Short answer: recursive self-improvement (RSI) is a feedback loop where AI makes the next AI better. A model writes training code, runs experiments, designs kernels or generates data; the improved model does that work faster; and the loop tightens. In November 2026 the loop is real but partial. Frontier labs report that AI agents now do a large share of their research engineering, and an independent assessment estimates one lab's progress at about twice its earlier rate, mostly because of AI. People still choose the experiments, approve the changes and hold the compute budget. Whether the loop closes in 2027 is an open question, and this page lays out the evidence on both sides, with every number linked to its source.
We will update this page as the signals move. The update log is at the bottom.
Definition
Recursive self-improvement is any process where an AI system improves the process that produces AI systems, and the output of that improvement feeds back in. Three parts make it recursive rather than just useful:
- The AI works on AI. Code, data, chips, training recipes or research ideas for the next model.
- The improvement feeds back. The better model is put back to work on the same problem.
- The cycle gets shorter or better each time. If it does not, you have a productivity tool, not a loop.
RSI is a spectrum. At one end, AI coding assistants make human researchers somewhat faster. At the other, AI systems run AI research end to end, and humans set goals and limits. Most of the debate is about where on that spectrum we are, and how fast we are moving along it.
RSI is the mechanism. An intelligence explosion is one possible outcome: RSI where each round of improvement buys more than the last, so capability grows faster and faster. RSI with diminishing returns gives steady progress, not an explosion.
History
The idea is sixty years old. The working systems are recent.
- 1965, I. J. Good. In Speculations Concerning the First Ultraintelligent Machine, Good argued that a machine better than humans at all intellectual work could also design better machines, causing an "intelligence explosion".1
- 2000s, Eliezer Yudkowsky. Proposed "seed AI", an AI built to understand and improve its own design, and in 2013 wrote Intelligence Explosion Microeconomics, which framed the question as returns on reinvested cognitive work.2
- 2003, Jürgen Schmidhuber. Described the Gödel machine, a self-referential program that rewrites any part of its own code once it can prove the rewrite helps.3 Elegant, and impractical: the proofs are rarely tractable.
- 2014, Nick Bostrom. Superintelligence brought takeoff speed and the control problem into mainstream debate.4
- 2024, The AI Scientist. Sakana AI and collaborators built a pipeline that generates ideas, runs experiments and writes papers, at under $15 per paper. The papers were modest; the pipeline was the point.5
- 2025, AlphaEvolve. Google DeepMind's Gemini-powered coding agent found a scheduling heuristic that recovers 0.7% of Google's fleet-wide compute, sped up a kernel in Gemini's architecture by 23% for a 1% cut in Gemini training time, and found a way to multiply 4x4 complex matrices with 48 scalar multiplications.6 That is AI improving the training of the AI that powers it, in production.
- 2025, Darwin Gödel Machine. Researchers at the University of British Columbia and Sakana AI dropped Schmidhuber's proof requirement and replaced it with empirical testing. A coding agent that edits its own code improved from 20.0% to 50.0% on SWE-bench and from 14.2% to 30.7% on Polyglot.7
- 2026, lab-scale AI R&D. OpenAI said it reached its "automated research intern" goal in September,8 Anthropic published measurements of how much of its R&D Claude leads,9 and METR published an independent assessment of AI-driven acceleration at Anthropic.10
Mechanisms
There is no single loop. There are several, running at different speeds, and each has its own bottleneck.
| Loop | Example | What limits it |
|---|---|---|
| Code and kernels | AlphaEvolve's Gemini kernel speedup | Gains per kernel are small and one-off |
| Automated research | The AI Scientist; lab research agents | Compute for experiments, judging results |
| Data generation | Synthetic data checked by tests or proofs | Only works where answers can be verified |
| Hardware and scheduling | AlphaEvolve's Borg heuristic | Fab cycles take years, not weeks |
| Training recipes | Self-editing agents like the Darwin Gödel Machine | Each candidate costs a training run |
The fast loops are software: code, recipes, experiments. The slow loop is hardware. That split matters, because the most aggressive scenarios depend on software gains alone carrying progress for a while, without waiting for more chips.
