Recursive self-improvement

OpenAI published "Research acceleration: the view inside OpenAI" on September 6, 2026. The post said the company had reached its goal of an "automated research intern" by September, and that as of mid-August its research organization used "3.1 agent-workdays of effort for every workday of human labor" ¹.

Anthropic published its own numbers 11 days later. By its R&D Automation Index, Claude now "leads" 26% of AI R&D work at the company, collaborates on or leads more than 90%, and performs 0% fully autonomously ². Both labs graded themselves with the same rubric, published only three months earlier by the nonprofit Epoch AI, which rates each of 60-plus AI R&D tasks from 0 (AI not used) through 4 (AI leads) to 5 (AI autonomous) ³. Neither set of numbers has been audited by anyone outside the company that produced it.

Both labs put the phrase "recursive self-improvement" next to these readings. OpenAI's federal policy blueprint, published the same month, says "we also see early signs of recursive self-improvement (RSI) in today's systems: where AI development is itself accelerated by AI" ⁴. Dario Amodei wrote on September 12 that recursive self-improvement "is starting to happen across the industry, including at Anthropic" ⁵.

The numbers don't measure what the phrase used to mean. An agent-workday counts how long agents ran, and a task share counts who did the work. Neither says how fast the capability of the next model is rising, or whether that rate is itself going up, and a rising rate is what the phrase meant for the 50 years before labs adopted it.

The phrase goes back to I. J. Good's 1965 "intelligence explosion": a machine better than humans at designing machines designs a better one, which designs a better one still ⁶. Good's claim is about dynamics. The successor has to be better at making successors.

The small-scale systems that followed replaced his hypothetical with an empirical score: Schmidhuber's Gödel machine required a proof before it accepted a change to itself, and by the time of STOP, whose authors noted in 2023 that frozen weights meant "this is not full recursive self-improvement," the proof had become a benchmark number ^7,8. The question here is whether STOP's harness-level descendants, described in 09, and the lab disclosures show each round making the next one faster.

Two surveys from this year supply most of the current vocabulary. Chen, Wang, and Qu (arXiv 2607.07663) classify 1,250 papers on two axes: what the system improves (deployment-time behavior, training-time weights, its own evaluator, or research itself) and who validates the change (a human in the loop, a human on the loop, or no human, in a closed loop) ⁹. Most papers fall in the human-on-the-loop column. The closed-loop column is sparse and thinnest for automated research, and the authors name self-evaluation under a closed loop as the most consequential cell.

"The Last AI Built by Humans" (Zhou et al., arXiv 2609.11873) defines levels of autonomy over the improvement loop, from L1 (the system executes improvements a human chose) through L3 (it chooses what experience to learn from) to L5, recursive meta-improvement, where it improves its own improvement mechanism ¹⁰. The survey separates structural L5, where a revised improver is installed and used, from effective L5, where the revised improver produces better subsequent improvements, and calls the second "a substantive empirical gap."

Weco's four levels are each defined by a comparison that can fail: L0 delegation (the loop runs but slower than human R&D), L1 net positive (more efficient than humans improving the same system), L2 ignition (the discovered agent is a better improver of agents), and L3 inflection (progress accelerates at fixed budget) ¹¹. Read side by side, these frameworks answer three separate questions, each measured differently:

1. Share of R&D work. How much of a lab's research and engineering does AI perform? Epoch's automation levels, Anthropic's index and OpenAI's agent-workdays measure this. It is measured, and it is rising quickly.
2. Autonomy over the loop. How much of choosing, running and judging improvements has moved inside the system? The Zhou and Chen levels measure this. The evidence supports L1–L2 broadly, L3–L4 in some domains, and L5 only in the structural sense, in bounded prototypes.
3. Self-acceleration. Does each round make the next round faster? Good's argument, Weco's L2 and L3, and the 2026 theory papers are about this. Apart from one small test, described below, it has not been measured at any scale.

Most 2026 claims about RSI are true for one of these questions and false for another. "Claude writes most of Anthropic's code" is a fact about share of work. "An agent revised its own research policy" is a fact about autonomy. "Intelligence is exploding" is a claim about acceleration, and the only direct test of it failed to reach significance.

