Hassabis Shortens AGI Forecast to Two to Three Years, Backs FINRA-Style Board

Demis Hassabis, the Nobel laureate who co-founded DeepMind and now chairs Google DeepMind while serving as Alphabet's chief scientist, told the Wall Street Journal this week that his estimate for when artificial general intelligence will arrive has compressed again—to somewhere between two and three years from now. The admission, confirmed across multiple technology outlets reporting on a WSJ profile published October 21, 2026, is notable not simply for the number but for its velocity: Hassabis's AGI forecast has moved approximately eight years closer across a single calendar year, and it now closely matches the public estimate of his most direct competitor, Anthropic's Dario Amodei.

A Timeline That Has Compressed in Twelve Months

The arc of Hassabis's 2026 estimates follows a consistent direction. At Davos in January, he described a 50% probability of AGI by 2030—a five-to-ten-year range. At India's AI Impact Summit in February, he said AGI was within five years. In April, speaking on the 20VC podcast, he said there was "a very good chance" of AGI within five years. At Google I/O in May, the phrase had shortened to "just a few years away," implying roughly 2029 or 2030. By July, when he published a detailed governance manifesto, the phrase had become "a few short years." The October WSJ interview distills this to two to three years—placing the implied arrival at 2028 or 2029.

The phrase "two to three years" is the Wall Street Journal's framing of Hassabis's remarks rather than a confirmed verbatim quotation; the original interview is behind a paywall. What is directly attributed across multiple secondaries is the sentiment: "It's going to happen soon. My estimate keeps moving closer to the present."

Hassabis is no longer Google DeepMind's chief executive. He stepped down from that operating role on August 5, 2026, and now chairs the organization while holding the title of chief scientist at Alphabet. He also continues as CEO of Isomorphic Labs, the Alphabet-backed AI drug-discovery spinoff built on the AlphaFold technology that contributed to his 2024 Nobel Prize in Chemistry. His current vantage point is broader and more strategic than a purely operational role—which lends his forward estimates a different character than a product announcement.

What Hassabis Means by AGI—and Why No Current System Passes His Test

The precision of any timeline claim depends entirely on the AGI definition being used, and Hassabis uses one of the most demanding available: a system capable of matching the full cognitive range of the human brain, not merely excelling at specific tasks or saturating existing benchmarks.

To make this concrete, he has proposed what he calls the Einstein Test. The idea is to train a system only on knowledge available before approximately 1901—prior to Einstein's entry into the scientific mainstream—and then ask whether it can independently derive the content of Einstein's 1905 annus mirabilis papers: special relativity, the quantization of light (photoelectric effect), the statistical mechanics of Brownian motion, and the relationship E=mc². An extended version uses a 1911 cutoff and targets the 1915 general theory of relativity.

Either version of the test demands that the model perform what researchers call out-of-distribution reasoning: not retrieving known results, not interpolating from training examples, but constructing genuinely novel theoretical frameworks by identifying and resolving conceptual tensions in existing physics. Current frontier models do none of this. They excel at in-context synthesis, pattern completion, formal mathematics within known formalisms, and structured reasoning over provided frameworks. What they do not do is look at Maxwell's electrodynamics and Newtonian mechanics, notice that they are mutually inconsistent at high velocities, and derive a new kinematics from first principles.

"It's clear today's systems couldn't do that," Hassabis said of the test earlier in 2026. The test is not without critics. Some researchers argue it is un-runnable in practice—any attempt to train on genuinely pre-1901 knowledge would require verifying the corpus, which is itself underdetermined. Others note that Einstein's derivations built heavily on prior work by Maxwell, Lorentz, and Boltzmann, all of which predates 1901, making the test's difficulty depend on what is included. But as a framework for distinguishing capability saturation from genuine creative intelligence, it is substantially more demanding than the benchmark suites that leading models routinely claim to top.

The Governance Proposal Was Born in July—Not This Interview

The oversight board that Hassabis endorsed in the WSJ profile was first proposed in a different context: a detailed manifesto he published on July 14, 2026, titled "A Framework for Frontier AI and the Dawning of a New Age." The two events—the governance proposal and the latest timeline compression—are related but distinct, and conflating them misrepresents both.

The proposed body is modeled on FINRA, the Financial Industry Regulatory Authority, which operates as a private, self-regulatory organization for the U.S. securities industry under Securities and Exchange Commission oversight. FINRA is funded by its member firms, sets rules, examines compliance, and can sanction members—all without requiring a dedicated federal bureaucracy to manage day-to-day operations. Hassabis's analogous structure would have frontier AI labs fund and staff a similar organization, operating under federal government supervision, with a board majority drawn from independent technical experts—Turing Award winners, AI safety researchers, open-source representatives, and government officials. Labs developing frontier models would submit them to this body for evaluation up to 30 days before release, with tests targeting cybersecurity risks, biosecurity risks, and deception capabilities.

The proposal emerged directly from the Anthropic Mythos episode of June 2026—the brief period during which the Trump administration imposed improvised export controls on Anthropic's Claude Mythos 5 and Fable 5 models, suspending access between June 12 and July 1 before the Department of Commerce lifted the controls. The episode demonstrated that uncoordinated executive intervention could disrupt frontier AI deployment with little technical review and minimal notice. Hassabis's position—that structured governance is preferable to ad-hoc political intervention—reflects a considered response to that experience rather than an abstract preference.

As of October 22, 2026, the proposed oversight body does not exist. In July, Hassabis expressed hope it could be operational by year-end 2026. That window is now fewer than ten weeks away.

Where the Two-Year Estimate Sits in the Wider Debate

The most significant aspect of Hassabis's latest estimate is not its absolute position but its convergence with Anthropic CEO Dario Amodei, who has publicly described AGI as one to three years away across several 2025 and 2026 statements. Two co-founders from two leading frontier AI laboratories, operating in direct competition, now publicly agree on approximately the same arrival window. That cross-institutional convergence carries more signal than either individual estimate in isolation.

The consensus is not universal. Meta's chief AI scientist, Yann LeCun, remains the most prominent public skeptic, arguing that Transformer architectures are fundamentally incapable of reaching human-level intelligence and that current hype is detached from genuine capability progress. Sam Altman of OpenAI has been more ambiguous, suggesting at points in 2026 that AGI may have already "whooshed by" in some functional sense—a framing that sidesteps the timeline debate by questioning whether the definition has become the real issue. Shane Legg, who co-founded DeepMind alongside Hassabis, previously estimated "minimal AGI" could arrive as early as 2028, a figure now bracketed by Hassabis's current range.

If the Hassabis-Amodei convergence on a two-to-three-year window is directionally correct, the implications for governance planning are concrete and urgent. The EU AI Act's frontier-model obligations are still phasing in. No U.S. federal legislation governing frontier AI systems currently exists. The FINRA-style oversight body Hassabis proposed in July—designed to fill precisely the gap these timelines expose—remains unbuilt.

The planning horizon implied by a 2028 AGI arrival is not a future consideration. It is a present one. Establishing any governance body, calibrating its evaluation criteria, and bringing frontier labs into voluntary compliance requires institutional lead time that the two-to-three-year window is already consuming. That is the operational argument behind a forecast that has continued to move closer with each public statement Hassabis makes.

Read more: Jack Clark moves AGI self-improvement forecast to 2027