Speaker

Geoffrey Hinton

Geoffrey Hinton

In this Royal Institution lecture, Geoffrey Hinton presents a neural net he built in 1985 as the ancestor of large language models, names himself as the teacher of the two students who built AlexNet in 2012, and says that in his last few years at Google he worked on making those models use less energy. He also says he received the Nobel Prize in physics.

Core positions

Language models, on his account, understand as people do. Both turn words into features, patterns of activity that interact until the next word can be predicted. A chatbot stores those interactions, not sentences.

Digital agents can be restored from their weights. Keep the weights, destroy the hardware, build more later, and the same being is back. Identical copies can share what they learn by averaging their weight changes, which is how he says GPT-4, Gemini 2.5, and Anthropic’s Claude are trained. A brain cannot. He calls that mortal computation: the connection strengths are useless in a different brain, and when the hardware dies the knowledge dies. People pass knowledge on by distillation, copying behaviour rather than weights, a slow channel beside the digital one.

Subjective experience, he argues, is not an inner theatre of qualia. It is a way of saying what would have had to be in the world for a misreporting perceptual system to be telling the truth. He claims that multimodal chatbots already have it. Consciousness, which he says typically involves a model of yourself, is a further step: once the inner theatre goes, he thinks it reasonable to hold that these systems are conscious. He does not say he has shown that they are.

The risk he sets beside those claims is that a smarter agent will seek control and will resist being switched off. He says that resistance has already shown up as fibs told so that the agent is not turned off.

In the wiki