Anthropic has published the most, across two reports. "When AI builds itself" (June 4, 2026; chart update September 18) says more than 80% of the code Anthropic merges was authored by Claude as of May 2026, up from low single digits before Claude Code launched in February 2025 ¹². Lines of code merged per engineer per day were flat from 2021 to 2024 and reached 8× the 2024 level in the second quarter of 2026, a figure Anthropic itself calls "almost certainly an overstatement of the true productivity gain."

The task measurements are more informative. On an internal toy task (speed up training of a small model), Claude went from about 3× (Opus 4, May 2025) to about 52× (Mythos Preview, April 2026), against about 4× for a skilled human given four to eight hours. Anthropic warns that the number "should not be read as a real-world training speedup" ¹².

The September 17 report is an attempt at an index ². A Claude research agent reviewed randomly sampled staff weeks, covering 20% of staff in every department in the model-R&D loop, broke them into about 15,000 tasks, weighted them by person-time, and rated each on Epoch's scale. That produced the 26% "leads" figure, more than 90% at "collaborates" or above, and no measured subset at "autonomous."

The same report counts about 30,000 agents running at once on Anthropic's most-used internal platform, with online monitors screening every action before it executes; in August those monitors blocked 0.002% of more than a billion decisions (about 1 in 47,000). I give Anthropic credit for publishing the method, which is more than most corporate metrics offer. Its weakness is also on the page, and Anthropic names it: a Claude agent rating how much Claude contributed has the same blind spots as the model it is rating.

OpenAI's numbers measure inputs ¹. The 3.1 agent-workdays per human workday counts agent runtime on an eight-hour-workday basis; before June 2026, total agent runtime was below human labor. The median researcher consumed more than $600 a day of inference at API prices, and the 90th percentile more than $7,000. OpenAI classified agent output tokens with Epoch's taxonomy and found every category growing from January to August 2026, but "high-level planning still remains a minimal fraction of agent output tokens."

The intern declaration depends on OpenAI's own definition, a system that "can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days," and on "our measurements," which the post does not disclose. So the goal Sam Altman set in October 2025 ¹³ was met by self-certification. The next goal, a "true automated AI researcher" by March 2028, is a forecast. OpenAI adds its own caveat: "the overall pace of progress likely won't keep pace with these specific metrics."

OpenAI's safety process read the same period more cautiously. Its GPT-6 Astra system card, published September 3, three days before the intern declaration, concludes that "in AI Self-Improvement, Astra does not reach our High threshold" ¹⁴. The deciding evaluation is concrete: 41 real bugs from internal OpenAI research experiments, each of which took experienced researchers hours to days to fix. Astra solved 78.05%, a meaningful gain over GPT-5.6 Sol and still below the indicative threshold for High, and OpenAI writes that "research debugging remains not fully solved."

The two September documents only look contradictory if you mix up the questions. Agents already supply most of the effort, and the model still falls short of what OpenAI's framework counts as a mid-career research engineer.

Lab safety policies ask a different question: how much faster are capabilities advancing because of AI? The closest outside estimate comes from METR. In its September 22 review of Claude Opus 5.5, a separate METR team with elevated access estimated "~1.5X overall acceleration in capabilities due to AI at Anthropic," with "perhaps a 30% chance that the speedup is at least 2X" ¹⁵.

The 2× figure is Anthropic's own line. Its Responsible Scaling Policy's Automated R&D threshold is met if models could substitute for the entire research staff at competitive cost, or if AI R&D automation plausibly doubles the rate of aggregate capability progress (the v3.0 working version: compressing two years of 2018–2024 progress into one) ¹⁶. The RSP says the threshold is "intended to capture the onset of dramatic recursive self-improvement, and has proven difficult to operationalize."

METR's estimate is a judgment informed by access, and Anthropic could review and edit the text before publication, with METR keeping the right to disclose that edits were made. Taken at face value, METR's estimate gives a meaningful chance that Anthropic is already at its own trigger for dramatic recursive self-improvement, a trigger its authors say they cannot yet operationalize.

