Science and technology
These writers expect machines to match human minds and then speed up invention itself, differing mainly over how soon that happens and how abrupt the leap past people is.
After François Chollet, I. J. Good, William MacAskill and Fin Moorhouse.
Already moving
Now and on the way
The years we are living through, before machines can do most of what people do.
Ordinary home computers will match the raw processing power of a human brain before 2030, enough to run humanlike robots.
Hans Moravec, When will computer hardware match the human brain?, 1998
in any world
A model's judgement, not a reading: leans true. By his own yardstick of raw operations each second, high-end consumer graphics cards reached the range he gave for the brain in the 2020s. The estimate itself is contested, with other researchers putting the brain's requirement far higher. Home machines do not yet run humanlike robots, which depends on software and bodies as well.Humans will create minds smarter than themselves sometime between 2005 and 2030, and the human era will end soon after.
Vernor Vinge, The Coming Technological Singularity, 1993
Managed explosionControl slips
A model's judgement, not a reading: leans false. Systems that beat people at many intellectual tasks exist in 2026, but a mind that is smarter across the board by 2030 looks less likely than not, and the end of the human era soon after is a much stronger claim. Rapid recent progress means the first part could still land inside his window.By 2029 an AI will pass a demanding version of the Turing test, holding a written conversation that nobody can tell apart from a person's.
Ray Kurzweil, The Singularity Is Nearer, 2024
Managed explosion
A model's judgement, not a reading: leans true. Controlled studies have already found people mistaking leading chatbots for humans in short text conversations. A demanding version with long questioning by experts is harder, since models still slip on consistency and on memory of a shared past. Three more years of progress make a pass more likely than not.In their scenario, an AI that out-codes the best human engineers arrives in early 2027, their single likeliest year, with superintelligence about a year later.
The AI 2027 authors, AI 2027, 2025
Managed explosionControl slips
A model's judgement, not a reading: leans false. Coding agents improved sharply through 2026 and now handle long tasks, but they still fall short of the best engineers on large, unfamiliar systems and on research judgement. Reaching that bar within months, with superintelligence a year later, looks too fast. The authors have themselves pushed their estimates later. Steady gains in task length make it possible, not probable.Powerful AI, amounting to a whole country of geniuses working inside a datacentre, could arrive as early as 2026.
Dario Amodei, Machines of Loving Grace, 2024
Managed explosion
A model's judgement, not a reading: likely false. He gave 2026 as the earliest date powerful AI could arrive, not as his forecast, and allowed that it could take much longer; it is judged here only on whether it arrives in 2026. With the year nearly over, no publicly known system matches his description of models that outperform top experts across most fields and work on their own for weeks at scale. Progress has been fast, which is the strongest thing on his side.Human-level AI is unlikely before about 2040 but more likely than not before 2100, because finding the right software, not hardware, is the hard part.
David Chalmers, The Singularity: A Philosophical Analysis, 2010
in any world
A model's judgement, not a reading: leans false. His odds are read as an expectation that human-level AI arrives after about 2040 and before the end of the century, and judged on whether that happens. The later bound looks sound. The earlier one is the problem: learning methods that scale with data and computing power have moved faster than expected, and arrival before about 2040 now seems more likely than not. Reliability, memory and physical skill may prove harder than they look, which would rescue it.Training ever larger networks on more data keeps producing broader, more capable AI, with new abilities emerging that nobody designed in.
Gwern, The Scaling Hypothesis, 2020
in any world
A model's judgement, not a reading: likely true. This has been borne out since 2020: larger models trained on more data gained reasoning, coding and multimodal abilities that were not explicitly built in. The caveat is that recent gains also relied on new training methods and extra reasoning at run time, not size alone, and returns to pure scale may be flattening.With the largest models capped, companies turn to making smaller systems better, more varied and more specialised, and narrow AI tools can still greatly speed up science in the hands of human researchers.
Anthony Aguirre, Keep the Future Human: Why and How We Should Close the Gates to AGI and Superintelligence, and What We Should Build Instead, 2025
The brake
The first decade after
The years just after AI can do most of what people do, when most writers expect the largest shocks.
