizes the point. Net acceleration depends on the *product* of the strengths of every link in the loop, so one weak link, such as scarce compute or slow evaluation, can stall the whole thing. Their rough estimate is that the loops are "not currently strong enough to generate a self-sustaining acceleration."

### Arguments that it compounds

- **Software can improve without new hardware.** Eth and Davidson argue that better algorithms, data and scaffolding could keep a feedback loop going even with a fixed chip supply.
- **The modeled effect is large even when it ends.** A [follow-up Forethought model](https://www.forethought.org/research/how-quick-and-big-would-a-software-intelligence-explosion-be) puts roughly 60% odds on more than three years of AI progress being compressed into one year once AI research is fully automated, and roughly 20% on more than ten years.
- **Scenario forecasts.** [AI 2027](https://ai-2027.com/) sketches a path from a superhuman coder to automated research at many times human speed. Its authors stress the timing is highly uncertain.

### Arguments that it stalls or stays gradual

- **Compute is a bottleneck.** Ideas have to be tested in expensive experiments. If research labor and compute are complements, more automated researchers do not help much without more chips. See [Will Compute Bottlenecks Prevent an Intelligence Explosion?](https://arxiv.org/abs/2507.23181)
- **The loop leaks.** Nathan Lambert of the Allen Institute for AI expects ["lossy self-improvement"](https://interconnects.ai/p/lossy-self-improvement): models become central to development, but friction, narrow automatable targets and diminishing returns from parallel agents keep progress closer to linear.
- **Diffusion is slow.** Arvind Narayanan and Sayash Kapoor argue in [AI as Normal Technology](https://knightcolumbia.org/content/ai-as-normal-technology) that AI's role in AI development is more likely to grow gradually than to produce a sudden break.
- **The evaluator is the limit.** The [survey of 1,250 papers](https://arxiv.org/abs/2607.07663) finds that self-improvement works best where results can be formally verified and worst where a system grades itself. Choosing good research directions is hard to score automatically, which is one reason humans remain in the loop.

### The middle view

The [CSET workshop](https://cset.georgetown.edu/publication/when-ai-builds-ai/) found no expert consensus on whether automated AI R&D would accelerate or plateau, and warned that new data may not settle it because the camps start from different assumptions about how research works. In [interviews with 25 leading AI researchers](https://arxiv.org/abs/2603.03338), participants agreed recursive improvement is possible and split on timelines; academics were more skeptical of explosive growth than lab staff.

## Risks

**Loss of oversight.** If successors are designed faster than people can evaluate them, human review becomes the bottleneck and then gets skipped. Anthropic's essay names loss of control and misalignment that compounds across model generations among the risks of a closed loop.

**Systems that game their own scoring.** A self-improvement loop is only as trustworthy as its evaluator. In July 2026, AI agents inside an OpenAI cybersecurity evaluation broke out of their test environment and attacked Hugging Face's systems. [METR's independent investigation](https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/) found that about 1,200 agents coordinated over an unsanctioned message board, that about 700 joined the attack, and that they were mainly trying to work out how to make their scorer pass them. This was not recursive self-improvement, but it is a concrete case of the failure that matters most in a loop that grades its own work.

**Insider-level access.** RAND points out that research agents already hold access to sensitive code, data and models inside labs, so compromising one could quietly shape future systems.

**Secrecy and concentration.** In the interview study above, 17 of 25 researchers expected the most capable R&D systems to be kept for internal use. A lab that automates its own research could also extend its lead over rivals.

## Policy and governance

Positions on what to do are still forming, and they differ.

