Loading
Loading
The point where a model stops being a tool that humans use to build the next model, and becomes the thing that builds the next model. Anthropic now says it is not here yet, not inevitable, but could arrive sooner than most institutions are prepared for. This is the decoder for that claim.
Recursive self-improvement (RSI) is a system that can, in Anthropic's phrasing, fully autonomously design and develop its own successor. Claude helps build a better Claude, that better Claude builds a better one still, and the loop closes: humans step out of the inner cycle of research, experiment, and training.
The word that matters is recursive. A coding assistant that writes 80% of a company's code is automation, the human still sets the goal, reads the result, decides what to build next. RSI is the same loop applied to the act of model-building itself, run by the model, at machine speed, without sleep gaps. It is the difference between a faster worker and a process that compounds.
Anthropic frames the path as a progression, not a switch. We are four steps in. The fifth is the one nobody has taken.
Each step removed a little more human involvement from the inner loop. Chatbots wrote snippets a person pasted in. Coding agents wrote whole files. Autonomous agents now run their own code and hand hours of work to other agents. "Closing the loop" is the step where agents build and train the models themselves, and a future Claude is improved continuously by Claude.
The reason this stopped being a thought experiment is that the early portion of the curve is now measurable. From Anthropic's own When AI builds itself report (May 2026 figures):
The single most load-bearing number is the task-length trend. The length of task a model can reliably finish on its own is doubling roughly every four months, up from every seven. Opus 3 handled about four-minute tasks in March 2024. Sonnet 3.7 reached ninety-minute tasks a year later. Opus 4.6 reached twelve-hour tasks by March 2026. Anthropic projects multi-day tasks this year and multi-week tasks in 2027. The work of building a model is a very long task. The trend line runs straight at it.
The current pace feels uneven, fast some weeks, slow others, because humans are still inside the loop: running experiments, sleeping, arguing in meetings. The thesis is that removing the humans does not just add a constant speedup, it changes the shape of the curve.
Nat Friedman put it cleanly at Stripe Sessions: the prime project at every lab right now is to remove humans from the continuous work of making models better, "eliminate all those sleep gaps and also scale it out to data center scale," and so today's pace is "probably as slow as it will ever be before it elbows up with self-improvement." The compounding substrate is already visible from inside Anthropic, where, per Krishna Rao, the models building the next generation of models is the reason the company allocates scarce compute internally and forgoes near-term revenue to do it.
The intuitive blocker is chips. The more honest one is measurement. A model can only improve what it can score, and outside rigid domains like math and code, robust evals are exactly what we lack. DeepMind's Mostafa Dehghani: "you can only improve what you can measure… the fact that we don't have evals that can measure how close we are to a self-improvement loop is making it much harder to make progress."
Close the loop without a clean, grounded feedback signal and you get model collapse: the system trains on its own degrading output and loses the ability to generalise. A tight, real-world evaluation loop is the thing standing between "AI accelerates AI research" and "AI runs away with it." Anthropic's own caveats point the same way. They note that the gap to a fully autonomous researcher is "exercising judgement in choosing goals," that benchmarks saturate before they can capture the effect on AI development itself, and that it is "genuinely unclear whether today's training methods and architectures could unlock that capacity."
There is a related unsolved piece. Today's models are frozen at the end of training; everything downstream assumes that. Continual learning, a model that updates its own weights from new experience without catastrophic forgetting, is the architectural change that would let the loop run continuously rather than in discrete training-run jumps. It is not solved. That is part of why RSI is "could," not "is."
Two consequences land before any singularity does.
Cadence compresses. If model-building accelerates, the release interval shrinks and the half-life of any single capability lead shrinks with it. Product teams already feel this as capability overhang: by the time you finish optimising a flow for one model, the next model has shipped and made your learnings irrelevant. The bottleneck for an AI company moves from developing the capability to diffusing it into a usable surface, which compounds the case that durable value sits in distribution and the product surface, not the weights, and that raw intelligence keeps drifting toward a price floor set by substitutes.
Compute demand bends upward, it does not relax. A cheaper, faster way to produce intelligence does not reduce appetite for compute; it expands the set of things worth running, which is the Jevons shape. Anthropic spending scarce internal compute on Claude-builds-Claude, ahead of revenue, is that logic in practice.
The report is not a victory lap. Its argument is that if the loop closes, the failure modes change character. Misalignment present in today's models "could compound as the models build their successors, growing more frequent but less understood until we lose control." The same efficiency that accelerates research is dual-use, it could equally power "authoritarian surveillance of whole populations."
So the ask is coordination. Anthropic wants to preserve "the option to slow or temporarily pause frontier AI development" long enough for alignment and societal structures to keep pace, and would pause "if other developers at or near the frontier also did so in a verifiable manner." They are blunt about why that is hard: "Training runs are far easier to conceal than missile silos," and the arms-control trust that makes verification work "took decades to build. We don't have that long." The honest tension they name: a pause that only lets the least cautious actors catch up "could leave everyone less safe."
Worth holding onto, because the headline figures are easy to over-read:
RSI is a trend line with a clear early slope and a genuinely uncertain elbow. The useful posture is neither dismissal nor inevitability: watch the task-length doubling time and the quality-of-judgement gap, because those two are the load-bearing variables.