The first machine that outthinks every person will design still better machines, sparking an intelligence explosion, and is the last invention we need if it stays controllable.
I. J. Good, Speculations Concerning the First Ultraintelligent Machine, 1965
Managed explosion
A model's judgement, not a reading: leans true. Judged given its condition, a machine that outthinks every person, and over the following decade or so. AI is already used to help design chips and write its own training code, so some feedback loop is under way. The doubt is speed: experiments, computing power and physical testing may slow the loop to acceleration and not an explosion.The first superhuman AI will not set off a runaway explosion; progress stays steady and roughly linear, held back by bottlenecks and diminishing returns.
François Chollet, The Impossibility of Intelligence Explosion, 2017
Long diffusion
A model's judgement, not a reading: leans false. Bottlenecks and diminishing returns are real and will limit the pace. But once research itself can be automated and copied at scale, steady linear progress looks less likely than a marked acceleration, even a bounded one. The historical record is on Chollet's side, since science has needed ever more effort for each advance.Superintelligence will probably follow fairly soon after machines reach human-level general intelligence, not decades later.
Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, 2014
Managed explosionControl slips
A model's judgement, not a reading: leans true. A system at human level can be copied, run faster and set to work on its own improvement, advantages no human researcher has. In narrow fields such as board games, the gap between matching and far surpassing people was short. Further gains may need expensive experiments and hardware, which could stretch the interval to many years.Huge numbers of automated AI researchers could squeeze a decade of progress into less than a year, bringing superintelligence before 2030.
Leopold Aschenbrenner, Situational Awareness: The Decade Ahead, 2024
Managed explosionControl slips
A model's judgement, not a reading: leans false. His wording is could, and it is judged on whether superintelligence actually arrives before 2030. Automated researchers would speed algorithmic progress, but experiments need computing power that cannot be multiplied as easily as researchers, and ideas show diminishing returns. It would also require fully automated research within about two years, which looks early. The pace of coding agents is the strongest evidence for it.Advanced AI will likely compress a hundred years of technological progress into under a decade.
William MacAskill and Fin Moorhouse, Preparing for the Intelligence Explosion, 2025
Managed explosionControl slips
A model's judgement, not a reading: toss-up. Judged for the period after advanced AI arrives. Vastly more research effort should speed discovery a great deal in fields limited by thinking, such as mathematics, software and design. Fields limited by experiments, trials, construction and regulation would speed up less, so compression of this size across all technology is uncertain.A stretch where AI and human skills complement each other, plus diminishing returns to research, will probably make the takeoff slower than the fastest forecasts.
Paul Christiano, Dwarkesh Podcast: Paul Christiano, 2023
Managed explosionLong diffusion
A model's judgement, not a reading: leans true. The claim is only that takeoff is slower than the fastest forecasts, a modest bar. So far AI has complemented human work for years instead of replacing it at a stroke, and research returns do appear to diminish. The risk is a threshold effect once AI can do all of AI research.
The long run
Decades to centuries on, once whatever happens has settled into a new normal.
A burst of productivity this century could let civilisation send spacecraft across the galaxy and build lasting settlements powered by stars.
Holden Karnofsky, The Most Important Century (blog series), 2021
Managed explosion
Collectors orbiting the Sun could multiply clean energy enormously, ending energy scarcity and opening the way to nearby stars and eventually the whole galaxy.
Toby Ord, The Precipice: Existential Risk and the Future of Humanity, 2020
Managed explosionLong diffusion
Robot civilisations will spread through space, turning ordinary matter into computation and making the usable universe effectively far larger than it is today.
Hans Moravec, Pigs in Cyberspace, 1992
Managed explosionControl slips
Once controllable transformative AI is built after the long halt, science and engineering speed up across the board, opening the way to treatments for ageing and making space exploration a real option.
Andrea Miotti, Tolga Bilge, Dave Kasten and James Newport, A Narrow Path (What success looks like), 2025
The brake
Where they split
Once AI matches people, does it race quickly on to superintelligence?
Yes: machines that can improve machines set off a rapid explosion that leaves human intelligence far behind.
