Google DeepMind CEO Demis Hassabis has established a rigorous benchmark for artificial general intelligence that far exceeds current industry standards. He argues that true AGI requires AI to replicate creative leaps, such as deriving special relativity from pre-1911 knowledge, rather than merely mimicking or regurgitating existing data.
Hassabis explicitly stated that recent milestones, including AlphaFold’s Nobel Prize-winning protein folding achievements and solutions to complex mathematical Erdős problems, do not constitute AGI. These systems operate within defined problem spaces, whereas true general intelligence must create entirely new paradigms.
This distinction sharply contrasts with competitors like OpenAI, which often define AGI by economic output potential. By setting such a high bar, Hassabis implies that most current claims of AGI are premature.
Regarding timelines, Hassabis has adjusted his projections, moving from an initial estimate of three to five years in early 2025 to a refined target of around 2030, plus or minus one year. This shift underscores the complexity of achieving genuine cognitive autonomy in machine learning systems.