Hassabis Says AGI Is Two Years Away. His Own Lab Says It Can’t Jump Yet

In October 2026, Demis Hassabis moved the AGI timeline to two to three years.

He had changed it before. Early 2025: five to ten years. April 2026: around 2030. October 2026: two years. For a cautious researcher who has shaped the current AI agenda, cutting his own forecast by two-thirds in less than two years matters more than the number itself. Hassabis does not usually change his language to raise money or win headlines. When he does, he has probably seen something in the lab that forced a recalibration.

His definition of AGI has stayed the same.

He has described a thought experiment. Take an AI and cut off its knowledge at 1901. Give it what humanity knew then. See whether it can independently reach Einstein’s 1905 breakthrough, special relativity.

The test is harsh because of what it does not ask. It does not ask the model to restate relativity. It does not ask it to search inside a problem someone else has defined. It asks whether the model, with only 1901 physics, can propose a new conceptual framework. Hassabis has also separated two abilities. One is solving a conjecture that already exists. The other is proposing a valuable new conjecture. The first finds a path on a map. The second draws a map that did not exist. That distance separates “smart” from “capable of original discovery.”

At almost the same time Hassabis set out this standard, his own lab published a paper arguing that current large language models cannot pass it.

The paper by Google DeepMind researcher Tom Zahavy has a blunt title: LLMs can’t jump. It borrows a model of discovery Einstein sketched. Experience leads to an intuitive jump. The jump produces new axioms. Deduction follows, then verification. Using Charles Sanders Peirce’s categories, the paper argues that LLMs are strong at induction and deduction. They extract patterns from data. They derive results once axioms and rules are given. Scientific discovery, however, depends on abduction. That means inventing an explanatory framework to make sense of an inconsistency. Abduction offers no guarantee. It may lack data. It gives a hypothesis worth testing.

The paper uses general relativity as a case study. When Einstein proposed it, Newtonian mechanics was not facing a large data error. The anomalous precession of Mercury’s perihelion existed. Many physicists did not think Newton’s law was wrong. They looked for another planet near the Sun. A hypothetical body called Vulcan. From an optimization standpoint, Newton’s theory had low loss. Adding a hidden planet was a local patch. Rewriting absolute time, Euclidean space, and gravitational action into curved spacetime would introduce a vast and unfamiliar mathematical structure. A system built to minimize short-term prediction error would likely choose the patch. The paper uses this case to reject the idea that creativity is compression. Major paradigm shifts often do not grow from an error gradient. They happen where no supervisory signal points the way.

General relativity also gained decisive experimental support after it was proposed. Einstein did not start with a huge anomaly dataset or a falling training loss. He felt an inconsistency in theory. Special relativity had rewritten time and space. Electromagnetism described action through continuous fields. Newtonian gravity still relied on action at a distance and a privileged inertial frame. The inconsistency felt wrong, but he could not yet name it. That is the intuition problem and the jump problem. More reasoning length, more sampling, and more verification can produce a stronger searcher. They do not produce a system that redefines the problem.

The paper blames the gap on a lack of embodied simulation. Einstein used physical intuitions, such as the elevator thought experiment, to anchor abstract symbols. A text-only model lacks a route from lived sensory experience to new axioms. The paper argues for physically consistent multimodal world models with action control and counterfactual intervention. If AGI is to achieve Einstein-style discovery, it will likely need a new architecture, not a bigger version of the current one.

That creates a contrast worth facing. Hassabis uses the Einstein test as the ultimate AGI yardstick. His own lab’s paper argues that current models cannot pass it. A charitable reading says the lab breakthrough Hassabis has seen points to a new architecture beyond plain-text LLMs. World models, embodied systems, or something not yet public. If “within two years” means Einstein-test discovery on the current route, then LLMs can’t jump is a barrier that cannot be ignored. The paper does not say models are too small or data is insufficient. It says this kind of system lacks a mechanism for jumping from experience to new axioms. More compute will not supply that mechanism.

Hassabis is not alone in moving the date. Dario Amodei of Anthropic argues that a closed loop is forming and the timeline may be one to two years. Yann LeCun argues that scaling current Transformer systems cannot reach AGI. Andrew Ng says AGI is “not possible” in 2026 and that the term has largely become marketing. These voices show that the AGI timeline is not a purely technical prediction. It depends on the definition and on the route you believe will cross it first.

Hassabis describes the present as “Jagged Intelligence.” A model can win a gold medal at the International Mathematical Olympiad and still make a basic logical error. It can approach human level in code generation and mathematical proof, where feedback is clear. It can look clumsy where a problem must be redefined. Capability grows unevenly. It jumps in some dimensions and lags in others.

A more accurate reading of “within two years” may be this. In verifiable, searchable domains with clear feedback, progress can be faster than expected. Mathematical proof, code generation, and specific scientific computation have loss functions and verifiers. Models can improve through search and sampling. The abductive ability the Einstein test demands still faces constraints under current architectures. Two years is short. It makes the forecast worth tracking. The jump described in LLMs can’t jump remains the largest question mark.

If Hassabis is right, the breakthrough will likely come at the architecture level, not from a longer extension of the current LLM line. The marker will not be how many questions the model can answer. It will be whether it can ask a question we never thought to ask.

Before that, the timeline matters less than the crack behind it. A researcher measures AGI with his lab’s theoretical standard. His lab’s paper says the current ruler cannot reach the mark. Moving the forecast from five years to two does not close that gap. A new architecture or a new definition will. Two years from now, we may know which arrives first.