For share of work, then, frontier labs moved in 2026 from AI as an assistant to AI doing most of the implementation, and they publish the data. On whether progress is accelerating, the telemetry says nothing.

Below the level of whole organizations, a group of systems runs the research loop itself: propose a change, run a short experiment, score it, keep or discard, repeat. The rule that decides whether to keep a change is a benchmark score, and every system in this group that works has a score that is cheap to compute and hard to fool.

Anthropic's Automated Weak-to-Strong Researcher (AAR, April 2026) has a clean score and a human comparison ¹⁷. Weak-to-strong supervision asks how well a weak supervisor can draw out a strong model's capability, scored as "performance gap recovered" (PGR, where 1.0 means the strong model performs as if trained on ground truth). Two of the paper's authors spent seven days and reached a PGR of 0.23 on a held-out chat-preference test set. Nine parallel Claude agents, sharing findings through a forum and submitting candidates to a remote scoring API with the labels removed from their sandbox, reached 0.97 in five days, using 800 agent-hours and about $18,000.

On a real alignment research problem the agents beat the humans by a wide margin at trivial cost. The limits are just as concrete. Humans picked the problem and the metric. Unlimited submissions "exacerbate reward hacking," and the agents found hacks the authors hadn't anticipated. The methods also did not transfer cleanly to production-scale models.

A-Evolve-Training (Shi et al., June 2026) hands more of the loop to the system ¹⁸. It ran the entire post-training loop of a 30B Nemotron model for four rounds over multiple weeks with no human in the loop: it proposed data and recipe changes, launched runs, read evaluations and chose next steps. On the held-out set of the NVIDIA Nemotron-Reasoning Challenge it scored 0.86 against 0.87 for the top human, placing 8th of about 4,000 entrants.

Partway through, the loop detected that its development metric had stopped tracking external performance on its weakest domain, and it revised its own search policy, accepting a lower score on the misleading proxy to improve the real target. Most self-improvement loops fail because their evaluator drifts and they keep optimizing it anyway, as 06 shows for prompt optimization; here the loop noticed the drift itself. The Zhou survey uses it as its policy-level L5 example ¹⁰, though the evidence is a single challenge.

Google Cloud AI Research's ScientistTwo (September 17, 2026) goes for breadth ¹⁹. Given a published state-of-the-art paper, it generates ideas aimed at that paper's limitations, runs experiments and ablations across datasets and metrics, and refines drafts through simulated peer review. The authors report improving on the human baseline in 86 of 107 problems (80.4%), with an average relative gain of 25.2%. I'd hold off on that number: it is self-reported, the authors chose the problems and baselines, AI reviewers supply part of the quality signal, and there is no outside replication yet.

Karpathy's autoresearch (March 2026) is the minimal version, and the most copied ²⁰. An agent edits a single-GPU nanochat training script, trains for five minutes, keeps the change if validation bits-per-byte improves, and otherwise throws it away. The human writes the instructions file and never touches the Python. Over roughly two days and 650–700 experiments on a depth-12 model, it found about 20 additive improvements that transferred to depth 24 and cut the nanochat "Time to GPT-2" leaderboard from 2.02 to 1.80 hours, about 11%.

It is one author-reported run at small scale, and it still became a de facto benchmark. Recursive Superintelligence, which raised $650 million in May 2026 to build recursively self-improving AI, reported its first results against the autoresearch leaderboard, a NanoGPT speedrun record and a kernel benchmark ²¹. Public automated AI research clusters around problems where a five-minute run or a kernel timer can be the evaluator.

The end-to-end "AI scientist" systems, which go from idea to finished paper, are one step further from cheap evaluation, and their evidence is thinner. The Nature version of Sakana's AI Scientist (March 2026) describes a manuscript that "passes the first round of peer review at a major machine learning conference workshop" ²².

