Recursive Self-Improvement: What Fable 5, Dario's Essay, and Anthropic's Own Data Actually Tell Us

Recursive self-improvement used to be a philosophy seminar topic. In 1965 the mathematician I. J. Good put the argument in one sentence: an ultraintelligent machine could design even better machines, so "the first ultraintelligent machine is the last invention that man need ever make." That framing, the intelligence explosion, anchored the debate for sixty years while staying safely hypothetical.

It is not hypothetical anymore, and you can date the shift precisely. In a 48-hour window this week, Anthropic released three documents that should be read as one argument:

1. The Claude Fable 5 and Mythos 5 launch - the product evidence
2. Dario Amodei's essay Policy on the AI Exponential - the policy response
3. The Anthropic Institute's When AI Builds Itself - previously unreported internal data on how much of Anthropic's own AI development is now done by AI

Read together, they are the clearest public picture yet of where the self-improvement loop actually stands. Not because Fable 5 is self-improving - it is not - but because for the first time a frontier lab is publishing the receipts.

Two Different Claims Hiding in One Term

"Recursive self-improvement" gets used for two very different claims, and most arguments about it are people talking past each other.

Strong RSI: an AI system fully autonomously designs and develops its own successor, without a human gating each cycle. This is Good's scenario, and it is the definition the Anthropic Institute uses. Their own verdict: "We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for."

Weak RSI: AI meaningfully accelerates the pipeline that produces the next AI - writing the training code, generating the feedback signal, optimizing the kernels, running the experiments. Humans still approve each generation, but each generation makes the next one cheaper and faster to build.

Weak RSI is not speculative. It now has quarterly metrics.

The Weak Loop Has Numbers Now

The Institute report is the first time a frontier lab has published internal data on this, and the numbers are blunt:

• More than 80% of the code merged into Anthropic's codebase is now authored by Claude (as of May 2026). Before Claude Code launched in February 2025, that figure was in the low single digits.
• Engineers merge 8x as much code per day as they did in 2024. Lines of code is an imperfect measure, and the report says so, but the inflection points line up exactly with agentic capability jumps.
• On a fixed kernel-optimization eval, models went from a ~3x training-code speedup (Opus 4, May 2025) to ~52x (Mythos Preview, April 2026). A skilled human researcher needs four to eight hours on the same task to reach 4x. The report's phrasing: "super helpful to superhuman in under a year."
• In Anthropic's automated researcher experiment, Claude agents ran an open AI-safety research problem end to end - proposing hypotheses, testing them, iterating. Two human researchers recovered roughly 23% of the available performance gap in a week; the agents recovered 97%. Humans chose the problem and the scoring rubric. The agents designed every experiment.

The same pattern shows up outside Anthropic. Constitutional AI replaced much of the human feedback that shapes model behavior with model-generated critiques back in 2022. DeepMind's AlphaEvolve used Gemini to discover a faster matrix-multiplication kernel that cut Gemini's own training time - a literal, measured turn of the loop. And METR's task-horizon data, the best public proxy for autonomous capability, now doubles roughly every four months, up from the seven-month doubling METR measured in early 2025. The curve everyone said to watch has already bent once.

What Fable 5 Itself Shows

Fable 5 is the first Mythos-class model - a tier Anthropic positions above Opus - released for general use. Four things in the launch material matter for the RSI question.

The autonomy is economically legible. Stripe reported Fable 5 performed a codebase-wide migration in a 50-million-line Ruby codebase in a day; Anthropic says the same work would have taken a team over two months. Frontier-lab engineering time is the input the loop consumes, and "months compressed into days" is the unit that matters.

Self-improvement through memory is measurable in-context. When Anthropic had Fable 5 play the deck-builder Slay the Spire, persistent file-based memory improved its performance three times more than the same setup improved Opus 4.8. The model "improves its outputs using its own notes." Not weight-level learning, but a system getting measurably better through its own accumulated experience.