The model below is a toy. It shows how a per-cycle gain compounds, and how each bottleneck damps it. The damping factors are made up for teaching. Play with it, then read the evidence.
Evidence and signals
What we can actually measure, as of November 2026:
Agents can work on longer tasks. METR measures the length of task, timed by a skilled human, that an agent completes half the time. In March 2025 that horizon had doubled about every 7 months for six years.11 Its January 2026 update, on a larger task suite, put the doubling time at 130.8 days for models since 2023 and 88.6 days since 2024.12 Longer horizons matter for RSI because research is a long task.
Labs report heavy internal use. OpenAI says its research organisation uses 3.1 agent-workdays of effort for every human workday, and that it reached its goal of an automated research intern, defined as a system that carries out well-defined research tasks under human direction.8 It also says over half of successful 4 to 8 hour tasks needed at least one human intervention. Anthropic reports that Claude leads 26% of its AI R&D work, up from under 1% in February 2026, with no measured subset fully autonomous.9 Both are self-reported.
An outside measurement. METR's November 2026 assessment, with access inside Anthropic, concluded the company is likely improving models on well-specified tasks at around twice its March 2024 to October 2025 rate, would not be surprised by anything from 1.3x to 3x, and attributes perhaps 80% of the acceleration to AI. It saw no evidence of significant bottlenecks in algorithmic progress or training compute over the next 12 months, while saying its view on bottlenecks had changed several times. Much of the evidence it found most useful was redacted by Anthropic.10
Read those three together and the honest summary is: AI is now a large input to AI research at the leading labs, and progress there has sped up. What is not yet shown is a loop that runs without people, or one that keeps speeding up.
Timelines and forecasts
Every row below links to its source. Lab statements are targets or self-reports. Scenarios are stories built to be stress-tested, not predictions.
| Source | When | Claim |
|---|---|---|
| Epoch AI | May 2024 | Returns to software R&D of about 0.83 for computer chess; above 1 elsewhere but not significant. No strong evidence for a software singularity |
| Anthropic | Mar 2025 | Expects powerful AI systems in late 2026 or early 2027 |
| METR | Mar 2025 | Task time horizon doubling about every 7 months over 6 years |
| AI 2027 | Apr 2025 | Scenario: superhuman coder in March 2027, AI R&D progress multiplier of 4x then, 10x by June 2027. Authors later said their medians were longer than 2027 |
| Forethought | Aug 2025 | Once AI automates AI R&D, about 60% chance of compressing more than 3 years of progress into under a year, about 20% for more than 10 years |
| OpenAI, via TechCrunch | Oct 2025 | Intern-level research assistant by September 2026, legitimate AI researcher by March 2028 |
| METR | Jan 2026 | Time horizon doubling in 88.6 days for models since 2024 |
| OpenAI | Sep 2026 | Says it reached the research intern goal; still targets an automated AI researcher by March 2028 |
| Anthropic | Sep 2026 | Claude leads 26% of its AI R&D work |
| METR | Nov 2026 | Anthropic progress at about 2x its earlier rate, 1.3x to 3x plausible, about 80% due to AI |
The same entries, filterable by source type:
Why 2027 keeps coming up: the AI 2027 scenario put its superhuman coder there, Anthropic's 2025 expectation for powerful AI ends there, and if METR's measured acceleration holds or grows, a lab doing most of its research with AI in 2027 is a straight-line extrapolation. Straight lines are where forecasting starts, not where it ends. The AI 2027 authors themselves moved their medians later after publishing.13
Counterarguments
The case against a fast loop is strong, and it is mostly about bottlenecks, not about whether AI is useful.
- Diminishing returns. Each improvement may be harder to find than the last. Epoch AI's estimate for computer chess put returns to research effort just below 1, the level where progress slows over time even with more effort.14 Forethought's model puts the median returns above 1 with a wide range.15 The sign of this one number decides between steady progress and acceleration, and nobody has measured it well for frontier AI.
- Compute. Faster idea generation does not mean more training runs. Experiments need GPUs, and GPUs come from fabs and power plants on multi-year schedules. A software-only loop runs into this ceiling.
