DeepMind's Hassabis Says AGI Could Arrive as Soon as 2029

Demis Hassabis, who runs Google DeepMind, used the days after Google's I/O conference to sharpen a forecast he has made before. He still expects artificial general intelligence, a system that can match human ability across most cognitive work, to arrive around 2030. What changed is the edge of his range. In remarks reported by Axios, he said 2029 now looks plausible, and he did not seem to be saying it to comfort anyone.

Hassabis described the current wave of capable AI agents as a kind of practice run, a chance for society to feel out what more powerful systems will demand before those systems actually show up. His point was less about the agents themselves and more about the short runway. If the serious version is three or four years out, the time to build institutions, laws, and economic cushions is now, not after it lands.

The part that gives even optimists pause

The most pointed comment was about recursive self-improvement, the idea of an AI system capable of meaningfully speeding up its own development. "All the leading labs are quite focused on that," Hassabis said, before adding the obvious caveat: "There are also risks with that type of system." Coming from someone whose lab is one of those leaders, that is not a throwaway line. It is a quiet acknowledgment that the field is actively working toward the capability that worries safety researchers most.

Hassabis was candid that his word choice was deliberate. He said he leaned on language that was, in his own description, a little provocative, precisely to provoke more urgency among governments, economists, and the public. It is an unusual move for a lab chief, who would normally have every commercial reason to soothe rather than alarm. The fact that he chose the other direction is itself worth noting.

How much to read into a date

A caution is in order. Predicted arrival dates for AGI have a long history of slipping, and definitions of the term vary enough that two experts can agree on 2029 while picturing entirely different things. Hassabis knows this, which is part of why the specific year matters less than the direction of his argument. The signal is not "mark your calendar for 2029." It is that the people closest to the work have stopped treating transformative AI as a distant abstraction and started treating it as a planning problem with a deadline.

That framing lines up with moves we have covered recently, from OpenAI publishing a governance framework for its frontier models to government labs probing what today's systems can already do. The labs are not waiting for AGI to start arguing about how to handle it. Hassabis is telling everyone else to do the same.