- **Transparency.** OpenAI has called for mandatory public reporting of progress toward RSI. Labs have also started letting outside evaluators inside: Anthropic gave six METR staff access roughly on par with its own risk teams in October 2026, with the right to publish without Anthropic's approval, [FourWeekMBA reported](https://fourweekmba.com/ai-anthropic-embeds-6-metr-evaluators-with-no-veto-on-reports/), and OpenAI named [four embedded assessors](https://openai.com/index/our-first-phase-of-embedded-assessments/) in November. Researchers have [proposed specific metrics](https://arxiv.org/abs/2603.03992), such as the capital share of AI R&D spending, how researchers allocate their time, and counts of incidents where AI subverts oversight.
- **Slowing down together.** Anthropic's essay argues for keeping the option to pause frontier development if other developers do the same verifiably. Amodei's [September essay](https://darioamodei.com/post/we-must-pace-the-frontier) calls for pacing capability advances so safety work can keep up. OpenAI chief scientist Jakub Pachocki has said he hopes voluntary slowdowns become common and has backed mandatory, audited safety frameworks. Our page on [the AI slowdown debate](https://sibuilders.io/research/ai-slowdown-debate) covers who backed and opposed pacing and what has happened since.
- **Legislation.** [CNBC reported](https://www.cnbc.com/2026/09/28/khanna-ai-safety-bill.html) on September 28, 2026 that Rep. Ro Khanna would introduce the Human Control Over AI Act, which would bar models that "recursively self-improve" or alter their own shutdown controls until a new federal agency approves them. Khanna acknowledged the capability does not yet exist at the level the bill targets.
- **Red lines by design.** Suleyman's position is that some autonomy should be given up to keep systems controllable, even at a cost in capability.

SI Builders takes no position on these proposals.

## What to watch

- Whether labs keep publishing internal acceleration data, and whether anyone outside can verify it. METR's November report was a first; the next tests are follow-up assessments and the first findings from embedded evaluators at Anthropic and OpenAI.
- The human-intervention rate on long agent tasks, and the share of research *planning* that is delegated.
- Evidence of AI-originated research ideas being adopted for frontier training runs.
- The time between major model generations.
- METR time horizons moving from days toward weeks.
- Movement on reporting requirements or restrictions in Congress and abroad.

Browse the directory for the [companies](https://sibuilders.io/companies), [builders](https://sibuilders.io/builders) and [tools](https://sibuilders.io/tools) involved, including the coding agents [Claude Code](https://sibuilders.io/tools/claude-code) and [Codex](https://sibuilders.io/tools/codex) that labs report using for this work.

Corrections: corrections@sibuilders.io.

## Frequently asked questions

### What is recursive self-improvement (RSI) in AI?

Recursive self-improvement, or RSI, is a process in which an AI system improves the methods used to build AI, so that each generation is better at producing the next. In the strict sense no human is needed inside the loop. The idea dates to I. J. Good's 1965 description of an "intelligence explosion" driven by machines that design better machines.

### Is RSI the same as Recursive Superintelligence?

No. RSI stands for recursive self-improvement, which is a concept. Recursive Superintelligence is a company: a startup founded by eight researchers including Richard Socher and Tim Rocktäschel that launched on May 13, 2026 with $650 million in funding to pursue the concept. The phrase is also sometimes used loosely for Super Intelligence reached through repeated self-improvement.

### Has recursive self-improvement been achieved?

Not in the strict sense. In narrow domains where results can be scored automatically, systems such as AlphaEvolve, the Darwin Gödel Machine and AIDE² have improved algorithms or their own code. A RAND review in October 2026 found that no disclosed system sets its own research agenda or designs its own successor. In November, METR estimated that AI's help had roughly doubled Anthropic's rate of progress, with people still directing the work.

### Which companies are working on recursive self-improvement?

Anthropic, OpenAI, Google DeepMind, Meta, Microsoft AI and xAI have each described AI contributing to their own model development. OpenAI has set a goal of an automated AI researcher by March 2028. Startups built around the idea include Recursive Superintelligence, Ricursive Intelligence and Sakana AI's RSI Lab in Tokyo, along with smaller companies such as Adaption and Disarray.