Autonomous research is past the demo stage. Anthropic's protein-design experts report Mythos 5 accelerated parts of drug design about ten times, and that the model with tools but no human assistance "matches or beats skilled human operators" across the full cycle: choosing binding sites, running tools, recovering from failures. Swap proteins for ML experiments and that job description is the strong loop's prerequisite.

Anthropic is policing the loop directly. One of Fable 5's three classifier categories is distillation: requests flagged as attempts to extract Fable's capabilities to train competing models get routed to Opus 4.8 instead. A lab building production defenses against model capabilities flowing into other models' training treats AI-improves-AI as an active threat surface, not a thought experiment.

What Dario's Essay Does With All This

The essay is worth reading against the launch, because it is the policy half of the same argument. Three points stand out.

He cites the loop as established fact, not forecast. The essay's opening evidence for "lightning pace" is that AI has gone from barely writing a coherent line of code to "writing most of the code at major AI companies" in four years - and the link behind that phrase is the Institute's RSI report. Later, the economic-growth section argues that "the iterative ability of AI to build even better AI may supercharge that growth even further." The loop is load-bearing in both his risk story and his growth story.

Automated R&D is now a named regulatory category. The essay's centerpiece is a proposal for FAA-style mandatory testing of frontier models, with government power to block deployment. The four risk areas: cybersecurity, biological weapons, loss of control, and "automated R&D that could accelerate these other risks." Recursive self-improvement is no longer a footnote in lab safety policies - it is one of four categories in proposed legislation, with Anthropic pledging financial backing for the bill.

The timescale mismatch is the whole point. The essay opens with Treebeard from Lord of the Rings: a wise institution that takes a day to say hello, facing a problem that moves in months. Dario's stated worry is that current policy actions are "at least a year out of step" with the technology. Pair that with the Institute's four-month doubling time and the argument writes itself: by the time a bill passes, the models it was written for are two generations old.

There is an obvious tension in shipping your most capable model and calling for mandatory pre-release testing in the same news cycle - we took that apart in The Dario Paradox. But on the narrow RSI question the documents are consistent: the lab's own data says the loop is partly running, the product monetizes the running parts, and the policy proposal targets the parts that have not closed yet. The Institute report even states that Anthropic would "slow down or temporarily pause" frontier development if other labs did so verifiably - while noting that the verification regimes that made nuclear arms control work took decades the AI timeline does not offer.

What Is Still Missing for the Strong Loop

The gap between weak and strong RSI is still concrete, and to its credit the Institute report is explicit about it:

• Judgment is the unclosed gap. Claude matches or beats humans at executing well-specified experiments, but "large performance gaps persist" in choosing goals - which problems to work on, which results to trust. Anthropic measured this directly: shown real research sessions at the moment a human took a wrong turn, Opus 4.5 proposed a better next step than the human 51% of the time in November 2025; Mythos Preview hit 64% by April 2026. Better, improving fast, and still far from running its own research agenda.
• Nothing touches its own weights. Every weight update on every frontier model remains a human-approved, human-budgeted training run.
• Compute is a physical gate. Frontier training runs on hardware with multi-year supply chains. A model cannot iterate on itself faster than the GPUs exist - and the report flags chip fabrication and grid expansion as candidate binding constraints.
• Amdahl's law is already biting. Speed up one stage and the bottleneck moves. Anthropic reports that human code review has become the new constraint now that code generation is effectively free. Evaluation - knowing a change actually made the model better - has the same shape.

The honest summary: a human-supervised flywheel, not a runaway loop. Each turn is getting faster, the models are doing more of the pushing, and the remaining human contribution is narrowing toward taste and direction-setting.

What It Could Mean

The Institute report sketches three futures, and it is unusually candid about which one it expects:

1. The trend stalls. The exponentials turn out to be S-curves: research judgment cannot be scaled into existence, or compute and energy supply bind first. Anthropic includes this "for completeness" but says it does not believe it is likely: "Every capability we can measure... has so far followed the same curve. We have not yet seen that curve bend."

