Geoffrey Hinton, perhaps the most important person in the recent history of artificial intelligence, recently sent me a video of Snoop Dogg.
In the clip of a discussion panel, the rapper expresses profane amazement at how artificial intelligence software, such as ChatGPT, can now hold a coherent and meaningful conversation.
“Then I heard the old dude that created AI saying, ‘This is not safe ’cause the AIs got their own mind and these motherfuckers gonna start doing their own shit,’” Snoop says. “And I’m like, ‘Is we in a fucking movie right now or what?’”
The “old dude” is, of course, Hinton. He didn’t create AI exactly, but has played a major role in developing the artificial neural network foundations of today’s most powerful AI programs, including ChatGPT, the chatbot that has sparked widespread debate about how rapidly machine intelligence is progressing.
“Snoop gets it,” Hinton tells me over Zoom from his home in London. The researcher recently left Google so that he could more freely call attention to the risks posed by intelligent machines. Hinton says AI is advancing more quickly than he and other experts expected, meaning there is an urgent need to ensure that humanity can contain and manage it. He is most concerned about near-term risks such as more sophisticated, AI-generated disinformation campaigns, but he also believes the long-term problems could be so serious that we need to start worrying about them now.
When asked what triggered his newfound alarm about the technology he has spent his life working on, Hinton points to two recent flashes of insight.
One was a revelatory interaction with a powerful new AI system—in his case, Google’s AI language model PaLM, which is similar to the model behind ChatGPT, and which the company made accessible via an API in March. A few months ago, Hinton says he asked the model to explain a joke that he had just made up—he doesn’t recall the specific quip—and was astonished to get a response that clearly explained what made it funny. “I’d been telling people for years that it's gonna be a long time before AI can tell you why jokes are funny,” he says. “It was a kind of litmus test.”
Hinton’s second sobering realization was that his previous belief that software needed to become much more complex—akin to the human brain—to become significantly more capable was probably wrong. PaLM is a large program, but its complexity pales in comparison to the brain’s, and yet it could perform the kind of reasoning that humans take a lifetime to attain.
Hinton concluded that as AI algorithms become larger, they might outstrip their human creators within a few years. “I used to think it would be 30 to 50 years from now,” he says. “Now I think it's more likely to be five to 20.”
Hinton isn’t the only person to have been shaken by the new capabilities that large language models such as PaLM or GPT-4 have begun demonstrating. Last month, a number of prominent AI researchers and others signed an open letter calling for a pause on the development of anything more powerful than currently exists. But since leaving Google, Hinton feels his views on whether the development of AI should continue have been misconstrued.