- Data. Synthetic data works where answers can be checked: code with tests, math with proofs. For open-ended skills there is no cheap judge, so there is no cheap data.
- Amdahl's law. If AI speeds up 80% of research tenfold and the remaining 20% (deciding what to try, reviewing results, waiting for training) stays human-paced, the whole process speeds up about 3.6x, not 10x. The slowest serial step sets the pace.
- Verification. Knowing whether a change is actually better takes evals, and evals for frontier capability are slow, noisy and saturating. METR notes the model it assessed scored above 50% on all 15 benchmarks it used, which makes measurement harder.10 A loop cannot run faster than you can tell good from bad.
- Measurement and incentives. Most of the 2026 evidence is self-reported by labs that benefit from looking fast, or reviewed under redaction. Self-reported uplift polls inside Anthropic reached about 12x, which METR calls "only very loose evidence".10
None of these says the loop cannot happen. Together they say it will probably be lumpier and slower than a simple compounding curve, which is what the toggles in the model above are for.
Risks and governance
The safety concern is not speed alone. It is that a fast loop leaves less time for people to notice and fix problems, and that each generation of AI is partly built by the previous one, so flaws can carry forward.
The main frontier labs now treat AI R&D capability as a tracked risk:
- Anthropic's Responsible Scaling Policy has AI R&D capability thresholds, originally framed as fully automating entry-level AI research work and causing dramatic acceleration in the rate of effective scaling. A 2026 revision clarified that the acceleration threshold means doubling the rate of progress in aggregate AI capabilities, not doubling researcher productivity.16
- OpenAI's Preparedness Framework lists AI self-improvement as one of three tracked categories, next to biological and chemical, and cybersecurity.17
- Independent review. METR's assessment of Anthropic is an early example of an outside group measuring AI-driven acceleration inside a lab, and also of the limits: redactions meant the public had to rely on METR's judgment.10
Open questions in policy are practical ones: how to define RSI precisely enough to regulate, how to verify lab claims without leaking capabilities, and who gets the unredacted evidence. We do not expect those to be settled in 2027.
What it means for enterprises and engineering teams
You do not need a view on 2027 to act on this. Whatever the timeline, the pattern teams will feel first is faster model turnover: more releases, bigger jumps between them, and shifting prices. That favours teams built to absorb change.
- Keep an eval harness from your own cases. When a model ships, you want to know in a day whether it is better on your work. We describe the method in how we test AI models on your work. A harness is also the defence against benchmark claims that do not hold on your data.
- Keep the model swappable. A thin layer between your workflow and the model lets you move when a better or cheaper model ships, and limits vendor risk if a provider changes terms, pricing or availability. Your advantage lives above that layer, in your method, which we argue in your AI moat is your method, not the model.
- Keep a person on decisions that matter. The labs building the loop still keep people on approvals. So should you. The verification bottleneck is real in your systems too, and a reviewed decision is cheaper than an unreviewed mistake.
- Look inside where you can. For high-stakes workflows, an open-weight model you can monitor internally can be worth more than a slightly better closed one. Our mechanistic interpretability deep dive covers what that buys you today.
- Ground answers in your own data. Better models still need the right facts. A retrieval layer with permissions, such as the one in our Graph RAG build guide, keeps that work when you swap the model underneath.
- Use agents in your own engineering, with gates. The loop at the labs is agents plus checks. We wrote about where we are on AI-native engineering and why verification is the work.
If you want help putting this into practice, Skcript Applied Labs builds production AI systems with these guardrails, and our AI consulting practice and AI pilot program are short, fixed-scope ways to start. The use cases show the kinds of systems we build.
Glossary
- Recursive self-improvement (RSI): A feedback loop in which an AI system improves the process that builds AI, so each improved system makes the next improvement faster or better.
- Intelligence explosion: I. J. Good's 1965 term for runaway growth in machine intelligence once machines can design better machines. A possible outcome of RSI if returns stay high.
- Seed AI: Eliezer Yudkowsky's term for an early AI designed to understand and rewrite its own code, as the starting point of recursive self-improvement.
