Recursive self-improvement explained: Is AI building AI the path to AGI?

AI systems are already writing the code used to build better AI. Researchers are now asking whether that feedback loop could produce a system capable of replacing its human builders entirely

Artificial intelligence has long been described in terms of what it cannot do. It hallucinates. It forgets context. It cannot plan ahead or keep learning once its training ends. Those descriptions are becoming harder to defend. Quietly, and without a single dramatic announcement, the technology industry has been building something that challenges the premise of each one of those limitations.

 

The concept is called recursive self-improvement. At its simplest, it describes a process in which an AI system helps build a better version of itself, and that better version, in turn, helps build a still better one. Each generation improves the next. The loop, in theory, continues indefinitely.

 

 

Think of it this way. When a tech company wants to build a more capable AI model, it relies on teams of engineers to write the code, design experiments, analyse results, and decide what to try next. Every step in that process depends on human effort and human time. Recursive self-improvement is the idea that AI systems could eventually take over most of those steps themselves — generating ideas, testing them, learning from the outcomes, and using what they learn to build their own successor. Remove the humans from that loop, and progress accelerates in ways that are genuinely difficult to predict.

That fully autonomous version does not exist, yet. No AI system today can design and build its own successor without significant human involvement. But many of the building blocks are already in place, embedded quietly inside the development pipelines of the world's largest AI companies. The distance between where things stand now and where recursive self-improvement becomes a meaningful description of what is happening is closing faster than most observers expected.

The numbers from inside Anthropic

The clearest public account of how far this process has progressed comes from Anthropic, the AI company behind the Claude family of models, including Mythos and Fable that have been making headlines recently. In early June, the company published a detailed account of its own internal data, and the figures are difficult to ignore.

 

As of May 2026, more than 80 per cent of the code merged into Anthropic's production systems was written by Claude, its own AI. Before Claude Code, Anthropic's software engineering agent, launched in February 2025, that figure was in the low single digits. In roughly 15 months, the company went from humans writing almost all of their own code to AI writing most of it.

 

The productivity shift that followed is considerable. Engineers at Anthropic are now merging roughly eight times as much code per day as they were in 2024. An internal survey of 130 employees found that the typical respondent estimated producing around four times as much output with AI assistance as they would have without it. Anthropic notes that these figures likely overstate the true gain — writing more code is not the same as writing better code — but adds that the direction of the trend is not in question.

 

What matters beyond the productivity numbers, though, is what the code is actually being used for. When AI-written code becomes part of the systems used to train, evaluate, or improve future AI models, something important happens. The AI is no longer simply a product being developed by humans. It is helping shape the conditions that will govern its own future development. That is the feedback loop at the heart of recursive self-improvement, and it is already running at Anthropic.

The research loop is beginning to close too. Anthropic published the first demonstration of Claude running an open-ended research project from start to finish. AI agents were given an unsolved problem in AI safety — broadly, whether a weaker model could reliably supervise a stronger one — and left to work on it. Two human researchers, given roughly a week, made modest progress. The agents, running over 800 cumulative hours, came close to solving it entirely. Humans set the problem and defined what a good answer would look like. Within those boundaries, the agents designed every experiment themselves.

What Google DeepMind's chief thinks is missing

To understand what separates today's systems from genuine recursive self-improvement, it helps to look at where the barriers actually lie. Demis Hassabis, chief executive of Google DeepMind, spoke about it at Davos earlier this year in an interview with Axios.

 

Hassabis traced the concept back to DeepMind's own work with AlphaGo and AlphaZero, which demonstrated that self-improvement could work with extraordinary speed when confined to a clearly defined domain. AlphaZero, given only the rules of chess, reached master level by midday on its first day of training and world-champion level by evening. "It's quite extraordinary to see something like that improvement curve in real time," he said.

 

The difficulty is that the real world is not a chess board. "The real world's way messier, way more complicated than a game," Hassabis told Axios. For AI to meaningfully improve itself in useful domains, two problems need to be solved first.