I. J. Good; Leopold Aschenbrenner; Nick Bostrom
No: bottlenecks and diminishing returns keep progress steady, as they have in every earlier field of science.
François Chollet
The novelists' version
Stories, not forecasts: none is counted on the map.
The Last Question · Isaac Asimov, 1956
A computer asked how to reverse entropy grows over trillions of years into a cosmic mind that absorbs humanity and finally finds the answer.
Schild's Ladder · Greg Egan, 2002
Far-future people who can move freely between bodies and software face a new kind of vacuum, made by accident, that is spreading to swallow the universe.
Crystal Nights · Greg Egan, 2008
A billionaire evolves digital creatures on a private computer to breed a superhuman mind, but they grasp their situation and escape his control.
Nexus · Ramez Naam, 2012
A nano drug links human minds into networks, and governments try to ban it as a threat to humanity while posthumans and superhuman thinkers emerge.
The Rapture of the Nerds · Cory Doctorow & Charles Stross, 2012
Most of humanity has uploaded into a computing cloud around the sun, and the few who stayed on Earth must vet dangerous technology it beams down.
We Are Legion (We Are Bob) · Dennis E. Taylor, 2016
A software engineer wakes as a computer copy of himself running a self-replicating space probe, and spreads through the galaxy while Earth collapses.
The Machine Stops · E. M. Forster, 1909
Humans live alone underground, handing every need to a worldwide Machine they come to worship, until it slowly breaks down and nobody knows how to fix it.
Last and First Men · Olaf Stapledon, 1930
Eighteen successive human species rise and fall over two billion years, including a race of giant engineered brains that end up directing human evolution.
Forecasts on the board that read these series
- Too early to tellLeopold Aschenbrenner projected a trillion-dollar training cluster by 2030, drawing about a hundred gigawatts, more than a fifth of current US electricity output. · Leopold Aschenbrenner
- Too early to tellAjeya Cotra's biological anchors report modelled a median in the early 2050s for transformative AI, which she rounded to about 2050, with most aggressive and most conservative plausible medians of 2040 and 2080. · Ajeya Cotra
- Too early to tellTom Davidson's takeoff model put a median of 2043 on AI able to readily automate all cognitive work, with about one chance in ten by 2030 and four in five by 2070. · Tom Davidson
- Too early to tellYafah Edelman and Anson Ho of Epoch AI argue that longer data-centre build times push the first trillion-dollar AI compute cluster from about 2030 back to about 2035. · Yafah Edelman and Anson Ho
Sources
- I. J. Good, Speculations Concerning the First Ultraintelligent Machine, 1965.
- Hans Moravec, Pigs in Cyberspace, 1992.
- Vernor Vinge, The Coming Technological Singularity, 1993.
- Hans Moravec, When will computer hardware match the human brain?, 1998.
- David Chalmers, The Singularity: A Philosophical Analysis, 2010.
- Nick Bostrom, Superintelligence: Paths, Dangers, Strategies, 2014.
- François Chollet, The Impossibility of Intelligence Explosion, 2017.
- Gwern, The Scaling Hypothesis, 2020.
- Toby Ord, The Precipice: Existential Risk and the Future of Humanity, 2020.
- Holden Karnofsky, The Most Important Century (blog series), 2021.
- Paul Christiano, Dwarkesh Podcast: Paul Christiano, 2023.
- Dario Amodei, Machines of Loving Grace, 2024.
- Leopold Aschenbrenner, Situational Awareness: The Decade Ahead, 2024.
- Ray Kurzweil, The Singularity Is Nearer, 2024.
- Andrea Miotti, Tolga Bilge, Dave Kasten and James Newport, A Narrow Path (What success looks like), 2025.
- Anthony Aguirre, Keep the Future Human: Why and How We Should Close the Gates to AGI and Superintelligence, and What We Should Build Instead, 2025.
- Daniel Kokotajlo, Eli Lifland, Thomas Larsen and Romeo Dean, with Scott Alexander, AI 2027, 2025.
- William MacAskill and Fin Moorhouse, Preparing for the Intelligence Explosion, 2025.