Google DeepMind’s Demis Hassabis puts AGI within three years, and asks rules to wait

Every headline about Demis Hassabis this week carries a countdown, and the countdown belongs to the Wall Street Journal. When WSJ Magazine asked the Google DeepMind chair how close artificial general intelligence is, his own answer was shorter and looser. “It’s soon,” he said. “I’m updating nearer and nearer.” The magazine turned that into a window of a few years, and the window became the news.

The gap between those two things matters less for whether Hassabis is right than for what kind of statement this is. Until recently he was the careful voice among the big lab bosses on timelines. He now holds an Alphabet job built around AGI arriving. And he has said openly that he picks his timeline language to make governments move. A forecast from that seat is also a lever, and the useful question is what it is meant to move.

What Demis Hassabis told WSJ Magazine

The profile, written by Ben Cohen, ran as WSJ Magazine named Hassabis its Innovator in Tech for 2026. AI Weekly points out that the two-to-three-year figure everyone is quoting is the Journal’s own phrasing, not a sentence Hassabis said. What he did say is about mission. “The ultimate goal was to build AI and understand the mind, then use it to understand science,” he told the magazine. “Step one, solve intelligence. Step two, solve everything else.”

Physics would come first. “I’m hoping that I’ll be sitting there with my AGI systems and some great human scientists, trying to work out the next level of physics,” he said. On curing cancer, he sees “hard things, but not impossible things.”

An AGI clock that moved four times in one year

The compression itself is real, and fast. At Davos in January, Hassabis put AGI five to ten years away, according to Startup Fortune and Tech Times. In New Delhi in February, at the India AI Impact Summit, it was roughly five years (Tech Times reports five to eight). At Sequoia’s AI Ascent in April he named 2030, a position he called “pretty consistent”. Speaking to Axios in May, he kept 2030 as his central estimate but allowed for 2029, Computerworld reported. Now the Journal’s two to three. The Free Press Journal notes that outlets do not even agree on the starting point, which tells you how loosely these numbers travel.

The seat changed too. In August, Alphabet moved Hassabis out of the DeepMind chief executive role, with Koray Kavukcuoglu taking day-to-day operations, and into a newly created chief scientist post. His memo to staff said he felt AGI “is close at hand.” After Google I/O in May, he told Axios he had chosen blunt language on purpose, to create urgency among governments, economists and the public, as Startup Fortune recounts. The timeline is a forecast and a message at once.

Same calendar, different finish lines

The number also hides how high Hassabis sets the bar. His AGI is a system with the full range of human cognitive abilities, not a strong score on narrow tests. His proposed check is the Einstein test, and it comes in two versions. In New Delhi he described training a system on knowledge available up to 1911 and asking whether it can derive general relativity on its own. The WSJ profile, as AI Weekly quotes it, sets the cutoff at 1901 and asks for the 1905 papers instead. No published result shows any system passing either.

That is why the convergence with rivals is thinner than it looks. Anthropic‘s Dario Amodei said at Davos in January 2025 that AI “better than almost all humans at almost all tasks” was likely within two to three years, the same span Hassabis has now landed on. Tech Times calls this researchers predicting different things that collide on the same calendar year. Yann LeCun, formerly Meta’s chief AI scientist, still argues that today’s large language models cannot get there at all, and raised more than a billion dollars for AMI Labs to prove another route.

Hassabis on regulation, one week after Dublin

The sharpest line in the profile is about rules. “What would be wrong is to try and regulate things now that we don’t even fully understand,” Hassabis said. “I think you’re almost certainly going to get the wrong regulation — and wrong regulation is worse than no regulation.”

A week earlier, at an AI summit in Dublin, Reuters heard a more anxious Hassabis. Recent hacks carried out by AI agents were “early warning shots,” he said, minor only because “systems aren’t that powerful yet,” and they could turn major “looking into next year and the next few years” without more security work. He called for practices borrowed from biology and nuclear safety, an eventual international standards body, and a certification process “proportionate with the risks.” On the same stage, Microsoft president Brad Smith told the industry to get over its “allergy” to regulation.

The two positions reconcile in one way. Hassabis wants measurement first, testing and certification, and rules only once the measurements exist. That is a coherent engineering view. It also runs on time, and his own clock is shrinking it. The “understand it first” period he asks regulators to respect and the period before AGI arrives, on his newest numbers, are now the same few years.

Hassabis is one of seven honourees in this year’s WSJ Magazine Innovator Awards, a programme that began in 2011; the magazine will name the rest over the coming two weeks, ahead of the awards event in New York on October 29.

Each time Hassabis moves AGI nearer, he also shortens the grace period he wants before anyone writes the rules. One more update nearer, and the grace period ends before the first rule is written.

Tags: Dario Amodei, Demis Hassabis, Google DeepMind, Alphabet, Artificial general intelligence