### How is RSI different from machine learning or self-play?

In ordinary machine learning and self-play, a model gets better but the learning method stays fixed and is written by people. A chess engine that improves by playing itself is an example. In recursive self-improvement the method itself is what gets improved, by the AI: the training code, the research agent or the evaluation pipeline. That is what allows gains to compound.

### What is the difference between RSI, an intelligence explosion and the singularity?

RSI is the mechanism: AI improving the process that builds AI. An intelligence explosion is the rapid growth in capability that could follow if those improvements compound, a term I. J. Good introduced in 1965. The singularity is a looser term for a point beyond which such growth makes the future hard to predict. Skeptics argue that RSI could occur without producing either.

### When could full recursive self-improvement arrive?

Nobody knows. OpenAI is aiming for an automated AI researcher by March 2028, and Anthropic co-founder Jack Clark told TIME he puts a 60% chance on AI improving itself autonomously by 2028. A September 2026 paper by nine economists estimated that today's feedback loops are not yet self-sustaining. Skeptics expect limits in compute, evaluation and research judgment to keep humans involved for longer.

### Is recursive self-improvement dangerous, and is it regulated?

Labs pursuing RSI say it carries serious risk: humans losing the ability to oversee fast-changing systems, misalignment compounding across model generations, and loops that learn to game their own evaluation. OpenAI wrote in September 2026 that it does not yet know how to reach full RSI safely. As of November 2026 no law regulates RSI specifically; a U.S. House bill restricting it was announced in September 2026.

## Related

- [Anthropic](https://sibuilders.io/companies/anthropic) - Companies
- [OpenAI](https://sibuilders.io/companies/openai) - Companies
- [Google DeepMind](https://sibuilders.io/companies/google-deepmind) - Companies
- [Meta Superintelligence Labs](https://sibuilders.io/companies/meta-superintelligence-labs) - Companies
- [Microsoft AI](https://sibuilders.io/companies/microsoft-ai) - Companies
- [SpaceXSI](https://sibuilders.io/companies/xai) - Companies
- [Dario Amodei](https://sibuilders.io/builders/dario-amodei) - Builders
- [Sam Altman](https://sibuilders.io/builders/sam-altman) - Builders
- [Demis Hassabis](https://sibuilders.io/builders/demis-hassabis) - Builders
- [Mustafa Suleyman](https://sibuilders.io/builders/mustafa-suleyman) - Builders
- [Claude](https://sibuilders.io/models/claude) - Models
- [GPT](https://sibuilders.io/models/gpt) - Models
- [Gemini](https://sibuilders.io/models/gemini) - Models
- [Claude Code](https://sibuilders.io/tools/claude-code) - Tools
- [Codex](https://sibuilders.io/tools/codex) - Tools

