Hassabis Says AGI in Two to Three Years: What the WSJ Reports and What His Timeline History Shows

The Wall Street Journal reports that Demis Hassabis, CEO of Google DeepMind, now expects AGI within two to three years, shortening his own earlier forecast of three to five years. The report, by Ben Cohen and published October 21, 2026, was surfaced through Techmeme's summary, which says Hassabis also discusses AI's potential for science.

Here is the honest limit of this post: we could not open the WSJ interview, which sits behind a paywall, and our searches did not surface a second outlet quoting it. So the headline claim is attributed to the WSJ as summarized by Techmeme, and we cannot confirm exact wording, context, or how Hassabis defines AGI in the interview. What we can do is verify the trajectory of his earlier public statements and explain how to read a shortened date.

TL;DR: claimed versus verified

How has Hassabis's timeline moved?

The shortened date fits a pattern. Using sources we could open:

Read as a sequence, the forecast has shortened from a five-to-ten-year window to roughly two to three in about nineteen months. That is a faster rate of change than the calendar alone, since every year that passes also removes a year from a fixed target. If the end date were fixed at 2030, a statement of "three to five years" in early 2026 and "two to three" in late 2026 would be consistent. The report describes the second as a decrease from the first, so the forecast itself is described as moving, though we have not verified the underlying quotes.

A glass ball with a small green star inside, illustrating how far an AGI timeline prediction can be trusted

What does Hassabis mean by AGI?

Definitions drive these predictions more than any date. In his July 2026 essay, as we summarized in Hassabis's frontier AI framework post, AGI means a system with all the cognitive capabilities the brain has. That is a demanding bar, and in earlier interviews he has pointed to gaps such as continual learning and long-term planning. It is stricter than definitions based on economic output or passing a task suite.

So "two to three years" from Hassabis is not directly comparable to a similar number from a leader who defines AGI as an agent that can do most knowledge work. For how the same question plays out on governance, see Altman's three endgames and Demis's plan.

Why does the AI-for-science angle matter?

Techmeme's summary says the interview also covers AI's potential in science, which is the area Hassabis has built his reputation on, from AlphaFold to the DeepMind work we followed in the DeepMind Institute survey on AI in science. For readers, the practical point is that "AGI" for him is partly a science story: systems that speed up discovery, in biology and materials in particular. If the interview makes specific claims about drug discovery or other timelines, those would be the more testable ones, and we will add them here if they can be verified.

Why should you be careful with the number?

It is a forecast, not a result. A two-to-three-year prediction tells you what a person with unusual access believes. It does not tell you what will happen. Forecasts of this type have missed in both directions for decades.

Incentives exist on all sides. A lab head has insight into internal progress and also benefits when investors, governments and talent take the field seriously. Researchers outside the labs, such as those behind the Hinton intelligence explosion paper, argue for caution from a different position. Both views are evidence about beliefs, not about timelines.

The reported change is a single data point. The Techmeme text describes a drop from "three-to-five" to "two-to-three". Without the interview we cannot tell whether he said it as a firm prediction, a hedged range, or a reaction to a particular recent system.

Other leaders disagree. Researchers like Andrej Karpathy have argued progress will come through different mechanisms, as in his remarks on models that stop speaking English, and David Silver's new venture bets on reinforcement learning as an alternative path, covered in the Ineffable superlearner post. A shared date does not imply a shared route.

What are people asking about this prediction?

Is this a change of mind or a change of wording? The Techmeme summary frames it as a shortened forecast, "down from his three-to-five-year prediction earlier in 2026." A move from a range to a narrower, earlier range can reflect new evidence, a different question from the interviewer, or a rounder way of speaking in a newspaper interview. The three explanations imply different things for readers, and only the full text can distinguish them.

Does a shorter timeline mean the risks are closer too? Hassabis has tied his optimism to risk warnings. In the May 2026 Axios interview, as relayed by Computerworld, he said today's agents should be viewed as a practice run for far more powerful systems, that governments and society are not taking this seriously enough, and that systems which improve their own development carry risks as well as faster research. A shorter timeline compresses the time available for the preparation he is calling for, which is why policy people care about the date more than the label.

Why do forecasts from lab leaders keep shortening? Several reasons are plausible, and none are proven by this report. Capabilities have advanced quickly. Leaders see unreleased systems. Public expectations have shifted, which makes a long timeline sound out of step. And as a deadline approaches, a fixed target date naturally yields a smaller number of years. Weighing these is the reader's job; the prediction alone cannot separate them.

What would count as evidence that the timeline is right? Not a model name. Stronger evidence would be systems that hold a goal over days of unattended work, learn from deployment without retraining from scratch, and produce discoveries that independent labs reproduce. The AI-for-science claim is the most checkable of these, because results either replicate or do not.

How have AGI predictions performed so far?

The record is mixed and worth stating plainly. Predictions made in the 2010s that human-level systems were a few years off were wrong. Predictions that language models would not reach their current coding, reasoning and tool-use ability were also wrong, often by large margins. Both kinds of error are common, which is why a single forecast should shift your planning only modestly. The useful habit is to track the forecasts of several informed people over time, as the table above does for Hassabis, and to ask what each would have predicted for concrete near-term benchmarks.

How can you check the claim yourself?

1. Read the primary interview. Subscribers can read the WSJ piece. Look for the exact sentence, the definition of AGI, and whether he gives a probability.
2. Compare against earlier statements. Use the table above and check that the earlier quotes match the same definition.
3. Look for second sources. A transcript, a DeepMind post, or Hassabis's own account would settle wording questions.
4. Watch concrete milestones. Long-horizon agent reliability, continual learning, and independently verified scientific discoveries say more about the timeline than any single interview.

What should builders do?

Do not plan to a date. Plan to a rate of change. Keep your stack model-agnostic so you can swap in stronger systems, invest in evaluation harnesses that tell you when a new model changes your results, and treat safety and monitoring as a standing cost rather than a one-time project. If you work on agents, the near-term question is reliability on long tasks, not a label.

Bottom line

A forecast of AGI in two to three years from the head of Google DeepMind would be a notable shift if the WSJ report holds as summarized. Until the full interview or a second outlet confirms the wording and definition, treat it as reported, directionally consistent with his earlier shortening, and not a measurement. We will update this post when more is verified.

This post reflects reporting available on October 21, 2026, and will be updated if the interview text is confirmed.

Related reading

• Demis Hassabis frontier AI framework and AGI timelines (July 2026)
• Jeff Dean leaves Google for Discovery Loop as Hassabis steps back
• Sam Altman's three endgames, Demis's plan and Eliezer's plan
• DeepMind Institute: bending the curve of discovery
• Hinton's intelligence explosion paper
• David Silver's Ineffable and the superlearner