Demis Hassabis says the age of thinking machines is almost here, and the field cannot agree whether he is right
Demis Hassabis, Co-Founder and Chief Executive Officer of Google DeepMind, has set out his conviction that artificial general intelligence is close and that its arrival would rank among the most consequential events in the history of technology. The argument appeared on Tuesday in a personal essay published to X and Substack, “A Framework for Frontier AI and the Dawning of a New Age,” in which he described the present moment as a pivotal one in human history and AGI, a system exhibiting all the cognitive capabilities of the brain, as “probably only a few short years away.”
The essay is notable less for novelty than for the directness with which one of the field's most credentialed figures commits to a timeline many of his peers regard as premature. Hassabis, who shared the 2024 Nobel Prize in Chemistry for the protein-structure work behind AlphaFold, has spent his career arguing that AGI, built and deployed responsibly, would prove among the most beneficial technologies ever invented. What he offers now is a fuller account of the scale he has in mind and the reasoning that underpins it.
A claim about the technology, made in the language of wonder
Hassabis reaches for elemental imagery to convey what he believes is coming. The technology, he wrote, bears closer comparison to the discovery of electricity or fire than to the internet or mobile, and the field is standing “in the foothills of the singularity.” He put the human achievement in plainer terms as well. “We've essentially found a way to make sand think,” he wrote. “It's miraculous.” His estimate of the impact is deliberately vast, reaching, by his reckoning, “10x the Industrial Revolution at 10x the speed.”
The benefits he sketches follow from that scale. Hassabis expects AGI to accelerate drug discovery, help develop new clean energy sources and produce advanced materials, and he raises the prospect of an era in which resources cease to be the limiting factor for human progress. The framing is characteristic of a scientist who has repeatedly described the technology as a tool for accelerating discovery rather than an end in itself, and whose own laboratory has built its reputation on scientific applications more than consumer products. The wonder, in his account, is not rhetorical decoration but the reason the stakes are as high as he believes.
That conviction is not universally shared among the people building these systems, and the disagreement has been neither muted nor polite. Late last year, Hassabis was drawn into a public exchange with Yann LeCun, Meta's chief AI scientist, who dismissed the notion of general intelligence built on today's language models as complete nonsense and delusional, to which Hassabis responded that LeCun was just plain incorrect. The exchange captured a genuine scientific divide over whether current methods can reach general intelligence at all, or whether some further conceptual breakthrough is required first.
Others occupy the ground between them. Oriol Vinyals, a Gemini co-lead, has held that current models are strong in some areas while still lacking the capacity to truly innovate, a limitation he regards as unsolved. Richard Sutton, one of the founding figures of reinforcement learning, takes a similar view and has launched a startup to pursue the missing piece directly. Even inside DeepMind the timelines vary, with Shane Legg, Hassabis's own co-founder, putting a minimal form of AGI as early as 2028. The spread among serious researchers, from those who consider the concept ill-founded to those who expect it within a few years, is itself the clearest measure of how uncertain the science remains, and Hassabis is candid about that uncertainty. “Nobody in the world knows for sure what is going to happen from here, and even the experts disagree,” he wrote.
Why a scientist is talking about guardrails
The essay is not only a statement of belief about capability. Hassabis argues that the same uncertainty which makes the timeline contested is itself the reason for caution, and that when the stakes are this large, “proceeding with cautious optimism is the sensible and correct strategy.” He points to risks that grow more serious as systems become more capable, in cybersecurity, in biological threats, and in the challenge of maintaining control over increasingly agentic, self-improving models, and he argues that advances on the frontier are outpacing the field's understanding of the technology it is producing.
His answer is a proposal for a testing regime, set out in the same essay, under which the most capable models would be assessed against evolving benchmarks before release, with a new standards body modelled on the Financial Industry Regulatory Authority that polices Wall Street under government oversight. The mechanism would begin voluntarily and could later become a condition of deployment. The detail matters less, for a reader trying to understand the significance, than what it represents: a working scientist at the head of a major laboratory arguing that his own field's most powerful future models, DeepMind's included, should be subject to outside scrutiny before they reach the market.
That position places him alongside Dario Amodei, Chief Executive Officer of Anthropic, who has made his own case for binding oversight, though the two differ on the form it should take. Amodei has drawn the analogy to the Federal Aviation Administration, a government body with authority to block an unsafe model outright, where Hassabis favours an industry-funded structure answerable to government. The difference is real, but for the purposes of this story it sits downstream of the shared premise that unites them, which is that the technology is approaching a threshold serious enough to warrant the conversation at all.
What it means
The significance of the essay, for the wider sector, lies in who is making the claim and how far he is willing to commit to it. Predictions of imminent AGI are commonplace at the edges of the industry; they are rarer, and carry more weight, when they come from a Nobel laureate running one of the three leading laboratories, stated plainly and attached to a concrete argument about consequences. Hassabis is a figure who commands respect across the field's warring camps, and his willingness to put a few-years timeline in writing, alongside an admission that the experts disagree, gives the claim a different standing from the same words spoken by a less cautious voice.
For enterprise decision-makers, the practical value is in separating the two things the essay does. One is a contested scientific forecast about when general intelligence arrives, on which no consensus exists and none should be assumed. The other is a claim about capability that is already demonstrable in the cyber and biological risk domains Hassabis cites, and that does not depend on the AGI timeline being correct to matter. An organisation does not need to accept the foothills-of-the-singularity framing to take the nearer-term capability seriously, and the essay is most useful read as an argument for doing exactly that.
Hassabis closes on the register in which he began, writing that a post-scarcity world would raise questions of meaning, purpose and economic organisation that “cannot and should not be left to technologists alone.” Whether his central prediction proves right within a few years or a few decades, the argument he is pressing is that the window to shape the technology exists now, and that “what we collectively do now will determine how the next phase of civilisation unfolds.” Though the field remains divided on the time, the value of using the time, he has fewer opponents.