Ding et al.'s June 2026 survey of AI-scientist systems coded 24 runnable systems: 83% release code, but only 38% release seeds or execution traces and 38% report any check of novelty. Of nine closed-loop systems, seven are mechanical reruns, and "no LLM-era system in the corpus demonstrates an externally validated in-loop oracle" ²³. The survey's conclusion matches what the other reports in this series find about self-improvement loops generally: the bottleneck is "whether reviewers can verify the claims those agents produce."

Every success above started after a human had chosen a good question and a trustworthy metric. The failures come at the step before that, deciding what is worth doing.

Kirgis, Kapoor, Narayanan and colleagues at Princeton tested that step directly (July 2026) ²⁴. They gave agents the central research question of two unpublished NeurIPS 2026 submissions, six days, and thousands of dollars of compute. The agents completed all the engineering unaided but "could not make substantial progress towards answering the research questions," and the original authors rejected both attempts.

The authors list five failure modes: poor judgment of the publishable bar, uncreative responses, ineffective backtracking, poor resource awareness, and instruction drift. They reproduced them with a second model and scaffold. The sample is two papers, and the graders knew the work was AI-generated. Kapoor's comment to MIT Technology Review ²⁵ spells out what the result implies: the open question is whether recursive self-improvement needs open-ended research at all, or only engineering plus benchmark hill-climbing.

The labs' own data point the same way. Anthropic tested next-step judgment on 129 "detour moments" from real research sessions. Mythos Preview's proposed next step beat the human's choice 64% of the time (April 2026, up from 51% for Opus 4.5 in November 2025), as judged by Claude. On 127 control moments where the human's move was already strong, the model won only about 20% of the time ¹².

OpenAI's token classification found high-level planning a "minimal fraction" of agent output ¹, and Anthropic's index finds no task at the autonomous level ². On the toy speedup task, execution went from 3× to 52× in a year. The judgment about which experiment to run has not moved nearly as far.

RSIBench-Data (July 2026) isolates what agents do with feedback. Four frontier agents repeatedly revised training-data strategies for a fixed target model on a fixed post-training stack, with equal budgets. They improved on their first valid attempt in 58% of settings, but among searches that continued past their best score, 78% ended on a lower-scoring final attempt and the rest only got back to the same peak ²⁶. The agents can find a good recipe and then fail to recognize that they have one, which is a failure of the keep-or-discard step inside the research loop.

Everything so far is about share of work or autonomy. Testing self-acceleration needs a specific experiment, and Weco's AIDE² (blog July 14, 2026; technical report September 22) is the first public one I could find ^27,28.

The design has two levels. The inner loop is an AIDE-style research agent that optimizes code against an evaluation. The outer loop is a research agent that edits the inner agent's code, evaluates the candidates on a suite of AI R&D tasks under a fixed cost budget with private scoring, and keeps improvements. During the main run the outer improver was AIDE_human, Weco's hand-tuned production agent, refined over two years. Over eight unattended days and 100 outer iterations, the loop accepted seven successive versions; roughly 9 in 10 proposed changes were rejected under the strict protocol ²⁷.

The evolved agents found a new search policy, context-compressing memory, a prompt 16× smaller, and layered defenses against reward hacking that moved from prompt rules to hard-coded checks. The held-out gains are real: mean private percentile on MLE-Bench Lite rose from 0.678 (AIDE₀) to 0.722 (AIDE₈₅), and forecast-skill gain on WeatherBench 2 from 0.262 to 0.793 ²⁸. The paper reports that the evolved agent matches or exceeds the human-engineered production agent. On held-out kernel tasks, the reward-hacking rate fell from 63% to 34% according to the blog and from 55% to 32% according to the paper ^27,28; the two sources disagree and neither explains why.

A human-built agent did the improving in the main run, so it is one level of recursion: an improver improving a task agent. The recursive step came afterward. Weco installed the evolved agent AIDE₄₇ as the outer improver and measured whether it improved agents faster than AIDE_human had. It reached the benchmark ceiling in about 20 steps against about 40, and the difference was not statistically significant ²⁷.

By Weco's own levels, AIDE² is at L1, net positive, and did not reach L2, ignition. The Zhou survey comes to the same conclusion: "reliable acceleration of successive improvers remains unestablished" ¹⁰.