2. Compounding efficiency, humans steering. Development becomes substantially automated while humans set direction and judge results. Hundred-person companies do the work of ten-thousand-person ones. The report says the evidence points here, and this is also the world Dario's economic sections assume: hypergrowth with serious labor displacement, which is why the essay pairs the FAA proposal with wage insurance and job-displacement frameworks.

3. The loop closes. If capability trends continue and models develop genuine research taste, systems begin building their successors, and progress becomes compute-bound rather than human-bound. The report does not predict this, but it stops treating it as science fiction - and admits the alignment question in that world is the one it is "least certain about."

The signals worth watching are now specific: the next-step-judgment number (51% to 64% in five months) and whether the METR doubling time shrinks again. The seven-to-four-month compression already happened. Scenario one requires that curve to flatten; nothing measured so far says it is flattening.

For developers the practical read is scenario two: the skills that compound are specifying tasks well, building evaluation harnesses, and reviewing work faster than agents produce it - because review, not generation, is the bottleneck the labs themselves are hitting. We made the cost-side version of this argument in Fable 5 vs Opus 4.8.

The Version You Can Run Today

You do not need a training cluster to build intuition for compounding improvement. The harness-level loop is available to any developer now:

• Skills that update themselves. A Claude Code skill that revises its own instructions after each session is a self-improving artifact - see Self-Improving Skills.
• Memory that compounds. Persistent notes agents consult and update across sessions, the same mechanism behind Fable 5's Slay the Spire result - see Continual Learning in Claude Code and Fable 5 memory tool setup.
• Agents that learn from their mistakes. Reflection, correction tracking, and skill accumulation - see Self-Improving AI Agents.

The loop you build at the harness level is bounded, auditable, and resets when you delete a directory. The loop the labs are measuring is none of those things, which is why the Fable/Mythos split, the distillation classifier, and the FAA proposal all exist. Both loops are real. Only one of them ships to your terminal.

FAQ

What is recursive self-improvement in AI?

Recursive self-improvement (RSI) is when an AI system improves the process that produces AI systems, so each generation builds a better next generation. The strong form - a model fully autonomously designing and developing its successor - does not exist yet. The weak form - AI accelerating the code, experiments, and infrastructure behind the next model - is now documented: Anthropic reports over 80% of its merged code is authored by Claude.

Is Claude Fable 5 a self-improving AI?

No. Fable 5 cannot modify its own weights or trigger its own training. It demonstrates in-context self-improvement: with persistent file-based memory, Anthropic measured task-performance gains three times larger than Opus 4.8 got from the same setup. It improves its outputs using its own notes, not its underlying model.

What does Dario Amodei's "Policy on the AI Exponential" say about recursive self-improvement?

The essay treats AI-accelerated AI development as established fact and proposes FAA-style mandatory third-party testing for frontier models, with "automated R&D" as one of four named risk categories alongside cybersecurity, bioweapons, and loss of control. It argues policy is at least a year behind the technology and that the gap is the central problem.

What signals would suggest strong recursive self-improvement is getting close?

Two measurable ones. First, the research-judgment gap: Anthropic's models went from beating human researchers' next-step choices 51% of the time (November 2025) to 64% (April 2026) on hard decision points. Second, METR's task-horizon doubling time, which already compressed from roughly seven months to roughly four. Another compression, or a lab crediting a model with a measurable share of its successor's capability gains, would be the strong-loop signal.

Sources

• Anthropic, Claude Fable 5 and Claude Mythos 5, June 9, 2026
• Dario Amodei, Policy on the AI Exponential, June 2026
• The Anthropic Institute, When AI Builds Itself, June 2026
• Anthropic Alignment, An automated weak-to-strong researcher, April 2026
• I. J. Good, Speculations Concerning the First Ultraintelligent Machine, 1965
• METR, Time horizons and Measuring AI Ability to Complete Long Tasks
• Anthropic, Constitutional AI: Harmlessness from AI Feedback, December 2022
• Google DeepMind, AlphaEvolve: A Gemini-powered coding agent, May 2025
• Anthropic, Claude Fable 5 / Mythos 5 system card