Demis Hassabis: Why AGI is Bigger than the Industrial Revolution & Where Are The Bottlenecks in AI

An AI system can excel at difficult questions while still failing simple variations of them. Google DeepMind co-founder Demis Hassabis identifies this inconsistency as one obstacle to general intelligence. Hassabis defines general intelligence as a system with all the cognitive capabilities of the human mind, so isolated impressive results are insufficient.

Hassabis also points to systems' inability to learn after training is finished and their weakness at planning over long time horizons. These are listed alongside scaling, which Hassabis says still brings substantial returns, though smaller than earlier. This gives a concrete way to interpret progress toward general intelligence: look for dependable abilities across situations, alongside improvements in continual learning and planning. Watch the account of the bottlenecks.