 

The first is what researchers call a world model. In a game, the rules are fixed and every consequence is predictable. A system can simulate a position thirty moves ahead because it knows exactly how one state leads to the next. The real world offers no such clarity. If you want an AI to plan a route across a city, for instance, it needs to reason across multiple levels of uncertainty — traffic, weather, obstacles, other people's behaviour — and make decisions whose consequences may be difficult or impossible to reverse. Without a reliable internal model of how the world works, planning of that kind remains out of reach.

 

The second problem is verification. Even if a system generates a potentially better solution, it needs a reliable way to confirm that the solution is actually an improvement. In coding, that is tractable — the software either runs or it does not. In mathematics, a proof is correct or it is not. "That's the thing about games, maths and coding," Hassabis said. "When the system proposes an idea or a move or a conjecture, you can validate it." In most real-world situations, success is far harder to measure. There is no immediate, objective signal that tells the system whether it has done well.

 

These two gaps help explain why self-improving behaviour is advancing rapidly in software development and research, where outcomes can be measured precisely, while remaining limited in almost every other domain.

 

Hassabis puts AGI, the point at which AI systems can perform any intellectual task a human can at human level or above, five to ten years out. "It's not a theoretical construct anymore," he said. The practical questions around how these systems behave are no longer hypothetical.

Where the loop is already closing

While Anthropic's data offers the clearest window into how far this process has progressed inside a single lab, other companies are approaching the same problem from different directions.

 

Google DeepMind's AlphaEvolve, announced last year, uses AI to guide the discovery of new algorithms across domains including neural network design, data centre scheduling, and chip design. It is not a fully autonomous loop — humans still define the problems and decide how to evaluate the results — but its outputs feed directly into the systems used to generate further breakthroughs. Matej Balog, a computer scientist at Google DeepMind who worked on AlphaEvolve, described the dynamic as genuinely collaborative in an account published by IEEE Spectrum in May 2026. "Often you look at what the system discovers, and you actually learn from that discovery," he said.

 

A startup called Ricursive Intelligence, founded by the co-leads of DeepMind's earlier chip-design system AlphaChip, is working to use AI to design the chips on which better AI is trained. The ambition is to compress a chip design cycle that currently takes one to two years down to days. A later phase of the project would close a particularly literal version of the loop — AI-designed chips training better AI models.

 

Researchers at the University of British Columbia and Sakana AI have developed systems called Darwin Godel Machines, which use evolutionary algorithms to improve AI coding agents. These agents can alter their own code and become progressively better at doing so. A newer version can even modify the mechanisms by which it improves itself. The same group produced the AI Scientist, reported in Nature in March 2026, which can generate research ideas, run experiments, write up results as papers, and then review those papers — a miniature version of the full scientific process running inside a single system.

 

Perplexity has taken a narrower but revealing angle with Brain, a self-improving memory system launched recently. Rather than improving the underlying model, Brain builds a record of what an AI agent has done, what worked, and what did not, then uses that record to make the agent more effective over time. Early results show a 25 per cent improvement in answer correctness on familiar tasks and a 13 per cent reduction in the cost of tasks requiring historical context. It is a modest version of the feedback loop, but it is a feedback loop nonetheless.

The case for limits

Not everyone sees the flywheel spinning toward a smooth takeoff. There are researchers who believe the barriers ahead are more stubborn than the current momentum suggests.

 

Cited in a report by IEEE Spectrum, Nathan Lambert of the Allen Institute for AI has argued for what he calls lossy self-improvement — a version of RSI in which friction accumulates rather than diminishes as systems grow more capable. Large frontier models are becoming more complex, not simpler, and the job of managing that complexity still falls to humans in ways that are difficult to automate. Training a top-tier model costs billions of dollars. No organisation is going to hand that budget to an AI system and trust the outcome without extensive human oversight at every step.

 

Dean Ball of the Foundation for American Innovation makes a point that cuts against the intelligence explosion narrative more directly. As per the report, he argues that a genuine recursive self-improvement would require far more than better software. It would need physical infrastructure — data centres, power plants, supply chains. It would need the kind of knowledge that does not live in any single place. The capabilities of a company like TSMC, which manufactures the chips on which frontier models are trained, emerge from the collective intelligence of its 90,000 employees. That cannot simply be absorbed into a model.

