Recursive Self Improvement

Recursive self-improvement (RSI) is the possibility that AI systems take on an increasing share of the work needed to create stronger AI systems, with improvements potentially feeding into successive development cycles ^[src-349].

Capability ladder

1. Tool-level improvement: models write, review or optimise code under human direction.

2. Automated experimentation: agents change bounded training code or parameters, run experiments against a fixed metric and retain improvements.

3. Model-assisted research: systems help choose experiments, interpret results and propose research directions while humans retain control.

4. Closed-loop model development: systems materially design, train, evaluate and secure successor models with limited human intervention.

5. Open-ended compounding RSI: successors repeatedly improve the process itself, producing sustained capability acceleration.

Current evidence boundary

• Public evidence supports the first two stages and some movement into the third. Anthropic reports increasing use of AI across its engineering and research work, while explicitly stating that full RSI has not arrived and is not inevitable.
• Anthropic’s automated-alignment experiment shows multi-agent hypothesis generation, experimentation and model training in a human-specified environment. Its lack of statistically significant production-scale transfer is an important boundary, not a footnote.
• Karpathy’s autoresearch demonstrates automated experiments in a deliberately narrow environment with a fixed file, time budget and validation metric. It is not evidence of open-ended autonomous research.
• METR’s task-completion horizons measure reliability on bounded software tasks whose duration is estimated from human experts. Extrapolation from those tasks to autonomous AI research remains uncertain.
• Forecasts such as AI 2027 are scenarios designed to expose assumptions and risks; they are not observations.
• Google DeepMind’s AlphaEvolve and Frontier Safety Framework provide counterevidence to claims that Google has abandoned automated AI R&D: they demonstrate bounded AI-generated optimisation in AI-training infrastructure and explicit governance for ML-R&D acceleration. Neither demonstrates a closed successor-model loop.
• OpenAI’s internal coding-agent monitoring and automated safety optimisation are compatible with movement along this ladder, but public first-party evidence is less explicit than Anthropic’s about RSI as the destination.
• The full Romero essay sharpens the central technical uncertainty: code and formal mathematics offer relatively objective evaluators, while research taste, scientific novelty, language quality and generalisation are much harder to score ^[src-349]. Automated experiment throughput is therefore not equivalent to autonomous research judgement.
• An agent product can create a commercial compounding loop—revenue funds compute and deployment creates workflow data—without creating a technical RSI loop. These mechanisms should be measured separately ^[src-349].
• Jack Clark’s 60%+ end-2028 estimate is a personal forecast, not an Anthropic company forecast. Anthropic’s reported internal code share and OpenAI’s account of Codex assisting its own development remain human-directed, company-reported evidence below the closed-loop threshold ^[src-349].

Governance implications

• Evaluation must distinguish task completion from research judgement, experiment selection, hidden failure modes and successor-system safety.
• Provenance, immutable experiment logs, rollback, compute budgets, approval gates and independent evaluation become more important as systems influence training or evaluation pipelines.
• Claims of RSI should specify the stage, degree of human control, scope of the optimisation target and whether improvements transfer beyond the sandbox.

Related entities

• Anthropic
• OpenAI
• Google DeepMind

Related concepts

• World Models
• Model Lab Differentiation
• Long Running Agents
• Continuous Agent Evaluation
• AI For Science

Source reference

• ^[src-349] Alberto Romero / The Algorithmic Bridge — “The Actual Reason Why Google ‘Fell Out’ of the AI Race Changes Everything” (2026-07-28; updated 2026-08-12)

Romero PDF batch: bounded loops versus RSI (2026-08-20)

• A launch claim about autonomous post-training is not sufficient evidence of open-ended self-improvement ^[src-353].
• AI-assisted mathematical discovery demonstrates search and formal reasoning under human objectives, not autonomous successor design ^[src-363].
• Karpathy’s autoresearch is a bounded, human-scaffolded optimisation loop ^[src-370].

Ornith-1.5 bounded curriculum update (2026-08-21)

• Ornith-1.5 couples model-generated tasks, task-specific scaffolds and solution rollouts with GRPO policy updates inside externally supplied environments and reward definitions ^[src-380].
• This is stronger evidence for automated curriculum and scaffold optimisation, but remains below closed-loop successor-model development: no public evidence shows autonomous architecture, optimiser, objective, compute or deployment changes, repeated open-ended cycles or general transfer ^[src-380].

Additional source reference

• ^[src-380] Ornith Team — “Ornith-1.5: From Self-Scaffolding to Self-Improvement” (2026-08-19)

September 18 evidence update

Verified on 2026-09-18; scope is the new sources below, not a refresh of all historical claims.

This batch adds serving optimization and search-policy commentary, alongside disagreement about research bottlenecks. It provides no independently verified demonstration of open-ended successor development. [src-490, src-493, src-503]

See Agent Delivery Evidence for attribution, availability and source summaries.