AI Accelerating AI Development

Sources#

• Chris Painter's testimony to the U.S. Senate on AI agent incidents
• Debating RSI, the US-China Gap, and Jaggedness with JS Denain of Epoch AI
• Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT
• How employment is changing at Anthropic
• Noam Brown – Agent swarms, alignment, & recursive self-improvement
• Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution
• Recursive Self Improvement for Coding Agents
• What if automating AI R&D triggers an intelligence explosion?
• What This OpenAI Insider Saw That Made Him Quit | The Ezra Klein Show
• When AI builds itself
• Why I still haven't bought into true RSI

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).

A rival lab's version, and the confound that breaks every number on this page (September 2026)#

The caveats above are about instrument bias — LOC overstates, self-reports inflate, the kernel multiple depends on headroom. Brown (Dwarkesh Podcast, 2026-09-17, practitioner-opinion) names a definitional confound instead, and it applies to every figure on this page regardless of how well the instrument is built:

"Are you comparing it to a speedup of three years ago? Is the question more, 'Given what we were doing three years ago, how much faster are we able to do it now?' versus 'Given what we're doing now, how much slower would it have been three years ago?' Those are actually two very different questions."

They diverge because the work itself moved. Brown's mechanism is jaggedness plus substitution: the models are "exceptionally good at looking over data sets and checking every single data point," so "you can disproportionately use the AIs for those things" — and "of course, if something is suddenly 100x faster and 100x better, you're going to do more of that thing." The forward counterfactual asks how fast today's task mix would have been three years ago, on work that partly did not exist then; the backward one asks how fast the old mix runs now, on work nobody does any more. Neither Anthropic's ~8× nor Kwa's back-solve states which it is computing, and the gap between them is not a measurement error that more careful instrumentation closes.

He adds a third confound this page's caveats do not carry: "if it's the human directing the AIs to do the work, how much do you attribute to the human? How much do you attribute to the AI?" — an attribution problem that a code-authorship share papers over by construction, which is the same objection the 94.2%-agent-authored figure already invites from the other direction.

And the datum he actually offers is an input measure, which is the interesting part. Rather than a productivity multiple, Brown cites OpenAI's internal-acceleration blog post reporting that "the top 1%" of its researchers were spending $7,000–8,000 a day on Codex as of early August, "on an exponential." That is the first non-Anthropic internal-acceleration number in this corpus, and it is deliberately not a claim about output: a spend curve says what was bought, not what it produced. Held against this page's other series it is the honest complement — the LOC and self-report figures measure an output with a contested denominator, and this measures an input with none. It is also the same activity-versus-capability distinction that made Ouroboros's four monthly series uninformative, arriving this time from an author who does not mistake it for a capability claim.

Tier it low. No denominator for "top 1%", no team size, no absolute headcount, no link to the post (which is not in raw/), and the figure is recalled in conversation from a blog post by an employee of the company that publishes it and sells the product being bought.

A second OpenAI datum, from the safety side (October 2026). David Robinson, OpenAI's system-card lead until October 2026 (What This OpenAI Insider Saw That Made Him Quit | The Ezra Klein Show, practitioner-opinion), describes the change as "night and day for our research teams specifically, who use far more agentic compute … than anybody else at the company — more than 100 times more than they did at the beginning of the year," citing the same internal blog post Brown recalls. He also gives the first mechanism in the corpus for where that compute goes. Research infrastructure is "pretty jank internally" because it constantly changes, so much of the work is "getting different pieces of machinery to talk to each other," which Codex now does well. His observable is a demand-side signal: traffic fell on the Slack channel where researchers asked colleagues to fix broken experiments, because they ask Codex instead. Tier it with Brown's figure. It is a growth multiple with no base, recalled from a post not in raw/, and the help-channel decline is unquantified. It does corroborate Brown's spend curve from a second OpenAI employee in a different function.

A third reading of the same post, by function and by task (September 2026). JS Denain of Epoch AI, on Lambert's podcast (Debating RSI, the US-China Gap, and Jaggedness with JS Denain of Epoch AI, practitioner-opinion), gives the most detailed account yet of what the OpenAI post shows. He is an outsider reading a public post, so this is a third-party summary of a document that is still not in raw/.

