AI Accelerating AI Development

Sources#

• Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT
• Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution
• Recursive Self Improvement for Coding Agents
• When AI builds itself

Summary#

The empirical half of the Anthropic Institute's When AI builds itself — the previously-unreported internal data showing that AI is already accelerating the development of AI at Anthropic. Where public benchmarks (Task Time-Horizon Scaling) show capability rising, this page collects the deployment-side evidence that the rising capability is already feeding back into Anthropic's own engineering and research throughput. It is the present-tense ground the Recursive Self-Improvement extrapolation stands on, and the concrete instance of the acceleration the AI R&D autonomy eval gates against.

The engineering / research split (and the autonomy ladder)#

Building a frontier model takes two kinds of work, and Claude has progressed differently on each:

• Engineering (writing code, standing up infra, overseeing training): Claude "can be handed an underspecified problem and figure out how to solve it; humans supply the goal, but they no longer need to supply the method."
• Research (choosing experiments, interpreting results, deciding what to try next): Claude "can already match or outperform skilled humans at executing a well-specified experiment."

Across both, the persistent gap is judgment in choosing goals — see Research Taste as the Human Bottleneck. The essay maps capability onto the seniority ladder Anthropic uses for its own people:

1. Execute a specified task — "The export button isn't working, please fix it."
2. Set the approach for a given goal — "Investigate why the network slows down under heavy load."
3. Choose which problems are worth working on — "What should the team build next quarter?"

Claude has climbed from rung 1 into rung 2; rung 3 is the frontier.

Engineering evidence#

Claude writes most of Anthropic's code. As of May 2026, >80% of merged code is Claude-authored, up from low single digits before Claude Code's Feb 2025 research preview. (Leadership has publicly estimated 90%+ including scripts/experimental code; the >80% figure is the more conservative lines-merged-to-production attribution.)

~8× output per engineer. Lines merged per engineer per day held flat 2021–2024, then climbed with two inflection points: 2025 (Claude began to run code, not just suggest it) and 2026 (models began working autonomously over longer horizons). In Q2 2026 the typical engineer merged ~8× as much code/day as in 2024. Caveat stated plainly: lines-of-code measures quantity over quality, so 8× "is almost certainly an overstatement of the true productivity gain" — but it indicates real acceleration, and Anthropic does not reward LOC.

• A March 2026 poll (130 research-team employees) put median self-estimated output at ~4× with Mythos Preview vs no AI (Anthropic believes true uplift was somewhat lower — developer self-estimates are known to overestimate).
• Work that wouldn't have happened otherwise: in April 2026 Claude shipped 800+ fixes that cut a class of API errors 1000× — estimated at four human-years of painstaking cross-context bug-solving.

Code quality reached parity. "Good code" = it works and another engineer can build on it. On works: the rate at which staff correct/redirect/take over mid-task has fallen steadily for a year, even on open-ended problems (session success on the hardest tier reached 76% in May 2026, +50pp in six months). On legibility: Claude-written code was "somewhat worse than human-written … in late 2025, is roughly at parity today, and we expect it to be strictly better within the year."

The automated reviewer. Every change is now read by an automated Claude reviewer before merge. A retrospective found it would have caught ~1/3 of the bugs behind past claude.ai incidents before production — "Claude is now catching the mistakes that [the best engineers in the world] missed." This is verification and the reviewer-in-fresh-context pattern operationalized at org scale.

Research evidence#

Three measurements, climbing the ladder from execution toward judgment:

• Kernel/experiment optimization (rung 1–2, now superhuman). A fixed eval run every release: given code that trains a small model, make it run as fast as possible while passing the same correctness checks. Opus 4 ~3× speedup (May 2025) → Mythos Preview ~52× (Apr 2026). A skilled human needs 4–8 hours to reach ~4×. "In under a year, Claude has gone from super helpful to superhuman" at optimizing within a defined experiment. (Caveat: the absolute multiple depends on headroom in the starting code and is not a real-world training speedup; the informative part is the like-for-like comparison across models and against humans.)
• Open-ended research, end to end (rung 2). The April 2026 automated weak-to-strong researcher: agents given an open AI-safety problem (can a weak model reliably supervise a stronger one?) proposed hypotheses, tested them, shared findings across parallel agents, and iterated. They recovered ~97% of the floor-to-ceiling gap over 800 cumulative agent-hours ($18k compute); two human researchers recovered ~23% over a week. Caveats: the result didn't transfer cleanly to production-scale models, and humans still chose the problem and wrote the scoring rubric — direction-setting was the only meaningful human role.
• Research next-step judgment (rung 2–3, the frontier signal). On n=129 real research sessions where the human took a detour, models saw only the pre-detour work and proposed a next step; a separate Claude (which saw how the session ended) judged who chose better. Opus 4.5 (Nov 2025) beat the human 51% → Mythos Preview (Apr 2026) 64%. Crucial caveat on judge bias: on a separate set of 127 moments where the human's move was already strong, models were judged better only ~20% of the time — so this is not a like-for-like human-vs-model comparison, but a trend on hard, ambiguous decisions.

