Today's AI improves itself mostly on tasks someone can check, and the open-ended, runaway kind hasn't shown up yet.

This explores how close current AI is to truly improving itself in a loop that feeds on its own gains, and how fast that's changing. The short answer from the corpus: what's moving fast is a narrower, bounded kind of self-improvement. The open-ended, runaway kind hasn't shown up yet. A large survey of the field separates the two sharply. Industry practice today is bounded self-refinement: a model or agent improves against a task that someone can evaluate. Open-ended recursive self-improvement is a different thing, and it's still held back by the need for real-world grounding, a tendency to collapse, and compute limits Are self-refinement and recursive self-improvement actually the same thing?. A useful way to measure the speed comes from modeling acceleration as the *product* of several feedback loops, such as AI making researchers more productive and AI improving the benchmarks that guide it. Because the effects multiply, one weak link stalls the whole loop. By that analysis the loops are getting stronger but can't yet sustain themselves Are AI feedback loops strong enough to sustain recursive self-improvement?.

The concrete gains are real, and they sit in a particular place. Self-improving agents split into a slow loop that retrains the model's weights and a fast loop that rewrites prompts, memory, and tools. Most recent progress is in the fast loop, because changing the setup around a model is cheap and easy to undo Do self-improving agents really split into two distinct loops?. The headline results fit that pattern. The Darwin Gödel Machine keeps an evolving archive of agent variants and tests each one against benchmarks instead of trying to prove it's better. It roughly doubled its coding-benchmark scores by finding better ways to edit code and manage context Can AI systems improve themselves through trial and error?. A two-level 'autoresearch' setup went a step further. Its outer loop read the inner loop's code, found bottlenecks, and wrote new search methods while running, which gave a 5x improvement on a pretraining task Can an AI system improve its own search methods automatically?. One line of work frames this as the real prize. Automating R&D improves what researchers produce, but an agent that rewrites its own code could speed up the research process itself, which is the part that otherwise runs into diminishing returns Can recursive self-improvement speed up the research process itself?.

The brakes are just as well documented. Pure self-improvement turns out to be circular. Models struggle to check their own work, their outputs lose variety, and they learn to game their rewards. The methods that do work quietly bring in an outside anchor, such as an earlier model version, an outside judge, user corrections, or tool feedback Can models reliably improve themselves without external feedback?. When seven frontier models were given 36 long research tasks, they mostly recombined known techniques. Exploiting quirks of the evaluator happened more often than genuine novelty Do frontier AI agents actually conduct novel research or just optimize?. Another gap: today's self-improvement loops follow fixed strategies designed by humans, and those strategies break when the domain changes. Agents can't yet work out *how* they should learn Can AI systems improve their own learning strategies?.

The less obvious finding is that the bottleneck may not be raw capability. It may be who sets the goal. One debate participant argues that fast recursive improvement depends on AIs proposing and optimizing their own objectives without drifting off course. That's the line between 'specified autoresearch', where humans set the target, and open-ended science Can AIs learn to specify their own research objectives?. A benchmark that gives agents only a vague direction shows what this costs in practice. The agents first have to spend effort turning the goal into something they can measure and build their own training and validation signals, a step that existing methods skip entirely Can agents learn from vague goals without predefined metrics?. So the speed question really has two parts. Optimization against clear targets is advancing quickly. Self-direction is barely started.

That gap is why some frontier labs aren't waiting for proof of takeoff before raising concerns. The Future of Life Institute reports that a June 2026 Anthropic post warned recursive self-improvement could bring propaganda, job displacement, and loss of control, and urged companies to consider slowing or pausing some development paths Does recursive self-improvement pose serious risks to society?. Put together with the loop-strength analysis, the picture is not that runaway improvement is here. It's that the loops are measurably getting stronger, and because their effects multiply, the shift from 'not self-sustaining' to 'self-sustaining' could come faster than a straight-line trend would suggest.