Hinton quits Google: when the pioneer changed his mind

What he said and when

Geoffrey Hinton, who shared the 2018 Turing Award with Yoshua Bengio and Yann LeCun for the work that made deep learning practical, announced at the start of May that he was leaving Google. His stated reasons were plain. He told MIT Technology Review that he was getting too old to do technical work that requires remembering lots of details, that he wanted to do more philosophical work on AI safety, and that he wanted to talk about safety issues without having to worry about how it interacts with Google's business. The New York Times interview published on May 1 reported that a part of him now regrets his life's work.

The part of the interview that matters for researchers is the description of what changed. Hinton said he had suddenly switched his views on whether these things are going to be more intelligent than us, and called it a sudden flip. The trigger was GPT-4 and the behaviour of large models on few-shot learning and reasoning tasks. On May 30 he was among the signatories of a 22-word statement organised by the Center for AI Safety, which reads in full: mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.

The argument behind the flip

Hinton's reasoning is more specific than the headlines suggest, and it is worth stating because it can be argued with. The brain has around 100 trillion connections. The largest language models have up to half a trillion, a trillion at most. Yet in his assessment GPT-4 knows hundreds of times more than any single person does. His conclusion is that these systems have a much better learning algorithm than we do, and that they are a completely different form of intelligence, and a better one.

That is an empirical claim dressed as a philosophical one. Connection counts are not a clean measure of capacity, a model's stored facts are not the same thing as a person's knowledge, and the comparison ignores how much of the model's advantage comes from reading a large fraction of the written record. But the shape of the argument is a researcher's argument, made from a quantity he has thought about for forty years, and the honest response is to contest the numbers rather than the standing of the person making it.

He was also careful about what models get wrong. Asked about hallucination, he said confabulation is a signature of human memory and that these models are doing something just like people. He was refusing to treat the failure mode as evidence that nothing is going on inside, which is a different position from excusing it.

What a public revision buys

The field has always had people warning about risk. What is new this month is that the warning comes from someone whose reputation was built on making these systems work, and who stayed inside the largest lab until he decided he could not say what he thought from there. That combination changes the audience. Politicians and journalists who dismissed the risk conversation as coming from outsiders now have to explain why the person who trained their engineers is saying the same thing.

It also sets a norm for how to change your mind. Hinton did not quietly adjust his estimates in a footnote. He said what he used to believe, said what he believes now, and named the observation that moved him. Whether or not the new belief is right, that is the correct procedure, and we would like more of it from senior people whose views have shifted since last November and who have said nothing.

Where we think the argument is weakest

The extinction statement is deliberately short, and its brevity is a feature for signatories and a problem for readers. It says nothing about mechanism, timeline or probability. Signing it commits you to the view that the risk deserves global priority, which is compatible with believing the probability is very small and the stakes are very large. The signatory list includes the heads of OpenAI, Anthropic and Google DeepMind alongside Hinton and Bengio, which critics have taken as either a conflict of interest or a bid for regulatory attention.

For an experimentalist, the more useful document is the interview, because it contains claims that could be tested. Does capability per parameter actually favour digital systems in a way that persists across scale? Does few-shot reasoning in current models reflect anything a psychologist would recognise as reasoning? Those are questions with answers. The statement is a political act, and it should be judged as one.

What we would want next

We would like to see Hinton's brain-versus-model comparison written up with the assumptions made explicit, so that people who disagree can say which number they disagree with. We would like the labs whose leaders signed the statement to say what they are going to do differently on Monday morning. And we would like the field to notice that the most senior person in it has just modelled the behaviour we ask of every graduate student, which is to update on evidence and say so out loud. Whether his updated view survives the next two years of evidence is exactly the kind of question the rest of us are here to work on.