Source Digest — Anthropic, "When AI Builds Itself" (Recursive Self-Improvement)

Status (June 2026): Complete standard digest. Captures the June 2026 Anthropic Institute essay that anchored the project's agent-automation verifier cluster — the steward's reading of it produced the "can civic taste/judgment be fast-tracked?" question. Promotes the inline citation in the Agent Automation and the Verifier memo §9 to a standing digest, closing ROADMAP verifier-cluster next-step 5. Companion to the Anthropic Advanced AI Framework digest (the same lab's policy proposal) — this is the same lab's capability report.

Why this digest

The verifier cluster is built on a claim borrowed from this essay: that the AI gains documented in software engineering ride on cheap, objective verifiers (tests, timers, reproduced numbers), and that the human residue is direction-setting / taste / judgment — exactly where the verifier is slow, contested, or absent. The cluster has been citing the essay inline; a project that demands link-and-quote discipline of its own sources should hold its genesis source as a proper digest. Capturing it also lets future exchanges test the essay's empirical claims against the AI Existential Risk digest and the AI Governance Practice digest, and gives Problem Map §11 a primary-source acceleration record.

Provenance caution carried throughout. This is a W3-type primary document for the claim "what Anthropic reports about its own acceleration." The most striking figures (8× code/engineer, >80% Claude-authored, the 3×→52× speedup, the 51%→64% next-step-judgment result) are internal, largely unaudited data reported by a self-interested party, with caveats the essay itself supplies (lines-of-code is a quality-blind measure; survey uplift is self-estimated and likely overstated; the weak-to-strong result "didn't transfer cleanly to production-scale models"). Treat as W3 "what Anthropic claims," never as W1 evidence the trends are real, durable, or generalizable.

Source identification

Thematic cluster 1: the recursive-self-improvement thesis

Anthropic is "delegating a growing share of AI development to AI systems themselves," and taken far enough that points to a system that can "fully autonomously design and develop its own successor" — recursive self-improvement (RSI). The essay is careful: "We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for." The named risk is loss of control: "If systems are capable of fully building their own successors, the ways we secure them, monitor them, and shape their behavior all grow much more important."

Thematic cluster 2: the evidence (the figures, with the essay's own caveats)

From public benchmarks.

• METR task-horizon doubling every ~4 months (up from ~7); Claude Opus 3 (~4-minute tasks, Mar 2024) → Sonnet 3.7 (~90-minute tasks) → Opus 4.6 (~12-hour tasks). Footnoted as a 50%-reliability measure.
• SWE-bench (real-world software engineering) went from low single digits to saturated in two years; CORE-Bench (reproducing published research — the prerequisite for original research) went from ~20% (2024) to saturated fifteen months later.

From inside Anthropic (the unaudited internal data).

• >80% of merged production code authored by Claude as of May 2026 (low single digits before Feb 2025); the typical engineer merged 8× as much code/day in Q2 2026 vs. 2024 — with the explicit caveat that lines-of-code is quality-blind and "almost certainly an overstatement of the true productivity gain."
• A March 2026 survey (n=130) self-estimated ~4× output uplift, which the essay says is "somewhat lower" in truth.
• An automated Claude code reviewer would have caught "roughly a third of the bugs behind past incidents on claude.ai before they ever reached production" — "Claude is now catching the mistakes that [the best engineers] missed."
• Research-loop optimization: ~3× (May 2025) → ~52× (Apr 2026) speedup on a fixed code-optimization task (a skilled human ≈ 4× in 4–8 hours) — with the caveat the absolute multiple depends on starting-code slack and "should not be read as a real-world training speedup."
• Open-ended research: agents recovered 97% of a weak-to-strong supervision gap over 800 hours / ~$18,000 compute (vs. ~23% for two humans in a week) — but "didn't transfer cleanly to production-scale models, and humans still chose the problem and created the scoring rubric."
• Next-step judgment: on a deliberately-hard set (n=129) where the human's move "had room for improvement," the best model beat the human choice 51% (Nov 2025) → 64% (Apr 2026) — with a bias check (n=127, where the human move was already strong) showing models judged better only ~20% of the time.