The same question from outside, and a null (METR, July 2026)#

Everything above is Anthropic measuring Anthropic. METR's expenditure-horizon note (2026-07-21, empirical) opens by cataloguing every way the field currently answers "how much is AI accelerating AI R&D?" and finding each one wanting — a useful frame for reading this page's own evidence:

• AI R&D benchmarks (RE-bench, MLE-bench) — mostly report no human baseline, or one at a fixed 8 or 40 hours, on problems "often not reflective of frontier-level AI R&D."
• Researcher-uplift studies — the experiment is very hard to run; the available numbers are self-reported (~4× from the Mythos Preview card, ~2× from Becker/METR), and both reports urge caution. Anthropic's ~8× merged-lines figure is real but "production of code is difficult to map to productivity in R&D" (Researcher Uplift from Code Output is METR's attempt to bridge exactly that gap).
• Qualitative reflections — e.g. 4 of 18 Mythos Preview respondents thinking it could replace an entry-level researcher with three months of scaffolding iteration. Useful, not quantitative.
• Contributions to frontier optimization problems — TTT-Discover, AlphaEvolve, LLM-assisted NanoGPT records: hard to size against human effort, and "reporting is biased towards successes."
• Capabilities-progress acceleration (time horizon, Epoch's ECI) — intrinsically lagging, observable only after the capability lands in a model, and requires separately estimating the human and compute contributions.

METR's contribution is a direct measurement of the fourth category, and it comes back deflationary. Six agents optimizing the NanoGPT speedrun autonomously from record #78, at up to $10,000 each, re-validate to gains of ~1–1.5% — "equivalent to 1-2 human contributions" against a cumulative human investment of roughly 1,650 hours (~$250,000) on the same problem. Two of the six found nothing at all once noise was removed. The maintainer would merge ~70% of the ideas but only 50–60% of the speedup, and calls the rest "mostly fiddling with knobs."

This does not contradict the page's internal evidence — it measures a different thing. Anthropic reports hybrid throughput (humans using Claude); METR measures autonomous optimization with no human in the loop, and says explicitly that "it is still possible that agents could dramatically augment human progress." The two numbers are the two ends of the same question, and the gap between them is the size of the human's remaining contribution — which is Research Taste as the Human Bottleneck priced in dollars.

An authorship figure that looks bigger and measures much less#

Ouroboros/Hope (arXiv 2608.08311, case-study) reports 94.2% of 1,085 self-modification commits agent-authored over a 161-day deployment. That reads as this page's >80% figure exceeded, and it is not comparable to it on three axes at once — worth writing down, because the two numbers will be quoted side by side:

• Different unit. Anthropic's is a share of merged lines of code; this is a share of commits. A commit-share is insensitive to size, so one line and ten thousand count the same.
• Different denominator. Anthropic's denominator is a frontier lab's whole production codebase. This one is a single agent's edits to its own repository — 175,755 published LOC total, all of it harness. It is a narrower object than "AI-adjacent engineering work," never mind AI development.
• Different authority. Anthropic's Claude writes code humans directed it to write; Hope selects its own changes, and the human role is a review gate that blocks 63.5% of recent self-edit attempts. Higher agent-authorship and a majority-blocking reviewer are the same system described twice.

And it is self-reported by the system under study, in a paper the system is credited as contributing to. The usable reading is that it belongs to this page's present-tense claim — AI compressing AI-adjacent engineering — at the scale of one harness repository, not that the authorship trend has moved. The corpus's other instance of the same shape is Cline's campaign, and both are outside-Anthropic data points about scaffolding rather than about model development.

