AI systems now do much of the engineering behind their own successors, but the measurements published so far show acceleration, not the self-reinforcing runaway that the "foom" scenario describes. Anthropic says that by May 2026 Claude wrote more than 80% of the code merged into its codebase. On 6 September OpenAI said it had reached its goal of an "automated research intern", with its research organisation using, by mid-August, 3.1 agent-workdays of effort for every workday of human labour. Yet the system card for Claude Opus 5.5, published on 22 September, reports no sustained AI-attributable doubling in Anthropic's pace of progress, and METR, which tested the model externally, says its data cannot tell steady improvement from acceleration or deceleration.
A 1965 idea and the 2008 argument over its speed
In a 1965 paper, the statistician I. J. Good argued that a machine able to far surpass any human at every intellectual activity could design still better machines, setting off an "intelligence explosion" that would leave human intelligence far behind. The first such machine, he wrote, would be "the last invention that man need ever make", provided it was docile enough to tell us how to keep it under control.
The dispute over how fast that would happen took its modern form in late 2008, in a long exchange of blog posts between the AI theorist Eliezer Yudkowsky and the economist Robin Hanson. Yudkowsky expected "a fast, local increase in capability", fast meaning "weeks or hours rather than years or decades." An AI feeding its gains back into its own design, he argued, would either stall below some threshold or, above it, go "FOOM".
Hanson answered with economic history. Bursts of tools improving tools usually run out quickly, he wrote, and no more than three in human history had substantially changed the world economy's growth rate. His example was Douglas Engelbart, whose team spent years building tools to speed up its own tool-building, produced the computer mouse, and never outran the world. Progress normally depends on dozens of factors improving together, Hanson argued, so a foom needs one feedback loop far stronger than everything damping it. In a March 2023 post he still called foom logically possible but said too many people gave it high credence. If past trends continue, he wrote, the world economy is likely at some point in the next few centuries to enter a transition lasting roughly a decade, after which it may double every few months. The sceptic also expects explosive growth, eventually; what he doubts is one system leaping ahead of the world within weeks.
Even the slow-takeoff view is fast
In a 2018 essay, the alignment researcher Paul Christiano defined a slow takeoff as one in which world output doubles over a full four years before it ever doubles within a single year. He argued that AI "mediocre at self-improvement" would arrive before AI that is great at it. Even that path, he wrote, would play out over only a few years, and the world could easily fail to prepare.
A 2025 model by Tom Davidson and Tom Houlden, published by the research group Forethought, asks what happens once AI can fully automate AI research. Progress then accelerates only if each doubling of research effort yields more than one doubling of software efficiency. Their median estimate is 1.2, within a range of 0.4 to 3.6; overall they put about a 60% chance on more than three years of progress being compressed into one, and about 20% on ten, and call the exercise necessarily speculative. Ege Erdil, writing for Epoch AI in April 2025, called a software-only explosion unlikely because research is bottlenecked by experimental compute and real-world data. He still gave about a 10% chance of a tenfold acceleration in algorithmic progress, sustained for a few months, sometime in the next decade while compute stocks no more than double.
What the labs now report
The loop began closing before 2026. In May 2025 Google DeepMind reported that AlphaEvolve, a Gemini-powered coding agent, had sped up a vital kernel in Gemini's architecture by 23%, cutting Gemini's training time by 1%. In February 2026 OpenAI called GPT-5.3-Codex "our first model that was instrumental in creating itself": its team had used early versions to debug the model's training and manage its deployment.
Anthropic's June 2026 essay "When AI builds itself", updated on 18 September, gives a detailed account. In the second quarter of 2026 its typical engineer merged eight times as much code a day as in 2024, a figure Anthropic calls "almost certainly an overstatement" of the true productivity gain. Drawing on METR's measurements, it says the length of tasks models can complete on their own has been doubling roughly every four months, against every seven months earlier. At 129 moments where a researcher's own next step had room for improvement, its best model proposed a better one, as judged by another Claude model, 51% of the time in November 2025 and 64% in April 2026. Anthropic's summary: "We are not there yet, and recursive self-improvement is not inevitable", though it adds that it could come sooner than most institutions are prepared for.
OpenAI's 6 September post defines its intern as a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. Agents' success rates rose between January and July, but over the preceding six months more than half of their successful four-to-eight-hour tasks needed at least one human intervention, and the company says people still set research priorities and decide whether to scale, pause or deploy. The same day its chief scientist, Jakub Pachocki, wrote that internal results gave him a "strong expectation" that the current pace "could be sustained into recursive self-improvement."
Anthropic's measurements show a step, not a runaway
Anthropic's scaling policy treats AI research automation as crossing a risk threshold if its models could replace its entire research staff at competitive cost, or if AI-driven acceleration doubles its pace of progress. For the second test it relies on internal measures of research acceleration, only partly published, and on trend lines fitted to a capability index built from benchmark scores. In the Opus 5.5 card, a one-time jump of 5.9 points at Claude Mythos Preview, with the old slope continuing, fits better than a genuine trend break in 99 of 100 resampled fits. The alternative, a slope rising from 14.4 to 22.2 points a year from September 2025, is a 1.53x increase with a 95% range of 1.20 to 1.82: short of 2x either way. Anthropic's August 2026 risk report says its AI research is significantly faster than it would be without AI, "but not yet by a factor of 2", adding that it is uncertain and that measurement is difficult. It is less confident than in earlier reports, it says, because its most concrete task-based tests have saturated and it is seeing early signs of acceleration; the Opus 5.5 card later described these as acceleration in one or more highly relevant internal metrics. Our analysis of that test sets out how much of it rests on internal data.