• Rate. Researcher spend on Codex rises about 2× a month in the post's plot. He offers a measurement-artifact caveat: researchers may have used ChatGPT before Codex, so part of the curve could be migration rather than new use.
• By function. Codex spend "in other parts of the company kind of had a huge surge in the spring and then kind of plateaued in the summer". For researchers, the data team and engineers "it keeps growing, even accelerates sometimes." This is the first breakdown in the corpus showing internal AI use rising in the R&D functions after it flattened elsewhere.
• By task. OpenAI classified sessions by task type. Usage grew on engineering and troubleshooting, but "for a lot of the high-level strategic decision-making, they both anecdotally and also when they look at sessions, don't seem to find a huge uplift in making better compute allocation decisions or deciding on research directions." This is the first lab-side instrument in the corpus pointed at the top rung of this page's ladder, and it reads null (Research Taste as the Human Bottleneck).
• Serving efficiency. He notes that OpenAI reports savings on serving costs from better kernels, with few details given. That is the efficiency-first reading Lambert's essay predicts.

His verdict on the whole public record: "I don't think this is strong evidence that in six months we have a software intelligence explosion." He also doubts that the labs privately see something that warrants "being way more freaked out". He speculates about what an internal early warning might be: team KPIs such as "compute multipliers in the pre-training team", combined with intuitions about how team outputs add up, giving a predicted trend in end-model performance. "It's unclear to me that the recent discourse we've seen is evidence of things going crazy on those metrics." Tier it with Brown's and Robinson's figures: relayed, with no denominators.

A newer Anthropic series, quoted second-hand: 1% to 26% autonomous (September 2026)#

The 22-author assessment on Software-Driven Intelligence Explosion (What if automating AI R&D triggers an intelligence explosion?, arXiv 2609.36054, practitioner-opinion) opens with lab figures. Its co-authors include Anthropic's Jack Clark and OpenAI's Jakub Pachocki. One of its figures is new to the corpus.

• Code share, restated. Citing the essay this page compiles (Favaro & Clark), the paper puts the AI share of approved code at "low single digits to over 80% between January 2025 and May 2026". That matches the >80%-of-merged-code figure above and is not a new measurement.
• Autonomous share of R&D work: 1% → 26%, March to August 2026. The paper cites Favaro & Wright, Measurements for understanding the pace of AI development inside frontier labs (Anthropic, September 2026, not in raw/). It describes the measure as "the proportion of R&D work autonomously completed with only high-level human supervision". That is rung 2 of the ladder above, measured as a share rather than shown by example. The paper does not say how "work" is counted (sessions, tasks, or hours) and gives no intermediate points.
• What 26% is consistent with. (Wiki arithmetic, not the paper's.) Suppose 26% of the work now costs no human time and humans remain the bottleneck on the rest. Amdahl's law then caps the human-labor speedup at 1/(1 − 0.26) ≈ 1.35×. That is the same order as METR's "~1.5X" for Opus 5.5 and the Risk Report's "by less than a factor of 2" (AI R&D Autonomy Evaluation (AECI)). So the new series does not contradict the lab's own acceleration readings. It shows how a share that grew 26-fold in five months still yields a small multiple. The multiple rises steeply only as the share nears 1.
• OpenAI and Google, through METR. Via METR's Frontier risk report (February to March 2026), the paper quotes OpenAI: "AI assistance is used in practically all parts of the company … with code-executing agents used in training, evaluating, and securing future models". It quotes Google: "AI is used in almost all work that involves writing code or configuration, technical design, research ideation". Neither quote has a number.
• A capability limit from OpenAI's own card. GPT-6 "fails some of OpenAI's research debugging tasks that experienced human researchers can complete (albeit in hours or days)", per the GPT-6 Astra system card as the paper reads it.

The paper also cites OpenAI's Research acceleration: The view inside OpenAI (September 2026) as one of the labs' voluntary AI-R&D indicator reports. That is plausibly the internal-acceleration post that Brown, Robinson and Denain relay above, but the paper quotes no figure from it, so the identification is unconfirmed. Tier all of this as second-hand lab data. The paper is practitioner-opinion, and every number in this section is a lab's self-report, quoted by authors who include that lab's staff.

(2026-10-26: a second source cites both posts, and quotes a figure from each.) METR president Chris Painter's Senate testimony (Chris Painter's testimony to the U.S. Senate on AI agent incidents, 2026-09-30, practitioner-opinion) quotes the Anthropic post directly: "Claude 'leads' 26% of Anthropic's AI R&D work", up from 0–1% in February–March. He defines "leads" as completing tasks end-to-end from high-level human input, without ongoing supervision. That matches the paper's 1%→26% and its rung-2 reading. It also quotes the OpenAI post, the first figure the wiki has from it: "as of mid-August, in total, the research organization uses 3.1 agent-workdays of effort for every workday of human labor." The two numbers measure different things. Anthropic's is the share of work agents finish without supervision. OpenAI's is agent effort per unit of human effort, which says nothing about how much of that effort became accepted output. Neither post is in raw/, so both stay second-hand lab self-reports. Painter uses the OpenAI ratio as an oversight number, not an acceleration number: agent activity already outruns the humans who could read it.