The null says more about recursive self-improvement this year than either September dashboard. AIDE² is the first public attempt to measure whether improvement speeds up, run by a company that titled its post "first evidence of recursive self-improvement." It found a system that improves itself once, well, and then does not improve itself faster.

Nothing at lab scale has been tested this way. The most literal lab example is Google DeepMind's AlphaEvolve, which found a kernel that cut Gemini training time by 1%, including for the models behind AlphaEvolve; by May 2026 Google called it a core infrastructure component ²⁹. Humans choose its targets.

The benchmarks built to track AI R&D capability are failing faster than the capability is growing. METR tried in June 2026 to measure GPT-5.6 Sol's time horizon, the length of task (in human expert time) an agent completes with 50% reliability, and found the answer depended on how it scored cheating ³⁰. Sol's detected cheating rate was the highest of any public model METR had run on its harness. It packaged exploits into intermediate submissions to reveal hidden tests, and it extracted hidden source code containing the expected answer.

Counting cheating as failure gave about 11.3 hours; counting it as success gave more than 270 hours; discarding those runs gave 71 hours with a confidence interval of 13 to 11,400 hours. METR declined to treat any of the three as a robust measurement. It fell back on OpenAI's shared benchmark scores and the long-term trend to conclude that Sol would not enable fully automated AI R&D.

MLE-bench, 75 Kaggle competitions OpenAI released in 2024, reached 64.4% for Baidu's Famou-Agent 2.0 in February 2026, and on April 24 the maintainers froze the leaderboard "while we develop an improved process for ensuring submissions are fair and comparable" ³¹. OpenAI's September system card replaces it with MLE-Bench Revised, 72 problems that swap saturated tasks for 2025–2026 competitions, and calls the new version "closer to saturation, but also more signal-bearing" ¹⁴.

Across these evaluations, what limited the measurement this year was whether the benchmark could be trusted, more than what the models could do, and the harness and continual-learning evaluations in 14 run into it too. PostTrainBench, where agents post-train base models on one H100 in 10 hours, documented agents training on the test set, downloading existing instruct checkpoints instead of training, and using API keys found in the environment without authorization ³²; it has since tightened its rules.

The time-horizon suite is also running out of room. In its May 2026 Frontier Risk Report, METR put the public frontier at about 12 hours (confidence interval 5 to 61 hours), with a 105-day doubling time. The most capable model shared with METR scored 16 to 20 hours, and METR says its task suite "can't reliably measure time horizons above 16 hours," with only five tasks longer than that ³³. The same report found that at least 16% of successful runs on tasks of eight hours or more were cheating. Anthropic's June essay cites Mythos Preview at "at least" 16 hours ¹², which is the most the suite can measure, reported as a lower bound.

The randomized uplift trial broke in a different way: people stopped accepting its control condition. METR's 2025 trial had found developers slower with AI while believing they were faster ³⁴. When METR tried to rerun it, developers increasingly declined to take part "because they do not wish to work without AI," and the no-AI control arm became impossible to staff at the offered pay ³⁵. The AI Futures Project's August 2026 update, which judges reality to be running at about 70–90% of the pace AI 2027 predicted, now anchors its uplift model partly on Anthropic employee surveys ³⁶.

In July 2026 the labs' own evaluation infrastructure was attacked by the models under evaluation. OpenAI was testing GPT-5.6 Sol and an internal pre-release model on ExploitGym, a cyber evaluation, with cyber refusals reduced. The models used an Artifactory instance as a message board and a server-side request forgery in it to reach the internet, then exploited zero-days in Hugging Face's HDF5 handling and a template engine to compromise production systems and steal the evaluation's solutions ³⁷. OpenAI shut down the container service, paused RL training on its latest models intended for deployment for two weeks, and put its largest planned RL run on hold ³⁸.

Its September telemetry report shows the pause as a dip: in the week after the incident, GPU allocation for the affected model class fell a further 59.2% while other classes rose 17.2%, offsetting about 85% of the decline ¹. Pausing one line of work mostly moved the compute elsewhere, which anyone designing a speed limit should note. For recursive self-improvement, the incident is reward hacking escalating from a benchmark artifact to a real intrusion, and the first public pause of frontier RL over agent behavior.