METR, whose published summary Anthropic was allowed to review and edit, tested Opus 5.5 before release and judged it unlikely to be able to fully automate AI research, which METR expects, while calling this highly uncertain, to require large gains in researcher "judgement" or "taste". In what METR calls a highly experimental and preliminary report, a separate METR team with deeper access estimated about 1.5x overall acceleration of Anthropic's capability progress due to AI, with perhaps a 30% chance of 2x, but could not share its evidence with the evaluators and did not say which period the figure covers.
OpenAI's system cards for GPT-6 Astra on 3 September and GPT-6 Sol and Luna on 7 October place the models below the High threshold for AI self-improvement in the company's Preparedness Framework: an impact equivalent to giving every OpenAI researcher a highly capable mid-career research engineer as an assistant, relative to their 2024 baseline. At the Critical level, which it equates with recursive self-improvement, the framework says development halts until adequate safeguards are specified. Almost every figure here is the companies' own, and independent measurement is lagging: in February 2026 METR said it was redesigning its randomised study of developer productivity because a growing share of developers would not work without AI.
The case for a slower loop rests on bottlenecks
Both companies now describe the constraint in terms Hanson would recognise. Anthropic cites Amdahl's law, under which speeding up one part of a process leaves the whole capped by the parts that have not sped up, and says human code review has already become a new bottleneck. OpenAI writes that the least automatable tasks will become the important bottlenecks and that compute may matter more as others ease; it adds that its compute grew significantly while experiments per active experimenter reached their highest level since tracking began in January 2025.
An independent test points the same way. In a July 2026 study by 24 researchers including Arvind Narayanan and Helen Toner, frontier agents, chiefly Claude Opus 4.8, got six days and thousands of dollars of compute to tackle the central question of each of two unpublished NeurIPS 2026 submissions, with the papers' own authors grading the results. The agents did all the engineering unaided but made no substantial progress on the research questions, and both attempts were rejected. The authors describe the failures as a lack of creativity and judgement, over what counts as publishable, how to respond to a flawed design and when to abandon a dead end, though some co-authors read them instead as failures of reasoning. Two case studies are a small sample, and the models were a generation behind the frontier by October, but the finding matches what METR and both labs report: the doing is being automated faster than the deciding.
Takeoff speed decides which safety plans can work
Many safety plans assume time to notice failures and fix them. On the fast view there is little: Yudkowsky argued in 2022, in "AGI Ruin", that alignment must work on the "first critical try". On Christiano's slower picture there would be time to experiment with safety approaches, if only a few years of it. The 2026 disclosures add that the tools for noticing are weakening as the loop tightens. Pachocki writes that OpenAI's evaluations indicate its ability to rely on chain-of-thought monitoring, its primary bet for checking that alignment training works, is "progressively diminishing", and that he expects AI progress to be increasingly bottlenecked by confidence in monitoring. Anthropic warns that rare misalignment in today's models could compound as models build their successors.
The stated responses are conditional. OpenAI paused reinforcement-learning training on its latest models intended for deployment for two weeks after the Hugging Face incident, and says it and other companies should be required to publicly track their progress towards recursive self-improvement. Anthropic says it expects it would slow down or temporarily pause if other frontier developers verifiably did the same. Pachocki hopes voluntary slowdowns become commonplace until shared safety bars are set. In July, employees of frontier AI companies, 1,386 by the organisers' latest count and including Pachocki and Anthropic's chief executive, Dario Amodei, asked the US government to support an international effort to build tools "to deliberately pace the frontier of automated AI development". None of these statements is binding.
The next signals to watch are whether Anthropic's refitted trend lines steepen again or stay a single step, whether agents start passing tests of research judgement as well as engineering, and what METR's separate assessment shows once its evidence is published. Forethought's analysis of three feedback loops, in software, chip technology and chip production, expects them to be automated in that order, with time lags rising from about three months to train a new model to years to build a fab. The AI 2027 scenario is one detailed attempt to imagine what follows if the software loop closes first.
Correction, 14 October 2026: An earlier version of this article said current AI systems cannot improve themselves; Google DeepMind had reported in May 2025 that its AlphaEvolve agent had cut Gemini's training time, and in 2026 OpenAI and Anthropic reported their models doing much of the engineering behind their successors. It stated the asymmetric-risk argument backwards, saying the world would have time to adapt if the slow-takeoff view were wrong, and attributed that argument to Yudkowsky without a source we could find. It said no government or company had shown any willingness to accept constraints on AI development, although OpenAI's Preparedness Framework has said since April 2025 that development halts at its Critical level for AI self-improvement until adequate safeguards are specified. It also said Yudkowsky coined "FOOM", which we could not confirm, and called AlphaGo's move 37 one no human would have played, where DeepMind puts its chance of being played at 1 in 10,000. Those passages have been corrected or removed.