The same firm, measured by its economists (October 2026)#

Wilmers, Catipon, Massenkoff & McCrory, How employment is changing at Anthropic (Anthropic Institute, 2026-10-16, empirical, first-party) is the second Anthropic self-study on this page and the first by labor economists rather than by the policy team. It measures the whole firm, not only R&D, and it brings three things the essay above does not have: agent-hour telemetry, a within-person design, and a code-quality test.

• Agent hours, directly. Employee-directed interactive agent runtime per human work-hour rises from ~0.12 (February 2026) to 6× by August, and to 19× in research organizations (Figure 2, chart viewed). Counting fleet and autonomous agents too, total agent work hours passed human hours in March. The human denominator is headcount × an assumed 40-hour week. Against Painter's OpenAI figure above (3.1 agent-workdays per human workday, mid-August), the units differ and neither scope is fully stated, so the two ratios are not a ranking of labs.
• Commits per engineer 5.1×, indexed to December 2025, with the series ending around June 2026 (Figure 3, Loess-smoothed). The prose says engineers commit at "5x their velocity as six months ago". This is not the essay's ~8×, which is merged lines per engineer per day against 2024. The units differ (commits against lines) and so do the bases (December 2025 against 2024). Since Figure 3's series was already rising before December 2025 (index ~50–130), the two are compatible and say nothing about each other's size. Neither is superseded. Data scientists' queries (3.1×) and legal contracts per employee (3.1×) also rise, and so do deals per salesperson (1.7×), hires per recruiter (1.4×) and invoices per AP processor (1.4×). The report concedes the confound itself: "It is easier to close deals when you sell a strong model."
• A within-person design, with its selection stated. 192 engineers who adopt coding agents are compared with 74 who do not, using person and calendar-month fixed effects. Treatment is a person's first month with at least five agentic sessions. Code added rises about 1.0 log points in the adoption month and about 1.85 two months after (CI ~0.95–2.75), with flat pre-trends (Figure 4, values read off the plot). For the first fully autonomous agent, the pre-trend is already rising (about −0.43 at month −4), so the authors read it as engineers who were already ramping, not as an effect. They name the selection channel: "if employees on the cusp of driving some big new project decide to tool up for it… this would bias our estimates."
• No measurable code-quality penalty. For the same engineer, month and directory, agent-written code against human-written code shows no difference in lines removed within 90 days, removal by a bug-fix, explicit reverts, new lint suppressions, TODO/FIXME/HACK notes or flaky-test markers. All six confidence intervals include zero. This is a different instrument from the essay's legibility claim, and it agrees with it. The report adds its own caveat: "AI errors may be harder to identify than human ones."
• Harder tasks after each model release. An automated grader rates deidentified sessions by depth, scope and number of systems involved. The share of tasks above the previous model's 90th difficulty percentile rises from 10% to 13% after a release and stays there, while the easy end barely moves. Employees take the release as a reason to attempt more, which is Brown's "you're going to do more of that thing" measured at one firm.

The same report's other half is where the human time goes when agent hours reach 6×: about 80% of interactive session time spent waiting for a person, and meta-work doubling to 22% of sessions (Human Attention as the Agent-Output Bottleneck).

Connections#

• Noam Brown — the source for the two-counterfactuals distinction, the substitution confound, and the OpenAI Codex internal-spend figure; also the rival lab's 3× median for the accelerated regime, on Intelligence Explosion Dynamics
• Intelligence Explosion Dynamics — where the acceleration figures on this page become a curve-shape argument, and where Brown's hedged 3× sits beside Kwa's ~2.5× as the corpus's only two quantified estimates of the same order
• 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
• Software-Driven Intelligence Explosion: where the present-tense acceleration on this page becomes the starting point of a growth model. The same source supplies the 1%→26% autonomous-work series
• Human Attention as the Agent-Output Bottleneck — the human side of this page's telemetry, from the same firm: agent hours at 6× (19× in research), ~80% of interactive session time waiting on a person, and reading loads far beyond a work week
• Frontier-Lab Safety Culture — a second OpenAI voice on internal acceleration: research teams at >100× their start-of-2026 agentic compute, and a falling help-channel load as researchers route infrastructure fixes to Codex

