Demis Hassabis Sees AGI Arriving Within Two to Three Years

Google DeepMind Chairman Demis Hassabis now estimates that artificial general intelligence could emerge within two to three years, shortening a timeline he put at three to five years earlier this year. The latest range points to around 2028 or 2029.

Hassabis’s proposed “Einstein test” would train a model only on knowledge available before 1901, then ask whether it could independently reproduce four breakthroughs Albert Einstein published in 1905. They covered the photoelectric effect, Brownian motion, special relativity and mass-energy equivalence.

The test’s design limits the model to knowledge that predates Einstein’s work. It would assess whether the model could independently reproduce those specific results, rather than simply rely on later physics knowledge. Hassabis has presented the test as a way to judge whether AI has reached the level of general intelligence he has in mind.

Hassabis describes AGI as a system with the full range of human cognitive abilities, including original thought. He has said he wants to use AGI to understand the nature of reality and try to derive the next layer of physics.

His timeline has shifted over the past year. Hassabis previously estimated that AGI was five to 10 years away. He later put it around 2030 or as early as 2029, before giving the latest estimate of two to three years.

The benchmark also sits within a debate over how to define AGI. One argument is that an AI matching people across the full range of cognitive tasks would already outperform humans in dozens of fields, making AGI a poor finish line.