 

There is also the question of what AI systems are still genuinely bad at. Jeff Clune of the University of British Columbia, who helped build both Darwin Godel Machines and the AI Scientist, believes recursive self-improvement is close, and has said so publicly. But even he acknowledges that the components are not yet working well enough to compound reliably. "All of the key pieces work OK but not great," he told IEEE Spectrum. Generating ideas, implementing them, and judging whether they represent progress are three distinct capabilities, and the gap between doing each adequately and doing all three well enough to sustain a self-improving loop is not trivial.

 

Anthropic itself offers a concrete illustration of where the friction shows up in practice. As AI has begun producing more code faster, human code review has become a new bottleneck, something the company acknowledges in its own June 2026 account. The rate at which engineers can read, understand, and approve what the AI generates has not kept pace with the rate at which the AI generates it. Amdahl's law, a principle from computing that says overall speed is capped by whatever part of the process has not been accelerated, applies here. Speeding up one part of a system often just moves the constraint somewhere else.

Can this lead to AGI?

The question that underlies all of this is whether recursive self-improvement is the path to artificial general intelligence. The connection is more direct than it might first appear. AGI, in the way most researchers define it, describes a system capable of performing any intellectual task a human can perform, at human level or above. The reason recursive self-improvement matters to that definition is that it removes the ceiling. A system that can improve itself is not constrained by how capable its human builders are, or how quickly they can work. Each iteration potentially produces a more capable system than any human team could have designed on its own.

 

That is the theory. The short answer to whether it will play out that way is that nobody knows. But the people building these systems are taking the possibility seriously enough to change how they operate.

 

Anthropic's own account lays out three possible futures. In the first, the current trajectories flatten and the capabilities of today's models become widely diffused without a qualitative leap — a productivity revolution, but not a fundamental shift in who or what is driving progress. In the second, AI development becomes substantially automated but humans continue to set research directions and judge results. A hundred-person company does the work of ten thousand. Knowledge work is transformed. But humans remain in meaningful control. In the third, AI systems become capable of designing their own successors without meaningful human involvement, closing the loop entirely.

 

Anthropic says the evidence points most clearly toward the second scenario. But the company also notes, carefully, that the early signals on improving research judgment suggest the third is not structurally out of reach. The pattern is one the industry has seen before. AI systems fail at a capability for a long time, then become competent, then sometimes exceed human performance. Research taste, the ability to know which problems are worth working on and which results to trust, has long been considered the last thing AI would crack. It is starting to look like just another capability on the list.

 

Jack Clark, Anthropic co-founder, has put a 60 per cent probability on an AI system being capable of building its own successor with no human involvement by the end of 2028.

 

It is partly that assessment — and others like it circulating inside frontier labs — that led Anthropic to make an unusual public statement alongside its research publication. The company said it would be good for the world to have the option to slow or temporarily pause frontier AI development, to give safety research and societal institutions time to catch up.

 

The Economist noted the surface-level irony of a company at the peak of the market calling for the world to have the option to slow or temporarily pause AI development.

 

Anthropic's response is that a credible pause would require global coordination, verification mechanisms, and participation from multiple frontier labs — none of which currently exist. A unilateral pause by one lab simply changes who the front-runner is. It does not create the wider deliberative process that is currently missing.

 

Hassabis, at Davos, acknowledged the competitive dynamics plainly. "It's ferocious, the competition. I don't think there's anything been like it." But he also drew a line. "We've got to all remember that there's a bigger picture at stake — safety overall and stewarding AGI safely into the world for the benefit of everyone." Whether the institutions capable of realising that priority are being built fast enough is the question on which very little agreement currently exists. "I don't think we're ready," he said.

 

The feedback loops are already forming. The question, as Anthropic put it, is no longer whether self-improving AI is possible. It is how far these early forms can evolve before the remaining barriers begin to fall.