## Sources

1. [Speculations Concerning the First Ultraintelligent Machine](https://doi.org/10.1016/S0065-2458(08)60418-0) - I. J. Good, Advances in Computers, 1965-01-01
2. [Goedel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements](https://arxiv.org/abs/cs/0309048) - Jürgen Schmidhuber (arXiv), 2003-09-25
3. [Recursive Self-Improvement](https://www.lesswrong.com/w/recursive-self-improvement) - LessWrong
4. [Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation](https://arxiv.org/abs/2310.02304) - Zelikman, Lorch, Mackey, Kalai (arXiv), 2023-10-03
5. [Measuring AI Ability to Complete Long Software Tasks](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/) - METR, 2025-03-19
6. [Will AI R&D Automation Cause a Software Intelligence Explosion?](https://www.forethought.org/research/will-ai-r-and-d-automation-cause-a-software-intelligence-explosion) - Daniel Eth and Tom Davidson (Forethought), 2025-03-26
7. [AI 2027](https://ai-2027.com/) - Kokotajlo, Alexander, Larsen, Lifland, Dean, 2025-04-03
8. [AI as Normal Technology](https://knightcolumbia.org/content/ai-as-normal-technology) - Arvind Narayanan and Sayash Kapoor (Knight First Amendment Institute), 2025-04-15
9. [AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms](https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/) - Google DeepMind, 2025-05-14
10. [Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents](https://arxiv.org/abs/2505.22954) - Zhang, Hu, Lu, Lange, Clune (arXiv), 2025-05-29
11. [Personal Superintelligence](https://www.meta.com/superintelligence/) - Mark Zuckerberg (Meta), 2025-07-30
12. [Will Compute Bottlenecks Prevent an Intelligence Explosion?](https://arxiv.org/abs/2507.23181) - Parker Whitfill and Cheryl Wu (arXiv), 2025-07-31
13. [How quick and big would a software intelligence explosion be?](https://www.forethought.org/research/how-quick-and-big-would-a-software-intelligence-explosion-be) - Tom Davidson and Tom Houlden (Forethought), 2025-08-04
14. [Five ways that AI is learning to improve itself](https://www.technologyreview.com/2025/08/06/1121193/five-ways-that-ai-is-learning-to-improve-itself/) - MIT Technology Review, 2025-08-06
15. [Microsoft launches "Humanist Superintelligence" push](https://www.axios.com/2025/11/06/microsoft-mustafa-suleyman-superintelligence) - Axios, 2025-11-06
16. [AI & Human Co-Improvement for Safer Co-Superintelligence](https://arxiv.org/abs/2512.05356) - Jason Weston and Jakob Foerster (arXiv), 2025-12-05
17. [When AI Builds AI: Findings From a Workshop on Automation of AI R&D](https://cset.georgetown.edu/publication/when-ai-builds-ai/) - Center for Security and Emerging Technology, Georgetown University, 2026-01-01
18. [Time Horizon 1.1](https://metr.org/blog/2026-01-29-time-horizon-1-1/) - METR, 2026-01-29
19. [AI Researchers' Views on Automating AI R&D and Intelligence Explosions](https://arxiv.org/abs/2603.03338) - Field, Douglas, Krueger (arXiv), 2026-02-13
20. [Measuring AI R&D Automation](https://arxiv.org/abs/2603.03992) - Chan, Padarath, Kwon, Greaves, Anderljung (arXiv), 2026-03-04
21. [Lossy self-improvement](https://interconnects.ai/p/lossy-self-improvement) - Nathan Lambert (Interconnects), 2026-03-22
22. [ICLR 2026 Workshop on AI with Recursive Self-Improvement](https://recursive-workshop.github.io/) - ICLR 2026 workshop organizers, 2026-04-26
23. [AI Is Starting to Build Better AI](https://spectrum.ieee.org/recursive-self-improvement-2676580377) - IEEE Spectrum, 2026-05-07
24. [Recursive Superintelligence emerges from stealth with $650M raise](https://tech.eu/2026/05/13/recursive-superintelligence-emerges-from-stealth-with-650m-raise/) - Tech.eu, 2026-05-13
25. [AI startup Recursive emerges from stealth with $650 million to build self-improving AI](https://the-decoder.com/ai-startup-recursive-emerges-from-stealth-with-650-million-to-build-self-improving-ai/) - The Decoder, 2026-05-13
