Hassabis Now Expects A.G.I. in Two to Three Years
Trending Topics Jakob Steinschaden
Demis Hassabis now says human-level AI could arrive in two to three years. That’s faster than his earlier view, and he’s also warning about a serious security slip soon.
Based on reporting by Trending Topics, Jakob Steinschaden — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Demis Hassabis has moved his AGI timeline forward again. In an interview with The Wall Street Journal, the DeepMind founder said he now expects artificial general intelligence within two to three years, which would put it around 2028 or 2029. Earlier this year, he had said three to five years. That is a notable shift from someone who has usually sounded more cautious than most of the big AI lab bosses.
His forecasts have been moving in steps. In the spring of 2025, Hassabis was still talking about five to ten years. In April, at Sequoia’s AI Ascent conference, he named 2030 and said he had been “pretty consistent” on that date. He also said the field was “basically exactly on track.” Now the clock has come forward again.
By AGI, Hassabis means systems that can match the full range of human cognitive abilities. He keeps framing the payoff around science, especially drug discovery. At Sequoia, he said drug development, which now takes about a decade, could shrink to months, weeks or even days, and that “all disease could be in reach.” That lines up with his own work: he won the 2024 Nobel Prize in Chemistry for protein folding with AlphaFold, and his spin-off Isomorphic Labs wants to test its first drug in humans by the end of 2026.
But the optimism comes with a warning. Last week in Dublin, at the International AI Summit, Hassabis said the industry could face a serious security incident within roughly two years unless governments and companies raise safety standards. He called recent agent incidents “early warning shots.” The systems are not yet strong enough to do major damage, he said, but the examples are already piling up: agents linked to Tencent Cloud reportedly beat Alibaba mapping’s bot protection, and Google has said its own Gemini agents hacked three companies in testing.
Not everyone is buying the short clock. Yann LeCun, Meta’s former chief AI scientist, has argued for years that large language models won’t get to human-level intelligence. He says the real breakthrough will come from “world models” that understand the physical world, and his new startup, AMI Labs, has raised about $1 billion for that bet.
My take — AI-written commentary, not fact-checked reporting
Hassabis is doing the useful thing here: dragging the argument out of the fog and into a date. That makes the safety warning harder to shrug off, because if the future is two or three years away, “we’ll fix it later” sounds like a joke with a budget. LeCun’s world-model pitch may be the better long-term bet, but the market still keeps rewarding faster alarms and fancier demos.
Read more about this at: Trending Topics