- Gödel machine: Jürgen Schmidhuber's 2003 design for a self-referential program that rewrites any part of its own code once it can prove the rewrite is useful.
- AI R&D automation: AI systems doing the work of AI research and engineering: writing training code, running experiments, analysing results and proposing changes.
- AI R&D progress multiplier: How much faster AI research moves with AI help than without it. A multiplier of 2 means one year of progress in six months.
- Task time horizon: METR's measure of AI capability: the length of task, timed by how long it takes a skilled human, that an AI agent completes with a given success rate, usually 50%.
- Software intelligence explosion: A scenario where AI improves AI through better algorithms and training methods alone, without needing more hardware.
- Returns to software R&D: How much algorithmic efficiency each doubling of research effort buys. Above 1, AI-driven research could accelerate; below 1, it slows over time.
- Superhuman coder: From the AI 2027 scenario: an AI system that can do the job of the best human coder on AI research engineering, faster and cheaper.
What we are watching
These are the measurable signals that would move our view, in either direction:
- Time horizon doubling time. METR's next updates. Faster than 88.6 days supports acceleration; slower weakens it.
- Independent acceleration estimates. Whether METR or others repeat the November 2026 assessment at other labs, and whether the roughly 2x figure rises, holds or falls.
- Share of R&D led by AI. Anthropic's 26% figure and OpenAI's agent-workday ratio, and whether either reports work that is fully autonomous.
- Human interventions per task. OpenAI's share of long tasks that needed a person. Falling interventions mean a looser loop.
- OpenAI's March 2028 target. Whether the automated AI researcher arrives, and how it is defined when it does.
- RSP and Preparedness threshold calls. Whether any lab states that a model crossed an AI R&D threshold.
- Compute and power build-out. Whether hardware keeps pace, or becomes the binding constraint the counterarguments predict.
- Benchmark saturation. Whether evaluators find new tests that can still measure frontier progress.
Further reading: Zvi Mowshowitz's commentary on the METR assessment, "AI R&D Is Accelerating Towards Potential RSI in 2027", collects reactions from researchers and policymakers. lesswrong.com
Update log
Last updated: 19 November 2026.
- 19 November 2026: First version. Covers evidence through METR's 5 November 2026 assessment of Anthropic.
References
Footnotes
-
I. J. Good, "Speculations Concerning the First Ultraintelligent Machine", Advances in Computers, vol. 6, 1965. doi.org ↩
-
Eliezer Yudkowsky, "Intelligence Explosion Microeconomics", MIRI, 2013. intelligence.org ↩
-
Jürgen Schmidhuber, "Gödel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements", 2003. arxiv.org ↩
-
Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, Oxford University Press, 2014. global.oup.com ↩
-
Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, David Ha, "The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery", 2024. arxiv.org ↩
-
Google DeepMind, "AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms", May 2025. deepmind.google ↩
-
Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, Jeff Clune, "Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents", 2025. arxiv.org ↩
-
OpenAI, "Research acceleration: The view inside OpenAI", September 2026. openai.com ↩ ↩2
-
Anthropic, "Measurements for understanding the pace of AI development inside frontier labs", 2026. anthropic.com ↩ ↩2
-
METR, "Independent Assessment of AI R&D Acceleration at Anthropic in October 2026", November 2026. metr.org ↩ ↩2 ↩3 ↩4 ↩5
-
METR, "Measuring AI Ability to Complete Long Tasks", March 2025. metr.org ↩
-
Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean, "AI 2027", April 2025, with later timeline notes. ai-2027.com ↩
-
Tamay Besiroglu, Ege Erdil, Anson Ho, "Do the Returns to Software R&D Point Towards a Singularity?", Epoch AI, May 2024. epoch.ai ↩
-
Tom Davidson, Tom Houlden, "How quick and big would a software intelligence explosion be?", Forethought, August 2025. forethought.org ↩
-
Anthropic, Responsible Scaling Policy and its update history. anthropic.com ↩
-
OpenAI, "Our updated Preparedness Framework", April 2025. openai.com ↩

