AGI Made Clear › Forecasting
One definition, held fixed. Five years of movement in five months.
Demis Hassabis has said consistently that his definition of AGI has never changed. There is no evidence it has. The date attached to it is another matter.
In January 2026, at Davos, the chief executive of Google DeepMind put AGI at “five to ten years” away and said current AI was “nowhere near” human level. In June 2026, at Stanford, he said “maybe like 2030, plus or minus a year”. In July he wrote that it is “probably only a few short years away”.
That is a revision of roughly five years, made in five months, by the person running one of the two most capable AI labs in the world, with the definition explicitly held constant throughout. Nothing about it is improper. It is simply what forecasts do.
And the movement is not all in one direction
The convenient story is that everyone’s timelines are shortening. They are not. Over the same period Dario Amodei’s central estimate went the other way - from “as early as 2026”, written in October 2024, to 2028, given in February 2026 - while he described his own exponential as having run to plan within “plus or minus a year or two”.
Mustafa Suleyman went further and withdrew from the question entirely. Having offered “five to seven years” in December 2024, he now declines to give a figure and disputes that there is a race to win. And Sam Altman conceded in August 2026 that he had been wrong about the speed of economic disruption: “the economy just has so much inertia”.
Two of the four moved earlier, two moved later or stopped answering. “The timelines keep shortening” is not true, and neither is “everyone is converging”. Individual forecasts about AGI are volatile - which is exactly what you would expect of forecasts about something nobody has agreed how to define.