Will MacAskill's argument is that AI able to do the work of human researchers could deliver a century of technological progress in less than a decade, while the people and institutions who must respond keep working at today's pace. He set it out with Fin Moorhouse in "Preparing for the Intelligence Explosion", published by the research group Forethought on 11 March 2025. To make the claim vivid, the paper replays 1925 to 2025 at ten times the speed. The Cuban Missile Crisis, which in reality ran for thirteen days in October 1962, lasts 31 hours, and President Kennedy has 20 minutes to decide how to answer Khrushchev's ultimatum. The authors quote Robert F. Kennedy's reflection that a decision forced within 24 hours would probably have produced a course "filled with far more risks."
The 31 hours belongs to a thought experiment about the past. If every process sped up equally, history would simply run faster and come out the same. The danger lies in what the authors call asymmetric acceleration: technology can be sped up, while human thinking, learning and institutional schedules mostly cannot. MacAskill, a co-founder of Giving What We Can, 80,000 Hours and the Centre for Effective Altruism, is a senior research fellow at Forethought and was previously an associate professor of moral philosophy at Oxford.
Alignment is one grand challenge among many
Much of the paper catalogues what the authors call grand challenges: hard-to-reverse decisions that a compressed century would force in quick succession. Takeover by misaligned AI is on the list, and the paper says it is still underappreciated. So are new weapons of mass destruction, AI-enabled concentration of power, mechanisms that could lock in one set of values, the moral status of digital minds and races to claim resources in space. The authors reject the view that everything hinges on alignment, with an aligned superintelligence left to solve the rest. Some challenges will arrive before such a system can help, they argue, and some remedies, such as agreements to share power struck while nobody knows who will hold it, exist only beforehand.
The forecast does not depend on AI improving itself
The case for speed is arithmetic about research effort. Human research effort grows by less than 5% a year, the paper says, while on current trends the research effort supplied by AI grows at least 25-fold a year, mostly from cheaper inference and more chips. On the authors' growth model, a century in a decade requires total research effort to rise about 600-fold over the decade. Their conservative scenario for the decade after AI matches human researchers, with no software feedback loop, still gives about 5x a year. Allowing for slow experiments and scarce capital, they judge a century of progress in a decade "more likely than not."
The 80,000 Hours podcast's description of its March 2025 interview with MacAskill says the change is coming "probably in three to seven years." That is the show's summary. The paper says AI is likely to match human research effort within two decades and could approach it well within the next ten years, and that the explosion may well begin within a decade. In the interview MacAskill says he takes human-level AI within three to five years very seriously, and concedes that many of his views are not shared by superforecasters.
Three feedback loops, probably in this order
A companion paper, "Three Types of Intelligence Explosion" by Tom Davidson, Rose Hadshar and MacAskill, published on 17 March 2025, splits the mechanism into three loops. In the software loop, the subject of the long argument over hard and soft takeoff, AI improves algorithms, data and training methods. In the chip technology loop, AI takes over the research and design work of firms such as NVIDIA, TSMC and ASML. In the chip production loop, robots directed by AI build the fabs and the chips. The authors expect automation in that order, partly because of time lags: about three months to train a new frontier model, many months to bring better chip technology into production, years to build a fab.
The less obvious claim is that an explosion does not hinge on the first loop. Absent human bottlenecks such as regulation, the authors put the chance that the software loop alone sustains accelerating progress at about 50%, the chip technology loop at about 65% and the two combined at about 75%, and they call a full-stack explosion involving all three "very likely to accelerate eventually." On the podcast MacAskill said the industrial route is more likely than a software-only one, has a very high ceiling, and may favour authoritarian states, which can sustain high reinvestment and sweep regulation aside. A software explosion would most likely start in the US and concentrate power with the owners of chips and algorithms, the paper judges, while a full-stack one may be more likely to involve industrial powers such as China and the Gulf states.
What labs and evaluators have measured
In May 2025 Google DeepMind reported what amount to small contributions to the first two loops: its AlphaEvolve system had sped up a key kernel in Gemini's architecture by 23%, cutting Gemini's training time by 1%, and had proposed a circuit change integrated into an upcoming TPU chip. The wider measurements show fast gains on tasks and slower gains in output. METR, a research nonprofit, reported in March 2025 that the length of software task frontier AI agents can complete half the time, measured by how long skilled humans take, had been doubling roughly every seven months since 2019; its frontier risk report of 19 May 2026 fits a doubling time of 105 days to frontier public models released since January 2024. The same report says a METR trial of open-source developers using late-2025 public agents found productivity gains of about 4% to 20%, which METR expects is an underestimate, and that the companies it assessed did not report evidence of dramatic speed-ups in their overall pace of progress attributable to AI. Explaining why large productivity gains had not doubled its pace, Anthropic used the diminishing-returns logic the paper itself models: doubling research labour might raise output only 1.15 to 1.3 times.
