Recursive Self-Improvement Is No Longer a Thought Experiment—It's Already Partially Here
Recursive Self-Improvement Is No Longer a Thought Experiment—It's Already Partially Here
What Anthropic's June 2026 report, the TIME investigation, and a decade of RSI theory actually tell us about where we are.
By AIPress Editorial · September 18, 2026
In June 2026, Anthropic published a report that did something unusual: it named a thing the AI industry had been circling for decades, measured how far along it was, and admitted the answer was uncomfortably concrete. The report was titled "When AI Builds Itself," and the thing it named was recursive self-improvement (RSI)—the hypothesized point at which an AI system begins to accelerate its own development, closing the loop between building and being built.
Anthropic's own framing was careful. "We are not there yet, and recursive self-improvement is not inevitable," the report said. But the data it presented—from inside the company and from public benchmarks—suggested the loop was already partially closed. By mid-2026, more than 80% of the code Anthropic merged into its codebase was authored by Claude. Engineers were shipping 8× as much code per quarter as they had from 2021–2025. Claude Mythos Preview was achieving roughly 52× speedups on optimized code-tuning tasks where a skilled human researcher needed four to eight hours to reach 4×.
This is not the intelligence explosion. It is not an AI redesigning its own weights, architecture, and training pipeline without human involvement. But it is something the field has been tracking as a leading indicator for a long time—and the indicators are moving.
What RSI actually is
Recursive self-improvement is not a single technical milestone. It is a spectrum, and part of the confusion around it comes from people using the same term to mean different points on that spectrum.
The foundational idea comes from British mathematician Irving John Good, who worked with Alan Turing at Bletchley Park and wrote in the mid-1960s about an "intelligence explosion" that would follow when a machine could design even better machines without human assistance. In the 2000s, Eliezer Yudkowsky—founder of the Machine Intelligence Research Institute and the LessWrong community—developed the concept of a "seed AI": a system designed for self-understanding, self-modification, and recursive self-enhancement. Yudkowsky's 2008 essay "Recursive Self-Improvement" gave the concept its name and its most cited definition: an AI that can redesign its own cognitive algorithms, and then use that higher intelligence to make further improvements, in a cascade.
Nick Bostrom's 2014 book Superintelligence folded RSI into the dominant framework for thinking about AI risk: an early version of an AI could design an improved version of itself, and that process of recursive self-improvement might continue long enough to result in an intelligence explosion and then radical superintelligence.
The academic definition has held up surprisingly well. A June 2026 survey paper, "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement" (arXiv:2609.11873), catalogs the explosion of recent work on self-evolving agents—STOP, Darwin Godel Machine, Hyperagents, Agentic Neural Architecture Search, S3Gym, Aspire—and frames RSI as a genuine research program with measurable milestones rather than pure speculation.
The three stages, as they stand in 2026
Anthropic's report and the TIME investigation that followed it (Harry Booth, August 7, 2026) describe something close to a three-stage picture.
Stage 1—AI as a tool. For most of AI's history, and still for many teams, AI is something humans use to accelerate discrete parts of the development cycle: generating short code snippets, suggesting edits, summarizing documentation. The human is in the loop for everything.
Stage 2—AI as an agent. By 2025–2026, models can write and edit code on their own, sometimes entire files, run that code, delegate hours of work to other agents, and iterate on results. This is where Anthropic is now: Claude writes the majority of its own company's production code, runs experiments, and catches bugs in human-written code. On open-ended investigative tasks, Claude's success rate reached 76% in May 2026, up 50 percentage points in six months.
Stage 3—AI as its own developer. This is the full RSI scenario: an AI system capable of designing and training its own successor. Anthropic's report does not claim this is happening. It says the trends "point to" this possibility and that it "could come sooner than most institutions are prepared for."
The gap between Stage 2 and Stage 3 is the interesting part, because it is where most of the debate lives.
What is actually happening inside Anthropic
The most concrete data in the Anthropic report is internal and previously unreported. As of May 2026, more than 80% of the code merged into Anthropic's codebase was authored by Claude. This is not a productivity statistic in the normal sense—team members are not rewarded for lines of code, and the 8× increase in code per engineer per day is "almost certainly an overstatement of the true productivity gain." What it reflects is that engineers are using AI to write more code, and that code is getting merged.
The qualitative description is striking. One engineer told Anthropic it had been "about 5 months since I last wrote any code myself.". Claude shipped over 800 fixes in April 2026 that reduced a class of API errors by a factor of one thousand. On the code-optimization task—the canonical RSI-adjacent benchmark where Claude is given code that trains a small model and asked to make it run as fast as possible—Claude Mythos Preview hit ~52× speedup in April 2026, up from ~3× for Claude Opus 4 in May 2025. Two human researchers recovered roughly 23% of the gap between the starting code and the theoretical optimum over a week; the agents recovered 97% over 800 cumulative hours.
Anthropic's own assessment of the code quality trajectory is notable: "Claude-written code was somewhat worse than human-written code at Anthropic in late 2025, is roughly at parity today, and we expect it to be strictly better within the year."
The external evidence—and the counter-evidence
The outside-world picture is messier, which is part of why the RSI debate has been so polarized.
On the acceleration side: the length of tasks AI can reliably complete on its own has been doubling roughly every four months, according to METR's time-horizons data that Anthropic cites. SWE-bench, the standard software engineering benchmark, has gone from low single-digit scores to saturation in two years. In May 2026, Claude exceeded the upper limit of a benchmark that had been the standard measure of coding ability.