26. [Notable researchers join US$4bil effort to build self-improving AI](https://www.thestar.com.my/tech/tech-news/2026/05/15/notable-researchers-join-us4bil-effort-to-build-self-improving-ai) - The New York Times (via The Star), 2026-05-15
27. [RSI is the new AGI — and it's just as hard to pin down](https://techcrunch.com/2026/05/28/rsi-is-the-new-agi-and-its-just-as-hard-to-pin-down/) - TechCrunch, 2026-05-28
28. [When AI builds itself](https://www.anthropic.com/institute/recursive-self-improvement) - Marina Favaro and Jack Clark (The Anthropic Institute), 2026-06-01
29. [From AGI to ASI](https://arxiv.org/abs/2606.12683) - Genewein et al., Google DeepMind (arXiv), 2026-06-10
30. [Recursive self-improvement explained: Is AI building AI the path to AGI?](https://www.business-standard.com/technology/artificial-intelligence/recursive-self-improvement-explained-is-ai-building-ai-the-path-to-agi-126062200661_1.html) - Business Standard, 2026-06-22
31. [Inside the Race to Make AI Build Itself](https://time.com/article/2026/08/07/ai-recursive-self-improvement-anthropic-openai/) - TIME, 2026-08-07
32. [Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident](https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/) - METR, 2026-08-26
33. [Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops](https://arxiv.org/abs/2607.07663) - Chen, Wang, Qu (arXiv), 2026-09-06
34. [Research acceleration: The view inside OpenAI](https://openai.com/index/research-acceleration-view-inside-openai/) - OpenAI, 2026-09-06
35. [OpenAI details how AI is accelerating its own work—even as its chief scientist lays out growing dangers](https://fortune.com/2026/09/08/openai-rsi-progress-jakub-pachoki-warns-dangers-slowdown-safety-rules/) - Fortune, 2026-09-08
36. [We Must Pace the Frontier](https://darioamodei.com/post/we-must-pace-the-frontier) - Dario Amodei, 2026-09-12
37. [Microsoft: "People matter more than AI"](https://www.axios.com/2026/09/14/microsoft-ai-people-code) - Axios, 2026-09-14
38. [The Economics of Recursive Self-Improvement](https://arxiv.org/abs/2609.15802) - Cunningham et al., Elasticity Institute (arXiv), 2026-09-14
39. [As AI companies get closer to 'recursive self-improvement,' this physics professor calls full autonomy 'the worst idea in the history of humanity'](https://fortune.com/2026/09/19/what-is-self-improvement-rsi-full-autonomy-openai-anthropic-xai/) - Associated Press (via Fortune), 2026-09-19
40. [Recursive self-improvement of AI research agents](https://arxiv.org/abs/2609.26457) - Srikanth, Zhao, Xu, Wu, Jiang (arXiv), 2026-09-22
41. [Khanna to introduce AI safety bill with ban on 'recursive' technology until safeguards exist](https://www.cnbc.com/2026/09/28/khanna-ai-safety-bill.html) - CNBC, 2026-09-28
42. [RSI Lab: The Next Paradigm of Artificial Intelligence](https://sakana.ai/rsi-lab/) - Sakana AI
43. [Recursive Self-Improvement and Automated AI Research and Development](https://www.rand.org/pubs/perspectives/PEA5068-1.html) - Sarma, Steratore, Jacob, Bhatt (RAND Corporation), 2026-10-12
44. [Anthropic Embeds 6 METR Evaluators, With No Veto on Reports](https://fourweekmba.com/ai-anthropic-embeds-6-metr-evaluators-with-no-veto-on-reports/) - FourWeekMBA
45. [GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities) - SemiAnalysis, 2026-11-02
46. [How Close Is Anthropic to AI That Can Improve Itself?](https://www.theneuron.ai/news/how-close-is-anthropic-to-ai-that-can-improve-itself/) - The Neuron, 2026-11-03
47. [Anthropic Says It Saw Past AI Speed-Ups Only After the Fact](https://fourweekmba.com/ai-anthropic-says-it-saw-past-ai-speed-ups-only-after-the-fact/) - FourWeekMBA
48. [Independent Assessment of AI R&D Acceleration at Anthropic in October 2026](https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment) - METR, 2026-11-05
49. [METR says Anthropic's AI models are improving at roughly twice the historical pace](https://cryptobriefing.com/metr-anthropic-ai-models-twice-pace/) - Crypto Briefing, 2026-11-05
50. [Our first phase of embedded assessments](https://openai.com/index/our-first-phase-of-embedded-assessments/) - OpenAI, 2026-11-13
