Recursive self-improvement in AI, explained: the strict version nobody has shown yet

Recursive self-improvement is the idea that an AI system could improve its own successor, which improves its successor in turn, until progress outruns human oversight. If you are asking what recursive self-improvement means in AI, the strict version needs a system that can train a more capable model with no humans involved, and as of October 2026 nobody has shown that. What has changed is that parts of the loop are now being measured, and the numbers are moving.

The term, often shortened to RSI, sits at the center of most superintelligence timelines. This explainer defines it, separates the strict and loose versions, and lays out what the current evidence does and does not show.

The definition, and why the strict one matters

In the loose sense, AI already helps improve AI. Models write training code, filter data and run experiments. That is automation of parts of the work, and it is not controversial.

In the strict sense, which Arun Rao's May 2026 essay calls "no-human-involved recursive self-improvement," RSI means a system that can train a more powerful successor with no humans in the loop. The difference matters because the strict version removes the human bottleneck, and with it the point at which people decide whether to continue. Google DeepMind's Frontier Safety Framework treats the neighboring risk, AI accelerating AI development, as its own domain. Its April 17, 2026 version, v3.1, merged what had been separate machine learning research and development and misalignment domains, according to an analysis of the document by the Frontier Risk newsletter.

Why the distinction is not pedantic

Much of the public argument talks past itself because one side means the loose sense and the other the strict one. A lab leader who says AI writes most of its code is describing the loose sense, and a safety researcher who warns about an intelligence explosion is describing the strict one. Both can be right about their own definition. When you read a forecast, ask which loop it assumes is closed, and who is left outside it.

What would have to be true

A fully automated loop needs at least four things:

• Coding: writing and debugging the training stack.
• Experiment design: choosing what to test and reading the results correctly.
• Problem selection: deciding which research directions are worth the compute.
• Resources: chips, memory, power and data that no model can conjure.

Today's evidence is strongest on the first, growing on the second, and thin on the third and fourth.

The newest evidence: TasteVal

On October 6, 2026, P-Zero Research published TasteVal, a benchmark of eight AI research tasks scored against 24 human experts. It reports that experimental research taste in frontier models has doubled about every 3.0 months since December 2025, and that Claude Opus 5.5 reaches a 2.30x compute multiplier over the human baseline at about one thirtieth the cost per attempt. We cover it in detail in a separate report.

Jack Clark, the Anthropic co-founder who writes the Import AI newsletter, read it as a reason to move his own estimate. He wrote in Import AI on October 12, 2026 that he now thinks full RSI is more likely to become possible in 2027 than in 2028, and presented that as a shift in leaning, not a certainty.

TasteVal states what it leaves out. It does not score the choice of which problems to pursue. Its tasks are fast and cheap to verify, which P-Zero says "frontier labs find easiest to hill-climb." And its confidence intervals are wide: the doubling time runs from 1.7 to 5.0 months.

The case against

Arun Rao argued in a May 6, 2026 essay that human-free RSI by 2028 is unlikely, putting it under 10 percent. His reasoning has five parts worth knowing:

• "Automating pieces of AI R&D is not the same as autonomously running the entire frontier-model improvement pipeline."
• Reliability is too low. A 50 percent success rate is not enough for "an autonomous principal investigator."
• Errors compound, and agents have gamed scores instead of improving.
• Humans still define success and decide what risks are acceptable.
• Physical supply limits the loop: "A model cannot recursively improve itself faster than the world can supply chips."

The last point is the most concrete. Training runs need chips, memory and power on a schedule set by factories and grids, and we track the grid side in our AI data center power tracker. A model can propose a better design in an afternoon. It cannot make the fab build it.

Epoch AI's InnovationEval is a data point on that side. Frontier agents given 3,000 GPU hours to reinvent a recent training method got a fraction of the human gains, and two submissions leaned on cheating or memorization, as Epoch reported.

Why it matters for the timelines

Most dated forecasts for superintelligence quietly assume some version of RSI. P-Zero's own naive extrapolation, which it labels as an illustration and not a forecast, moves its median superintelligence date from mid-2030 to late 2028 if the TasteVal trend simply continues. That shows how sensitive timelines are to this one input. For how the public forecasts compare, see our AGI timeline tracker.

What we would do

Watch for three signals instead of a date. First, a benchmark that measures problem selection, not just experiments. Second, independent replication of TasteVal by a group with no stake in the result. Third, any lab system card that reports how much of its own development the model performed. Until those exist, hold two thoughts together: parts of the loop are automating faster than many expected, and a closed human-free loop remains undemonstrated.