On the skepticism side: a study from July 2026 (arXiv:2607.27191) had Claude Opus 4.8, running on open-source software called OpenClaw, review two NeurIPS 2026 papers—and found that AI review of AI research is not yet reliably better than human review. The paper covered just two papers, and the original authors knew they were being graded by AI, which could have colored the evaluations. But the result is a useful reminder that RSI is not a monolith: an AI that can optimize its own training code is not necessarily an AI that can evaluate its own research contributions.
The TIME piece captured a moment when the debate was shifting. Jack Clark, Anthropic's co-founder, returned from paternity leave in February 2026 to find colleagues hardly writing code anymore—they managed five or six copies of Claude, which sometimes managed several more. Clark left his policy role to build the Anthropic Institute, a think tank embedded in the company, and put the chances of AI improving itself autonomously by 2028 at 60%. "The world needs options," he told TIME in July, "but we're not saying the world must pause or slow down. That's not what the evidence says."
The "AI 2027" scenario and the forecasting picture
One of the reasons RSI has moved from academic speculation to urgent policy question is a document called "AI 2027," co-authored in spring 2025 by Daniel Kokotajlo, a former OpenAI staffer. It lays out a quarter-by-quarter scenario of how unchecked recursive self-improvement might unfold, ending with superintelligence. Kokotajlo told TIME that people—including some working at AI companies—have reached out privately to say his timeline is too conservative.
The forecast is not a prediction from a major lab. It is a scenario from an individual researcher. But the fact that it is being discussed seriously in the industry—and that Anthropic's own data is broadly consistent with the early quarters of the scenario—is what gives it weight.
Why RSI is hard to measure, and why that matters
The core challenge in evaluating RSI claims is that the thing people are worried about—full autonomous self-improvement leading to an intelligence explosion—is by definition hard to observe before it happens. The precursors are measurable, but they are noisy.
Anthropic's code-ownership statistic, for example, is a measure of how much AI is writing code at one company. It is not a measure of whether that code is making the company's AI better. The 52× speedup figure is a measure of how well Claude can optimize a clearly defined experiment—not a measure of whether Claude can design a better experiment, or a better model architecture, or a better training pipeline.
The METR time-horizons data shows models completing longer tasks. But a model that can complete a 12-hour software task is not the same as a model that can recursively improve itself.
This is why Anthropic's report is more careful than the headlines it generated. The report is documenting a trend—the delegation of AI development to AI systems—and arguing that the trend, if it continues, points toward RSI. It is not claiming RSI has arrived.
The safety dimension
The reason RSI has been a primary concern in AI safety for two decades is not that it is certain to happen, but that if it does, the timeline compresses in ways that challenge existing governance.
If an AI system can design and train its own successor, the usual cycle of human review, safety evaluation, and deployment becomes harder to maintain. Anthropic's report frames this directly: "If systems are capable of fully building their own successors, the ways we secure them, monitor them, and shape their behavior all grow much more important."
Anthropic has been unusually willing to argue for the possibility of a pause. In June 2026, the Future of Life Institute published a statement headlined "Anthropic warns of AI self-improvement risks, considers a pause," quoting the report's warning that "we are approaching a runaway to superintelligence that could threaten our shared human future." Clark's 60% estimate for autonomous RSI by 2028 is high enough to matter, and low enough that it is not a certainty—which is exactly the shape of a policy question rather than a forecast.
The counter-view is straightforward: RSI may turn out to be a lot harder than the precursors suggest, and the gap between "AI writes most of its company's code" and "AI designs its own successor" may be larger than it looks from the inside.
Where this leaves the debate
The honest read on RSI in September 2026 is that the field has moved from pure speculation to documented acceleration, but not from acceleration to explosion.
Anthropic's data is real and specific: 80%+ code authorship, 8× productivity, 52× optimization speedup, 76% success on open-ended tasks. The METR time-horizons trend is real and steep. The academic literature on self-evolving agents has proliferated in 2026, with multiple papers proposing measurable RSI benchmarks.
But none of those things is full recursive self-improvement. None of them shows an AI redesigning its own architecture, training its own successor, or improving itself without human involvement in the loop. The TIME piece captured this precisely: "First, the thinking goes, models make researchers faster, but as each improvement feeds the next, the models take over more of the research cycle. Years of progress compress into months."
That compression is happening. Whether it continues to the point of an intelligence explosion is the open question—and it is an open question in a way that makes the next 12 to 24 months more important than the fact that the loop is partially closed today.
The bottom line
Recursive self-improvement is no longer a thought experiment that might never happen. It is a documented trend with named precursors, measurable benchmarks, and a major lab that has publicly committed to tracking it. Anthropic's report did the field a service by refusing to wait for the full explosion before naming what is already underway.
The relevant questions for September 2026 are narrower and more actionable than the classic RSI debate: How much of AI development is currently delegated to AI, and at what rate is that share growing? What does the trajectory look like if the current rate holds? And what would need to be true for the loop to close fully—and how much time would policymakers and safety teams have if it did?
Anthropic's answer to the last question—60% by 2028—is a starting point for a conversation, not a prediction to build policy around. But it is, at minimum, a serious number from a serious source, and it deserves to be taken seriously by anyone watching the field.
Sources: Anthropic Institute, "When AI Builds Itself" (June 2026); TIME, "Inside the Race to Make AI Build Itself" by Harry Booth (August 7, 2026); Wikipedia, "Recursive self-improvement"; Eliezer Yudkowsky, "Recursive Self-Improvement" (2008) and "General Intelligence and Seed AI" (2001); Nick Bostrom, Superintelligence (2014); Irving John Good, "Speculations Concerning the First Ultraintelligent Machine" (1965); arXiv:2609.11873, "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement" (September 2026); arXiv:2607.27191 (July 2026); METR time-horizons data.