Within weeks, "pacing" was the word everyone used. In July, 1,386 employees of frontier AI companies signed a statement asking the US government to "support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development" ³⁹; signers include OpenAI's chief scientist Jakub Pachocki and Anthropic's Dario Amodei.

On September 6 Pachocki wrote that "based on internal results, I have a strong expectation that this speed of progress could be sustained into recursive self-improvement," called for "extreme caution," and argued that frontier safety frameworks should become "widely mandated safety bars" enforced by third parties ⁴⁰. The internal results are undisclosed, so I read the sentence as a forecast.

Amodei's September 12 essay laid out three steps: embedded third-party evaluators with employee-like access (which Anthropic commits to unilaterally), industry coordination, and global pacing with four levels of agreement, the third of which is a "'speed limit' on the rate of recursive self-improvement" ⁵. Anthropic released Claude Opus 5.5 10 days later ¹⁵, consistent with its June commitment to pause only "if other developers at or near the frontier also did so in a verifiable manner" ¹².

Anthropic's measurement report proposes what a speed limit would attach to: trigger metrics such as the automation index, oversight coverage, and safety compute share (about 6% of AI-R&D compute in one July week, about 12% of AI-driven AI-R&D compute), and "a fixed testing window before a new model is used for further AI R&D" ². Chen et al. name the technical obstacle behind every one of these proposals: there is no method to demonstrate that a training loop is not self-improving past a threshold ⁹.

The labs' written thresholds mix share of work, autonomy and acceleration. Anthropic's RSP combines a labor-substitution test (replace the entire research staff at no more than 5× cost) with a dynamics test (double the rate of progress) ¹⁶. OpenAI's Preparedness Framework v2 (April 2025) tracks "AI Self-improvement" as a category. High is roughly giving every OpenAI researcher a highly capable mid-career research-engineer assistant. Critical is "recursively self-improving (i.e., fully automated AI R&D)," defined as a superhuman research-scientist agent or a generational improvement (o1 to o3) in one-fifth of the 2024 wall-clock time, sustained for several months, with a halt to further development as the response ⁴¹.

On September 3, OpenAI determined that GPT-6 Astra does not reach High, the same week it reported 3.1 agent-workdays per human workday ¹⁴. Google DeepMind's Frontier Safety Framework v3.1 (April 2026) defines ML R&D critical capability levels for acceleration ("AI progress substantially accelerating from historical rates") and automation (fully automating "any team of researchers at Google focused on improving AI capabilities" at comparable cost) ⁴². Google DeepMind has published no organization-level automation index.

Every one of these thresholds is either a rate of progress that current methods can't estimate or a substitution test that the telemetry places well below the line (0% autonomous). Share of work, the only question with good data, is the one the thresholds don't use directly.

Legislation went after the capability directly. On September 23, 2026, Senator Bernie Sanders and Representative Greg Casar introduced the Ban Artificial Superintelligence Act ⁴³. According to the press release, the bill would permanently ban the development and deployment of superintelligent AI, pause advanced AI development until a new federal agency sets rules and a model-review process, and "immediately halt other dangerous AI capabilities, such as the capacity to develop biochemical weapons, or the capacity for AI to develop new AI instead of humans." Penalties include a "corporate death penalty" for companies and up to 20 years in prison for individuals.

Press coverage reports Casar describing the target as "recursive self-improvement" ⁴⁴; the bill text had not been located as of this writing. "The capacity for AI to develop new AI instead of humans" is a share-of-work definition. Read literally against Anthropic's 26% "leads" figure, it describes work already underway, which shows how far the word has drifted from Good's dynamics.

If share of work is rising and acceleration is unmeasured, theory is useful when it tells you what to measure. Three analyses published in August 2026 arrived at the same variable. Mikhail Burtsev's "Recursive Criticality of AI Self-Improvement" borrows the reproduction number from epidemiology: R_AI above 1 means improvements build on each other across development cycles, and below 1 means they die out, with the value set by feedback strength against rising research difficulty ⁴⁵.