The honest caveats#

The essay is unusually careful to bound its own evidence: LOC overstates productivity; self-reported uplift is biased upward; the kernel multiple is headroom-dependent; the W2S result didn't transfer to scale and used a human-chosen problem; the next-step test was run on deliberately-chosen weak-human moments. The load-bearing claim survives all of them: the human role is narrowing at each step, and the doing now costs almost nothing in human time (though still in compute).

Connections#

• Recursive Self-Improvement — the extrapolation this evidence grounds; "the loop is already tightening"
• AI R&D Autonomy Evaluation (AECI) — the formal capability gate (AECI, substitution threshold); this page is the deployment-side correlate the eval anticipated
• Research Taste as the Human Bottleneck — the persistent gap these measurements keep hitting: choosing goals, not executing them
• Task Time-Horizon Scaling — the external-benchmark companion (METR, SWE-bench, CORE-Bench) to this internal data
• Verification as the New Bottleneck — the automated reviewer and "review became the new bottleneck" are this thesis at org scale
• Harness Shrinkage as Models Improve — the same narrowing role: humans stop writing code, shift to direction and review
• The Bitter Lesson — "research progress is mostly a function of tools and resources" is the bitter lesson applied to R&D itself
• Agentic Loops Overtake Bespoke Systems — the same simple-loop-overtakes-bespoke dynamic, measured in formal math
• LLM-Driven Vulnerability Research — Project Glasswing as a worked example of AI-accelerated technical output
• AI-Native Startup Lifecycle — the diffusion of this acceleration into the wider economy: 100-person firms doing 1,000-person work
• Frontier Pause Verification — compounding acceleration is why "we don't have decades" to build a verification regime
• Researcher Uplift from Code Output — the economic translation of the ~8× figure: production functions turn 8× code output into a plausibly->2× serial researcher uplift (Kwa/METR), formalizing this page's own "LOC overstates" caveat as explicit verbosity / Cadillac-code discounts
• Expenditure Horizon — the autonomous-only counterpart measured from outside, and a null: $10K agent runs on a frontier training-speed problem buy 1–2 human contributions' worth of speedup, against ~$250K of cumulative human labour on the same problem
• Agent-Authored Harness Optimization — an outside-Anthropic data point of the same shape: Cline's benchmark hill-climb went from four engineers × two weeks (Jan 2026) to one prompt × 17 unattended hours (Jul 2026), on a task that is itself AI engineering
• Continuous Self-Modification Under Review — the authorship figure that is not comparable to this page's: 94.2% of 1,085 self-modification commits agent-authored, on a commit-share rather than a line-share, over one agent's own harness repository rather than a lab's production codebase, with a review gate blocking 63.5% of recent attempts

Open Questions#

• LOC, self-reports, and headroom-dependent multiples all overstate; what unbiased throughput metric would Anthropic's promised shift to "direct measurement of AI R&D acceleration and researcher uplift" (AI R&D Autonomy Evaluation (AECI)) actually use? Partially answered: Researcher Uplift from Code Output — Kwa argues code output (the 8× itself) beats per-hour code uplift because output already prices in marginal value through time reallocation and is robust to production-function assumptions; but it stays corrupted by verbosity, barely-useful "Cadillac" code, and fun-driven time-allocation shifts — so the metric it really points to is quality-adjusted code output, which still needs internal data LoC can't supply.
• The W2S result didn't transfer to production-scale models. Is that a temporary scaling artifact or a structural limit on autonomous research?
• The next-step judgment trend (51%→64%) is measured only on weak-human-move slices. What does the curve look like on a representative sample of research decisions?

Sources#

• When AI builds itself — §"Evidence from within Anthropic" (engineering + research evidence, productivity poll, kernel eval, W2S researcher, next-step judgment)
• Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution — Razzhigaev, Gritsaev, Kaznacheev, Dragunov, Yampolskiy & Kuznetsov (MSU / Skoltech / Joi Lab / AIRI, arXiv 2608.08311, 2026-08-08, case-study — downgraded from empirical at compile): Table 4's authorship and review counters only. Self-reported by the system under study; total author COI. Parse warnings and full treatment on Continuous Self-Modification Under Review
• Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT — Cunningham, Shetty, Cheng & Rush (METR, 2026-07-21, empirical): the introduction's five-way taxonomy of AI-R&D-acceleration evidence and what each one cannot support, plus the NanoGPT measurement of the "contributions to frontier optimization problems" category — autonomous agents at $10K per run producing 1–2 human contributions' worth of speedup. Full treatment on Expenditure Horizon