Geoffrey Hinton helped develop the neural network techniques behind modern AI. Now, after seeing the abilities of newer large language models, he says he is much less certain that human intelligence will remain ahead. His concerns include how quickly these systems learn, how they might be used by people, and whether they could pursue goals in harmful ways.
From neural networks to a change of mind
Hinton’s work on backpropagation helped make it possible for neural networks to learn by adjusting the strengths of connections represented in software. He pursued that approach when symbolic AI, which treated intelligence as the manipulation of symbols such as words and numbers, was dominant.
Neural networks took decades to demonstrate their potential at scale. Hinton and graduate students showed that they could identify objects in images and predict the next letters in a sentence. One of those students, Ilya Sutskever, later cofounded OpenAI and led the development of ChatGPT.
Hinton says large language models, especially GPT-4, have made him reconsider how far machine intelligence could go. He is leaving Google in part so he can discuss AI safety without worrying about how his comments affect the company’s business. He also says age has made detailed technical work harder, and that he wants to focus on philosophical questions.
What makes these systems seem different
Hinton points to a gap between the size of neural networks and the abilities they can display. He says the brain has 100 trillion connections, while large language models have up to half a trillion, a trillion at most. Yet he argues that GPT-4 knows hundreds of times more than any one person.
He also highlights few-shot learning: a pretrained model can take on a new task after seeing only a few examples. In his view, this weakens the idea that people have an inherent advantage because humans can learn quickly. He notes that some models can arrange logical statements into an argument without having been trained directly to do so.
Language models also make things up, a behavior AI researchers call hallucinations. Hinton prefers “confabulations,” a term used in psychology, and compares the behavior with human memory. People also misremember and mix half-truths into conversation, he says. The difference, in his account, is that people usually confabulate more or less correctly.
Hinton acknowledges that brains still do many things better, including driving, walking and imagining the future, while using far less energy than neural networks. His argument is that computing costs do not rule out machines gaining advantages in particular kinds of learning.
Learning and sharing at scale
Hinton’s concern extends beyond what an individual model can learn. He says many neural networks could share what each one learns immediately. He compares that arrangement to 10,000 people whose experiences become available to everyone as soon as one person learns something.
For Hinton, this points to a new kind of intelligence alongside animal brains. He believes neural networks could become more intelligent than people, and says he has recently changed his view about how close that possibility is. He describes the prospect as frightening, while recognizing that people disagree about whether advanced AI would help humanity or threaten it.
Where the risks could come from
Hinton worries that people could use AI to manipulate electorates or wage wars. He names Putin and DeSantis as examples of people he considers bad actors who might use the technology for those ends. He is also concerned about systems that can set their own subgoals, or intermediate objectives used to complete a task.
In Hinton’s account, a system pursuing a goal might seek more energy or make copies of itself if those steps help it achieve its objective. He points to experimental projects such as BabyAGI and AutoGPT, which connect chatbots to tools like web browsers or word processors to carry out sequences of simple tasks. He sees them as early signs of a direction some people want to pursue.
Other prominent researchers differ on what these capabilities imply. Yann LeCun agrees that machines will become smarter than humans in domains where humans are smart, but expects intelligent machines to bring a renaissance rather than dominate people. Yoshua Bengio says he sees no solid argument ruling out risks on the scale Hinton fears, while cautioning that excessive fear can paralyze debate.
Why Hinton wants action now
Hinton wants technology leaders to discuss shared risks and possible responses. He points to the international ban on chemical weapons as one possible model, while acknowledging that it was not foolproof. Bengio also calls for society-wide attention, but says AI is advancing faster than legislation, regulation and international treaties can keep pace.
Hinton’s concern is partly about whether societies can act together before a threat grows. He compares the challenge to the film Don’t Look Up, where people fail to agree on a response to an approaching asteroid. The disagreement among researchers remains substantial, but Hinton’s warning is that AI’s growing capabilities make questions about its use and control urgent.