Recursive self-improvement: how AI is starting to build AI

Recursive self-improvement is the idea that an AI system could help design a better version of itself, which then designs an even better version, and so on. It is the mechanism most researchers point to when they talk about a fast path from today’s AI to superintelligence. In 2026 it moved from thought experiment to something labs measure in their own engineering data.

What recursive self-improvement means

The idea is old. In the 1960s the mathematician I. J. Good described an “intelligence explosion”: a machine smart enough to improve machine design would trigger a feedback loop that leaves human intelligence far behind. The logic is simple. Human AI research is limited by the number of researchers and how fast they work. If AI can do that research, you can run thousands of copies around the clock, and each improvement speeds up the next one.

In practice, it is useful to split it into two things:

• Bounded self-improvement: a model refines its own answers, generates training data, or tunes its tools. This is already standard practice and it converges; it does not run away.
• Open-ended recursive self-improvement: AI systems choose research directions, run the experiments and train their own successors, with humans mostly watching. This has not happened yet.

How far AI labs have actually got

The most detailed public account comes from Anthropic’s essay “When AI builds itself”, which describes a steady handover of work from humans to models:

• 2021–2023: humans write the code.
• 2023–2025: chatbots suggest snippets that humans copy in.
• 2025–2026: coding agents write and edit whole files.
• Now: autonomous agents run code and hand hours of work to other agents.
• Next, maybe: agents that build and train new models themselves.

Some of the numbers behind that picture:

• More than 80% of the code merged at Anthropic by May 2026 was written by Claude.
• On an internal test where the model has to make training code run faster, speedups went from about 3x in May 2025 to about 52x by April 2026. Anthropic says a skilled human needs four to eight hours to reach roughly 4x.
• In one open-ended AI safety research project, agents recovered 97% of a performance gap that two human researchers closed only 23% of in a week, though humans still chose the problem and the scoring.
• A later report put AI’s share of Anthropic’s research and development at 26% by August 2026, up from 1% in March.

Academic work points the same way. A September 2026 paper on recursive self-improvement of AI research agents showed a research agent rewriting its own code to become more efficient, with gains that carried over to tasks it had never seen. The field is exploding: one survey counted roughly 500 papers on the topic in the second quarter of 2026 alone.

Where recursive self-improvement is stuck

Research taste

The biggest gap is judgement. Models are now very good at executing a well-defined experiment, and weaker at deciding which experiment is worth running. A Princeton-led study found that agents can solve the engineering problems of AI research but do not yet produce original work at top-conference level. Until that changes, humans stay in charge of direction.

Verification

Self-improvement works best where results can be checked automatically: faster code, passing tests, formal proofs. It works worst where “better” is subjective. Loops that grade themselves tend to reinforce their own mistakes.

Compute, energy and human review

Every experiment still needs chips and electricity. And as AI writes more code, human review becomes the new bottleneck, which is exactly what Anthropic says it ran into. Speed up one part of a system and the slowest remaining part sets the pace.

Three possible futures

1. The curve flattens. Progress hits limits in data, compute or research judgement. Even then, spreading today’s capabilities through the economy changes a lot of work.
2. Compounding acceleration with humans steering. AI does most of the doing, humans set direction. Small teams do the work of much larger organisations. Anthropic considers this the most likely near-term path.
3. Full recursive self-improvement. AI builds its own successors, and progress is limited mainly by compute. This is where the biggest upside and the biggest risks sit, and where the intelligence explosion debate lives.

The intelligence explosion debate

On 28 September 2026, research leaders from OpenAI, Anthropic, Microsoft and Meta published a paper warning that automating AI research could compress years of progress into months, and asked policymakers to set up oversight of self-improving systems before it happens. Anthropic has said it would slow down or pause if other frontier developers did the same in a verifiable way, and that building that verification is hard: training runs are easier to hide than missile silos.

Sceptics argue labs have an incentive to make their technology sound both unstoppable and dangerous, and that the bottlenecks above will slow things more than the forecasts assume. Supporters of stronger oversight argue that if the loop does close, there may be very little time to react. Both views are held by serious people.

What this means if you build software

I use AI agents every day to build client systems, and the pattern at the labs matches what I see at small scale: execution is getting cheap, judgement is not.

• Your value moves up the stack. Knowing what to build, for whom and why matters more than typing speed.
• Verification is the job. Tests, evals and monitoring are what let you trust work you did not write. The same applies to evaluating AI systems you ship to users.
• Design agents with checkpoints. Let them work long and autonomously on checkable tasks, and require a human decision on anything irreversible. More in agentic AI for business.

Frequently asked questions

Is AI already improving itself?

Partly. AI now writes most of the code at some frontier labs and runs experiments faster than humans, but people still decide research direction. Fully autonomous recursive self-improvement has not been demonstrated.

What is an intelligence explosion?

A scenario where AI that can improve AI triggers a rapid feedback loop, so capability grows much faster than it could with human researchers alone.

Could recursive self-improvement lead to superintelligence?

That is the main argument for why superintelligence could arrive quickly after human-level AI. Whether it happens depends on research judgement, verification and physical limits like compute and energy.

If you want AI agents doing real work inside your business, with the checkpoints to trust them, see my AI agents and RAG systems service or get in touch.