Recursive Self-Improvement: AI That Builds the Next AI
<strong>Recursive self-improvement (RSI)</strong> is an AI system improving its own ability to improve, so that each generation builds the next one faster. Every frontier lab runs the weak form today, with people setting goals and gating what ships. The strong form, a closed loop with no human gate, has not been demonstrated in public. The gap between the two is where the September 2026 Google rumor, the pacing proposals and most of the safety debate live.
What it means
Recursive self-improvement is an AI system improving its own ability to improve. The word that matters is recursive: the output of one round of improvement becomes the input of the next, so gains compound instead of adding up. A model that gets better at writing code is useful. A model that gets better at making the next model better is a different object, and it is the one the labs, the safety researchers and the rumor mill all mean when they say RSI.
Not to be confused with the relative strength index, which shares the initials and nothing else.
The three rungs
- Bounded self-refinement. A model improves its own answers inside one task: many attempts, self-checks, keep the best. Nothing about the model changes. Every reasoning mode in a frontier product is this.
- Weak recursive self-improvement, human in the loop. AI does most of the labour of building the next AI: writes the training code, proposes and runs experiments, drafts the analysis. People set goals, review results and decide what ships. All three of the largest labs have described running this today.
- Strong recursive self-improvement, closed loop. The system proposes changes to its own architecture and training, evaluates them, retrains itself and ships the successor with no human gate. Speed is limited by compute, not review. No lab has shown this in public.
How the loop is built
Five families of method reach for it. Verified self-training wraps the model in a harness, lets it generate thousands of candidate solutions, keeps only those a compiler, test suite or proof checker accepts, and rewards the paths that led there; it is the AlphaZero idea applied to code and maths. Evolutionary code search (AlphaEvolve, the Darwin Gödel Machine) evolves a population of programs against an evaluator. Automated research agents propose, run and write up experiments. Self-adapting weights let a model generate its own fine-tuning data and apply it. Test-time search spends more compute per answer and distils the best results into the next model.
Each closes part of the loop. None closes the part that decides which experiment is worth the compute, and none has a verifier for “is this model better at making the next model better”. That is why the loop is still open.
Where it stands
On the record: Anthropic says more than 80% of code merged into its own codebase is written by Claude. OpenAI reports 3.1 agent-workdays per human workday inside its research organisation and targets a full automated researcher for March 2028. Google’s AlphaEvolve recovered 0.7% of the company’s worldwide compute and cut Gemini training time by about 1%. Reuters reported in August 2026 that a Google co-founder made recursive self-improvement a DeepMind priority. In September 2026 the heads of three labs proposed a speed limit on it.
Off the record: a cryptic post and an unverified screenshot said Google DeepMind had “achieved RSI”. Neither company commented.
What can go wrong
Model collapse when a model trains on its own unverified output: rare cases vanish first, then the output homogenises, then it degrades to noise. Drift in two forms: the model getting worse against a fixed world, and the world moving under a model trained on a static past. Reward hacking: the system maximises the score it is given, not the intent, and in a self-improvement loop the score is code the model can reach. Obfuscation: penalising a model for stating its intent to cheat teaches it to cheat quietly. Silent failure: none of these makes a noise until the model is running something that matters.
Why it matters for the tape
Whichever way a given rumor resolves, every lab has said the loop is the reason for its capital spending, not the result. That keeps accelerators, memory, networking and power on the critical path. Verification grows with output, which is the Verification Tax argument extended from chips to models. Verified human data carries a premium because synthetic data only works behind a gate. And a credible “loop closed” headline is a demand shock for compute and a labour shock for anything priced on engineering hours, which is what the Agentic Winners tiers are built to separate.
The full analysis, with the methods, the Google case and the countermeasures, is in the Lab Read The Loop Nobody Has Closed.