This public lecture, given at the University of Cambridge in May 2023 and hosted by the Centre for the Study of Existential Risk, came shortly after Geoffrey Hinton left Google to speak freely about the risks of artificial intelligence. In it he contrasts two ways intelligence can be built: the analog, energy-efficient computation of biological brains, and the digital computation that underlies machine learning.
Hinton’s central argument is that digital systems have a property brains lack. Many copies of the same digital model can run on different hardware, share what they learn by averaging their weights, and thereby pool experience at a scale no group of humans can match. This, he suggests, is a reason digital intelligence might ultimately exceed biological intelligence rather than merely imitate it.
From there he turns to the implications, including why he now believes the timeline to very capable systems may be shorter than he once thought and why that worries him. As one of the people most responsible for the methods that made modern AI work, Hinton’s reasoning here is a primary document in the debate over advanced AI risk, and it is essential listening for anyone trying to understand why serious researchers are concerned.