An Anthropic line on the rate, quoted second-hand (September 2026). Nathan Lambert (Why I still haven't bought into true RSI, practitioner-opinion) quotes the Claude Fable 5.1 & Mythos 5.1 system card: "We believe that internal usage of recent AI models has been a key factor in maintaining the current rate of progress, but we do not yet see clear signs of dramatic acceleration beyond that rate." The card is not in raw/, so the quote rests on Lambert's transcription. It fits the Mythos Preview card's "well short of a sustained, AI-attributable doubling" and the essay's own hedges. Unlike the essay's "AI already makes Anthropic move much faster", it credits AI with sustaining the rate rather than raising it. Lambert reads both labs' published internal measurements the same way. In his reading the biggest automation gains are in "software engineering, monitoring logs, managing planned experiments, and other fairly routine (but not always easy) tasks", which is this page's engineering column and not its research column.

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? Settler (2026-10-21): a production-scale replication of the automated W2S run, or a lab report that it was attempted.
• 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? Settler (2026-10-21): the same next-step eval rerun on a random sample of research decisions rather than weak-move slices.

Sources#

• Chris Painter's testimony to the U.S. Senate on AI agent incidents — Chris Painter (METR president), Senate HSGAC subcommittee testimony, 2026-09-30 (practitioner-opinion). Cited only for fn. 29 (Anthropic's 26% "leads" figure and its definition) and fn. 30 (OpenAI's 3.1 agent-workdays per human workday). Both are second-hand: neither lab post is in raw/
• Noam Brown – Agent swarms, alignment, & recursive self-improvement — Noam Brown (OpenAI) with Dwarkesh Patel, Dwarkesh Podcast, 2026-09-17 (practitioner-opinion). §00:22:02 only: the attribution confound, the jaggedness-plus-substitution confound, the two-counterfactuals distinction, and the Codex internal-spend figure ($7,000–8,000/day for the top 1% of OpenAI researchers, early August, "on an exponential"). Weakest-tier provenance for the number — recalled in conversation from an OpenAI blog post that is not in raw/, with no denominator, team size or absolute headcount, by an employee of the company selling the product being bought. The three confounds are the durable part and are method claims rather than measurements
• 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
• What This OpenAI Insider Saw That Made Him Quit | The Ezra Klein Show — Ezra Klein with David Robinson (OpenAI system-card lead until his October 2026 resignation), The Ezra Klein Show, 2026-10-07 (practitioner-opinion, YouTube auto-captions with ingester-assigned speaker labels). COI: Robinson holds OpenAI equity; the host's employer is suing OpenAI. Cited for the >100× research-team agentic-compute figure (same un-ingested internal blog post as Brown's) and the Slack help-channel observation. Full treatment on Frontier-Lab Safety Culture
• Why I still haven't bought into true RSI — Nathan Lambert, Interconnects, 2026-09-19 (practitioner-opinion): the Fable 5.1 & Mythos 5.1 system-card sentence on internal usage maintaining rather than accelerating the rate (second-hand: the card is not in raw/), and Lambert's read of where the labs' internal automation is concentrated
• Debating RSI, the US-China Gap, and Jaggedness with JS Denain of Epoch AI — Lambert with JS Denain (Epoch AI), Interconnects podcast #19, 2026-09-22, practitioner-opinion. §00:00–16:49: Denain's reading of OpenAI's internal-acceleration post (Codex spend ~2×/month for researchers, the by-function plateau, the session analysis finding no strategic-decision uplift, kernel serving savings) and his guess at internal leading indicators. Third-hand on the numbers: an outsider summarizing a post not in raw/
• What if automating AI R&D triggers an intelligence explosion?: Chan, Mindermann and 20 co-authors (including Clark of Anthropic and Pachocki of OpenAI), arXiv 2609.36054, 2026-09-28 (practitioner-opinion). Section "AI is rapidly automating AI R&D": the restated code share, the second-hand Favaro & Wright 1%→26% autonomous-work series (Anthropic post not in raw/), the METR-relayed OpenAI and Google quotes, and the GPT-6 research-debugging limit. The 1.35× Amdahl bound is wiki arithmetic
• How employment is changing at Anthropic — Wilmers, Catipon, Massenkoff & McCrory (Anthropic Institute, Economics), How employment is changing at Anthropic, 2026-10-16, empirical (tier corrected from case-study). Cited for Figure 2 (agent hours per human work-hour, viewed), Figure 3 (output per person by function), Figure 4 (within-person adoption event study and the quality panel, values read off the plot) and the post-release difficulty shift. COI: first-party, like everything else on this page