Thematic cluster 3: the perspiration/inspiration split (the cluster's load-bearing idea)

The essay's central conceptual move: "perspiration [is] becoming increasingly automated. … much of what advances the frontier is automatable; large-scale research progress is mostly a function of tools and resources." The persistent human residue is direction-setting / taste / judgment: "large performance gaps persist when it comes to Claude exercising judgement in choosing goals." In the weak-to-strong demonstration, "Direction-setting was the only meaningful role a human played."

"An area of human comparative advantage, for now, is research taste and judgment, including choosing which problems matter, which results to trust, and when an approach is a dead end."

And the verifier/bottleneck observation the cluster turns on — once code quality reaches parity, "human review will become the bottleneck" (Amdahl's law: "overall pace is capped by the parts that haven't sped up"), so "together they would build the systems needed to verify that AI outputs can be trusted."

Thematic cluster 4: three futures and the coordination proposal

Three scenarios: (1) the trend stalls into S-curves but today's capability diffuses widely (the essay deems this least likely); (2) compounding efficiency gains with humans still setting direction (the essay's expected case); (3) full RSI, with humans moving "towards oversight, validation, and verification of an expanding 'virtual lab.'" On the third, alignment is "something we are least certain about."

The conclusion is a coordination/pause proposal: a credible slowdown would need "multiple well-resourced labs … in multiple countries, agreeing to stop under the same conditions," each able to verify the others have — harder for AI than missiles because "training runs are far easier to conceal than missile silos." Anthropic says it "would slow down or temporarily pause, if other developers at or near the frontier also did so in a verifiable manner," and is convening "policymakers, researchers, civil society, and other AI companies" — explicitly: "people outside AI companies should be involved in this deliberation."

Research context

Mapping to Project 2028

• The verifier cluster's genesis. This essay is why the Agent Automation and the Verifier memo exists: it dissects the essay's "cheap verifier → automation" structure and asks where civic judgment has cheap verifiers. The Cross-Lineage Review Harness, the Civic-Bench Design Memo (CORE-Bench analog), and the Historical Parallel Test backtest extension are the cluster's answers.
• The "never automate direction-setting" line. The essay's "direction-setting is the human residue" is the empirical correlate of the project's normative line — industrialize the verifiable tiers; never automate the strategic / normative / legitimacy tier (Principle 3, Principle 4).
• The automated-reviewer parallel. "Claude is now catching the mistakes that [the best] missed" is the software-domain analog of the harness's decorrelated-review result (Pipeline Run 001); the Decorrelation Metrics memo is the project's attempt to measure that catch rather than assert it.
• Problem Map §11. A primary-source acceleration record for Domain 11 ("AI and compute power are concentrating faster than governance can respond") — and the coordination/verification-regime proposal is a concrete instance of the governance gap the domain names.

Interpretive notes

• Self-interest is the load-bearing caveat (again). As with the framework digest, the author is a frontier developer reporting data that supports its own strategic narrative. The figures are striking and unaudited; the digest records them as claims and flags the self-interest rather than laundering it.
• The essay is itself a performative document (cf. the Reflexivity / Performativity digest): publishing an acceleration narrative partly helps bring about the coordination regime it calls for. Engaging it should mean testing the project's own commitments against it, not adopting its frame.
• The honest gift it gives the project is the perspiration/judgment cut — a clean way to say what the project may automate (the verifiable perspiration) and what it must not (direction-setting, legitimacy). That cut is the cluster's spine.

What verification does and does not cover

Per Research Protocol §4.4, this digest is a citation-integrity artifact: the source URL resolves to the Anthropic Institute essay; the quoted excerpts are verbatim; the tier is labeled honestly (W3 primary document for what Anthropic reports, with an explicit self-interest + unaudited-data caveat — never W1 evidence the trends are real or generalizable); and every Project 2028 linkage quotes the target section's current title verbatim per the authoring guard. It is not a truth check on the internal figures. §4.1 self-verification checks passed for the source cited (June 2026).