On 6 September 2026 OpenAI said that, by its own measurements, it had reached its goal of an "automated research intern," able to carry out well-defined research tasks under human direction, including some that would take a skilled researcher a few days. By mid-August its research organisation was using 3.1 agent-workdays of effort for every workday of human labour, though more than half of successful four-to-eight-hour tasks still needed at least one human intervention, and the company cautioned that overall progress probably will not keep pace with such metrics. On 22 September Anthropic's Claude Opus 5.5 system card reported that its internal measures do not show a sustained AI-attributable doubling in the pace of development. The better-fitting reading of its capability trend is a one-time jump; the alternative puts the increase in slope at 1.53x. A preliminary METR estimate quoted in the card is roughly 1.5x acceleration in capabilities due to AI, with perhaps a 30% chance of 2x, over an unspecified period, and our analysis of that test sets out the caveats.
These figures measure something narrower than MacAskill's claim, which concerns technological progress once AI matches human researchers, and they come largely from the labs themselves. As the nearest public evidence on the first stage of that sequence, they show AI doing a fast-growing share of research work while the central estimates of AI-driven acceleration, at one lab, remain below a doubling.
MacAskill favours slowing the explosion at its start
The paper's proposals are mostly about decision-making: AI tools that check claims, forecast and advise policymakers; removing rules that stop civil servants using frontier models; agreeing in advance on a trigger for the start of the explosion, such as AI doing half the work of machine-learning researchers; and a union of computer scientists who would refuse to work for companies or governments that cross certain red lines. It also lists slowing the explosion itself, for instance through mandatory pre-deployment testing.
His position on pausing is easily misread. "I do recommend slowing the intelligence explosion," he told 80,000 Hours in 2025. In a note published on 26 January 2026 he sketched one way to do it: a threshold defined by benchmarks and an expert panel, a US commitment made in advance to pause frontier AI development for one month once it is crossed, and a convention during that month to draft treaties, with input from countries that verifiably pause too. He concedes that the plan is "a very high-level sketch," written as part of a line of research he does not currently plan to pursue further. What he opposes is a pause on capabilities now. In an interview published on 22 April 2026 he argued that it would let laggards catch up while compute kept accumulating, so that if the pause broke, progress could suddenly be much faster. He also judged cheaper safety measures far more likely to happen than an international pause.
His list of concrete projects, written with Moorhouse and published on 26 March 2026, puts first an independent organisation to evaluate AI character, including whether models live up to their developers' own specifications. The AI advisers he wants leaders to rely on would need such checking, which is harder when models can recognise that they are being tested and their visible reasoning may not reflect what drives their answers. A version of his concern then came from inside the labs: a statement published in July 2026, which 1,386 employees of frontier AI companies had signed by mid-October on the campaign's count, asks the US government to back an international effort to build tools "to deliberately pace the frontier of automated AI development." It sets no threshold or trigger, as our report on it explains.
The case that it will be slower
Two published critiques dispute the speed. In "AI as Normal Technology", published on 15 April 2025, Arvind Narayanan and Sayash Kapoor argue that transformative economic and social effects will take decades, because adoption is limited by how quickly organisations and institutions can adapt. They reject fast-takeoff scenarios and expect automation of AI development to grow gradually. Both camps agree that institutions are slow; they disagree about whether that slowness throttles the technology or leaves society behind it.
Writing for Epoch AI in March 2025, Ege Erdil and Matthew Barnett estimated that R&D spending explains only about 20% of US labour productivity growth since 1988, and argued that because AI research needs compute for experiments as well as ideas, a software-only singularity, as they call it, should by default "be treated as an unlikely outcome." MacAskill concedes that the software loop may fizzle for lack of compute, which is one reason he argues the industrial route deserves more attention than it gets. The paper's answer to the uncertainty is insurance: much of what it recommends costs little if the explosion never comes.
Viatopia is his answer to what the transition should aim for
In an essay published on 7 January 2026, MacAskill defines viatopia as "an intermediate state of society that is on track for a near-best future," a waystation whose name, he writes, means "by way of this place." He sets it against utopian blueprints, which later generations have tended to find dystopian, and against fixing urgent problems one at a time, which he thinks cannot set priorities when many huge problems arrive at once. His provisional list of its ingredients includes material abundance, very low catastrophic risk, preserved options and enough stability that those structures cannot easily be overturned. It ties back to the paper's worry about value lock-in, and to a question that, the paper notes, safety advocates have treated as secondary: what AI should be aligned with in the first place.
Correction, 14 October 2026: An earlier version of this article said that MacAskill does not want to pause AI development and thinks a pause unachievable. He has said he recommends slowing the intelligence explosion and has sketched a one-month pause once a defined threshold is crossed; what he opposes is a pause on capabilities now. The earlier version also attributed to him a three-to-seven-year timeline taken from a podcast's episode description, presented co-authored papers as his alone, misstated his current role, put in quotation marks phrases that do not appear in his work, and asserted without a source that an AI-driven chip-design feedback loop was already running at several manufacturers. It described the three feedback loops as a single chain set off by the first, whereas the companion paper argues that an explosion could come through the later loops even if the first fizzles. Those passages have been corrected or removed.