Hassabis Pulls His AGI Forecast Forward to 2029

Demis Hassabis has moved his clock forward. Speaking at Google's developer conference in late May, the DeepMind chief executive said he still broadly expects artificial general intelligence around 2030, but now sees 2029 as a genuine possibility. He described the present moment as the "foothills of the singularity," and argued that society has only a few years left to get ready. Axios and Computerworld both reported the remarks.

Hassabis offered a useful way to think about the systems we already have. Today's AI agents, he said, are best understood as a "practice run" for something far more capable. The interesting part of his message was not the optimism, which is his stock in trade, but the warning attached to it. He said governments, economists, and the wider public are not taking the prospect seriously enough, and that he has been discussing safety measures with leaders at other top labs, though he declined to give specifics.

What a date like 2029 does and does not tell you

A forecast from the person running one of the three leading labs carries weight, and it deserves a careful read rather than a headline. Two things are worth holding in mind.

First, there is still no agreed definition of AGI. Hassabis tends to mean a system with broad, human-level cognitive ability across most tasks. Others set the bar at economic milestones, or at specific benchmarks, or at something fuzzier they will recognize when they see it. A prediction is only as precise as the thing it predicts, and "AGI by 2029" can mean very different futures depending on whose definition you use.

Second, timelines from inside the industry have a mixed record. The same confidence that the field has found the right technical path has been voiced before, at moments that later looked premature. None of that makes Hassabis wrong. It is a reason to weigh the claim against evidence rather than against the credentials of the person making it.

The gap that actually matters

Strip away the date and the more durable point is about preparation. Whether transformative AI arrives in 2029, 2032, or later, the institutions meant to absorb it, labor markets, regulators, schools, are not moving at the pace of the labs building it. That gap is real regardless of which year proves correct.

It is also why the louder debate about whether progress has stalled misses the mark. We recently examined the claim that AI has hit a wall, and Hassabis is firmly on the other side of it. The honest position sits between the two camps. Progress is clearly continuing, the endpoint is genuinely uncertain, and the readiness gap is worth worrying about whichever way the technology breaks. Yann LeCun's recent work on world models that reason about the physical world is a reminder that researchers still disagree, sharply, about what the path to general intelligence even looks like.