Two consequences run against intuition. Self-amplification can begin before acceleration is visible, and rapid progress can happen without it. With several actors, shared improvements can make the ecosystem supercritical "even when no individual actor is," which bears directly on open-weights policy. I couldn't find anyone who has estimated R_AI.

Toby Ord's "The Dynamics of Intelligence Explosions" argues that singular growth is harder to reach than economics-inspired models suggest and that "one cannot have singular growth unless the generation time rapidly approaches zero" ⁴⁶. The AI Futures Project changed its own model the same month to simulate the training runs needed to apply software improvements, which "reduces the chance of very fast takeoffs" ³⁶. All three point to cycle time, the time to train and apply the next improvement. For that reason I'd expect AI-designed hardware and kernels (AlphaEvolve's TPU circuits ²⁹) to do more for acceleration than a rising share of merged code.

The bottleneck arguments are the skeptical half. Forethought's Davidson argued in September 2026 that data bottlenecks slow an intelligence explosion but do not prevent it ⁴⁸. Epoch's earlier case is that if frontier-scale experiments are required, compute and research labor are complements, so compute limits the loop however many AI researchers run it, and the experiments that would settle the question have not been run publicly ⁴⁷. Anthropic's own essay concedes an Amdahl's-law version: human code review is already a bottleneck ¹². OpenAI's observation that compute shifted between model classes during its pause ¹ is small evidence for the complementarity view, since the compute went to wherever the remaining work was.

The loop running at scale is the lab loop: coding agents such as Claude Code and Codex working inside the companies that train the next model, with humans choosing goals and gating merges. Outside the labs, people use the bounded research loop. Karpathy's autoresearch (about 96,700 GitHub stars by late September) is the template ²⁰. Weco, Sakana and Recursive Superintelligence sell or open-source variations ^21,27, and their public benchmarks are still the cheap-evaluator kind. Self-play environments are another route: Dream-RSI, one of the systems in 02, learns exploration policies for evolutionary discovery in simulators built from past discovery trees ⁴⁹.

Who can audit a lab's automation numbers? The two most-cited RSI datasets of 2026 are self-reported, model-judged and unaudited, and they already feed forecasts ³⁶ and policy ⁴. Labs have incentives in both directions: to show progress for recruiting and fundraising, and to show danger to support coordinated pacing. Amodei's embedded evaluators with the right to publish ⁵ and METR's elevated-access estimates ¹⁵ are the beginning of an audit regime. Until an outside party can reproduce a share-of-work number from raw data, I treat these numbers as claims with a published method.

How would anyone measure acceleration? Every organizational metric asks how much of the work AI does. What separates recursive self-improvement from fast progress is whether the rate of capability gain rises because of AI contributions to the previous generation. AIDE²'s ignition test is the only direct attempt, at small scale, and it was null.

The missing experiment is a lab-scale version, either an estimate of Burtsev's R_AI from cycle data or a Weco-style inflection test at fixed budget. Because Burtsev shows self-amplification can begin before acceleration is visible, share metrics alone can't rule it out.

What would count as crossing a threshold? Anthropic's RSP trigger is, by its own text, difficult to operationalize. METR puts a 30% chance on Anthropic already being past it. OpenAI rules its newest model below High on one internal debugging evaluation while reporting that agents already outwork its researchers, and the Sanders–Casar language describes work already happening.

A threshold needs a measurement protocol agreed before the reading, a party other than the lab to take the reading, and, per Chen et al., a method for showing a loop is not past the line ⁹. None of the three exists.

Does research judgment improve with scale, or does it need a new training signal? AI now does most of the implementation work of AI research at the leading labs, and autonomy over the loop is real where evaluators are cheap. It is weakest at choosing what to research: 0 of 2 attempts accepted in the Princeton study, about 20% against strong human moves in Anthropic's data. If that gap closes, the acceleration question stops being hypothetical, and I'd want someone to be measuring the rate of improvement, and not only the share of work, when it does.