This is Geoffrey Hinton’s Romanes Lecture, delivered at the Sheldonian Theatre and posted on the University of Oxford’s official channel. The annual Romanes Lecture is one of Oxford’s most prestigious public lectures, and Hinton, often called a godfather of AI, used it to lay out why he now takes the prospect of superintelligent machines seriously.
Hinton argues that digital intelligence has properties biological brains lack: it can be copied, run in parallel, and share what it learns across many instances, which he suggests could make it a fundamentally more powerful form of intelligence. He also discusses the risks, including the possibility that a sufficiently capable system could pursue self-preservation, and his revised view that human-level and beyond capability may arrive far sooner than he once thought.
For readers trying to understand the AI safety debate, this lecture is valuable precisely because of who is giving it. Hinton spent decades building the neural network methods at the heart of modern AI, and here he explains firsthand why a pioneer of the field came to publicly warn about its trajectory.