Every seven months

In December 2025, Geoffrey Hinton said something on CNN's State of the Union that should have received more attention than it did. He said that AI systems' ability to carry out tasks, measured by how long the task takes a human, roughly doubles every seven months. That sounds abstract until you do the arithmetic. If today's frontier systems can handle tasks that take a human expert one hour, then in a year they can handle tasks that take four hours, and around 2030 tasks that take close to a month.

This is not a claim that machines are becoming humans. It is not even a claim about AGI. It is an extrapolation from a measurement — the same phenomenon that the research institute METR has documented with greater precision, although Hinton's formulation is more useful precisely because it is operational. The length of coherent work tasks an AI system can carry out is growing exponentially, on a timescale we can measure.

Most of the AI debate gets stuck on a different question: will "AGI" arrive, and if so when? That question rarely advances, because people use the word to mean different things. Hinton's observation has one property that makes it unusually useful in a setting where decisions actually have to be made: it does not require agreement on a definition. The curve exists regardless of what one calls the endpoint.

And here is what makes the question difficult to wave away. Three of the field's most authoritative voices point, from three very different vantage points, toward the same time window.

Demis Hassabis — the 2024 Nobel laureate in chemistry for AlphaFold and CEO of Google DeepMind — has pulled his own timeline inward. In spring 2025 he described genuine human-level AGI as five to ten years away, meaning 2030–2035. After Google I/O in May 2026, the window had narrowed: he now expects AGI around 2030, plus or minus a year, and sees 2029 as possible. When one of the field's more restrained voices moves his own estimate closer in less than a year, that itself is a data point worth noting. The caveats remain — he says several core problems still have to be solved: robust world models, memory, consistency and continual learning. He describes today's systems as jagged intelligence: models that can handle Nobel-level tests in one area and make errors a child would not make in another. His stance is not hype. He has said the technology is overhyped in the near term while its long-term impact is underestimated. But he is explicit that society is not prepared for the speed at which systems are now moving, and has likened the present moment to standing in the foothills of the singularity.

Geoffrey Hinton — the Nobel laureate in physics the same year, and the man who left Google in 2023 so he could speak freely about the risks — gives a wider interval, roughly five to twenty years, with the weight toward the lower half. He is also unusually direct about the downside: a ten to twenty percent risk that things end very badly over the coming decades. But the most important thing from Hinton is not the interval. It is the doubling.

Shane Legg — DeepMind's Chief AGI Scientist, and the man who helped popularize the term itself — is the most specific of the three. Since 2009, through several hype cycles and AI winters, he has held to a fifty percent probability of "minimal AGI" by 2028. The notable part is not only the forecast, but that Google DeepMind itself published the conversation with him on its official podcast in December 2025. An employer that wants to distance itself from such a number does not make it a featured conversation.

Legg's point is worth lingering on, because it is the most operational. He distinguishes between three thresholds: minimal AGI — an agent that can do the cognitive tasks a typical human can do — followed by full AGI and then superintelligence. And he notes the decisive point: minimal AGI is enough to change the conditions for knowledge work. Not the exotic levels. If an agent can do what a typical human does at a keyboard, large parts of administration, law, finance, coding and analysis are exposed — not in some distant future, but inside his 2028 window. Physical occupations, where one must handle a messy reality, have a longer horizon because robotics lags behind pure machine cognition.

It is easy to read this as experts agreeing that AGI will arrive. That is not what it is, and the precision matters. The three disagree about definitions, mechanisms and how bad things could become. There are also serious voices in the field who believe in a longer timeline, and there are real unsolved problems — continual learning, robust world models, physical robotics that remains far behind. The convergence is not consensus about the outcome. It is epistemic: three independent authoritative voices, with different technical and institutional perspectives, name roughly the same window, with a median somewhere around 2028–2030. If anything, that window has narrowed over the past year rather than widened. When three people who disagree about much else happen to point to the same years, that carries more weight than if one of them had said it alone.

What does one do with that? The dishonest reaction is to choose a date and plan for it. The equally dishonest opposite is to dismiss all of it as hype until proven otherwise. Both make the same mistake: they treat uncertainty as if it were a forecast.

The honest stance is duller and more useful. Treat the threshold as a variable, not a date. It has three possible states during the period: no passage at all, early passage around 2028–2031, or late passage in 2032–2037. We do not know which. But then one can ask a different question from "when will AGI arrive" — namely what is wise to do regardless of which state occurs. That is a far more manageable question.

There is an asymmetry that makes it urgent even for those who lean toward a late passage. Companies can adjust quickly: cancel one bet, scale another, redraw a department in a quarter. Institutions cannot. Education slots, transition systems, tax rules, power grids, government competence — these have lead times measured in years. That means the right reforms can arrive too late, not because they were wrong, but because the pace was not the one they were built for. At normal speed, the difference between being on time and being late may be five years. Under an early threshold passage, that difference shrinks to a quarter. The same decisions, the same reforms — but they only work if they have taken effect before the pressure becomes real.

That is the real risk. Not choosing the wrong future, but doing many things right and still arriving behind events.

The conclusion, therefore, is not panic. Not a universal basic income tomorrow, not a blanket AI ban in the public sector, not large new taxes introduced under pressure. All of that is most expensive if the signal turns out to be false — which is exactly why it does not belong in a preparedness track. The conclusion is reversible preparedness: instruments that are cheap to keep dormant and quick to activate. Mandates that can be switched on by a government decision, not by a two-year legislative process. Systems that can be scaled back again if the signal does not hold.

And — most important — place the indicators early in the chain. Unemployment statistics and GDP effects come too late; they measure a transition that has already happened. What arrives in time are the capability measurements themselves: the time horizon of agent tasks, signals from the labs, financial markets' repricing of exposed sectors. The cheapest thing a country can do is build the ability to see when the curve tips.

One final, self-exposing note. This text will age quickly, faster than most. If the doubling slows, if the core breakthroughs take longer, if 2028 comes and goes without a qualitative change — then this concern should be revised, and that would be good news. The point was never prophecy. The point is not to stand flat-footed in front of a curve that three serious people, from three different directions and independently of one another, pointed to while there was still time to prepare.

Every seven months. That is not a long time to get used to a new pace.

Sources and checkpoints

The sources below cover the central factual claims about timelines, measurement methods and expert statements. The conclusion about reversible preparedness is the author's.

Rolf Skogling runs ai-skiftet.se — a Swedish voice on how AI changes society, work and leadership.