Geoffrey Hinton doesn't hold back on the future of AI

I'm slightly sick, so I've got an extra jacket on because I'm cold. But I will try to answer your questions anyway.

Thank you.

Slightly embarrassing but very nice.

It blows me away too. It's changed dramatically.

So I was working in an American university, Carnegie Mellon University, and research was going very well there, but I was somewhat unhappy with Reagan's policies towards Nicaragua. I thought mining the harbors in Nicaragua was a bad thing to do. Most of my colleagues in Pittsburgh thought it was perfectly reasonable because it was the American hemisphere and they should own it. I thought this was terrible. Little did I know.

Well, there was someone called Stuart Russell who was raising alarms well before that. So I was a fairly late comer to AI safety. Many of my colleagues have been thinking about it for more than 10 years. I was lucky in that people paid attention to what I said but I was a latecomer to it.

Yes. So very recently for example we've seen that one of the big tech companies actually did have some moral principles and sort of fairly mild moral principles. They didn't want their AI being used for mass surveillance of Americans. Just of Americans as far as I can tell. And they also didn't want it being used for weapons that decided by themselves to kill people. Which seems sort of... it's hard to argue with that. The defense department, that nice man Peter Hegseth, wanted to be able to use them however they liked and Anthropic said no. And so they've now gone after Anthropic. OpenAI, I think on the Thursday OpenAI said they stood with Anthropic and on the Friday they took their business. So I've now stopped using ChatGPT. I love ChatGPT. But I'm moving over to Claude because Sam Altman's moral flexibility is a bit too much for me.

Yes. I was just giving a lecture in one of the hospitals. I wasn't thinking that it was going to a worldwide public. And AI had got much better at interpreting what was in images. It was getting better rapidly. And it seemed to me obvious that just for interpreting medical images, AI would be better than people in about 5 years. That was overenthusiastic. It's about comparable with people at many images now. Better than people at some images. Quite a few in fact. But what happened was instead of AI just replacing people at the interpretation of images, people now work with AIs. They have many many different AI systems that have been approved for image interpretation and I think there's a lot more image interpretation going on. So there were two things I got wrong. Actually there were three things I got wrong. One, I got the time scale wrong by about a factor of two, maybe three. If you're, you know, factor of two or three, you know, physicists don't worry about things like that. I also got the conservativeness of the medical profession wrong. It's hard to overstate how conservative they are. So for example, if AI were to kill somebody by failing to diagnose a cancer, they think that was terrible. If a few hundred people die because they didn't use AI to diagnose the cancer, they ignore that. That doesn't count. So, I think it's a very asymmetrical decision they're making. They should think about people who are dying because they're not using AI. Then the third thing I got wrong was that healthcare is an elastic market. So in many markets when you get AI to do things as well as people or better than people will lose their jobs. But in healthcare we could absorb endless amounts of it particularly when we're old. I could use 10 full-time doctors just myself. But when we get AI, we'll get a lot more healthcare. And so I don't think it'll put people out of work. If you made doctors 10 times more efficient, we just all get 10 times more healthcare. And that would be great. So you have to decide whether the market's elastic or not.

I think quite a lot of it depends on politics. There's one case I know I heard about recently where a company developed a system that can you take a photograph with your iPhone or your Android and you give it to their system and it tells you whether that blotch on your skin is a melanoma or whatever. It tells you what it is. And it's as good as good people at doing that. And so it can act as a front end to decide whether you should go to a dermatologist. It can also be used in hospitals. The company couldn't... the company's dissolved now. It's in limbo with debts because they couldn't get it used in Ontario because the Ontario government wouldn't make a health code for it. So hospitals wouldn't buy it because there's no way they could charge three to five dollars for using it to look at something and say what it was. So there's quite a lot of that goes on I think.

Yeah.

Yes. Education. So we can all do with more education. I could do with learning some physics. It's slightly embarrassing that I don't know much. So in education, I think it's going to be great. It's slowly... we don't, it's not really good at being a tutor yet. It's getting there, though.

So what we know is that if you have a private tutor, you learn about twice as fast as learning in a classroom because quite a lot of what goes on in the classroom is a teacher in broadcast mode. The teacher is telling you... you don't have one to one, you can't afford it, right? So a teacher is teaching 20 kids in a private school or 35 kids in a state school. And what the children are hearing is the answers to questions that they didn't just ask. Now, when I interact with a chatbot, as soon as I'm curious about something, I ask it and it tells me and I absorb the information because that's what I wanted to know. And a lot more education should be like that. It should be people following their curiosity. And teachers can't do that. Teachers can do that in one on one and it's very effective. But there aren't enough teachers to go around, but there will be enough AIs to go around. So, I think we'll find education, particularly for kids like me who have slight ADHD. If people are talking about something I'm not interested in, I find it very hard to attend. And I think AIs will make education much more efficient. It's going to be a while before they're really good at understanding what it is that the pupil is misunderstanding.

It's not quite there yet. It doesn't really understand fully what the pupil doesn't understand but in the end it will because it'll have had a lot of training data.

When students use pocket calculators, they're cheating.

I mean, the future is going to be people working with AIs. So, what teachers ought to be examining is how well the student can get something done with AI. At least for the next few years, things are going to be like that. It may be a bit later on AI is just going to take over, but for the next few years, it'll be people working with AI. And so, that's what's going to be how people are going to be employed. That's what they ought to be examining.

Yes.

So my belief is and the belief of most of the experts is that sometime within the next 20 years we'll get general purpose artificial intelligence that can do pretty much any intellectual job people do. Anything you do on a computer is going to be able to do better than a person. And things like interviewing people about something is going to be able to do that just fine.

Well, my livelihood used to be sort of coming up with scientific ideas and programming. That's all gone.

I don't know.

Do we have 80 minutes?

How about 15?

Okay, I will do my best. So your brain has a lot of brain cells in it, neurons, and you learn to do things. And the question is how do you learn to do things? How do you learn to make a network of brain cells perform some tricky task like for example taking an image and saying whether it contains a bird or not? And so roughly speaking and this is grossly simplified but this is the rough idea. You have layers of these brain cells. Neuroscientists call those different cortical areas. In artificial neural nets we call them layers and the early layers for doing vision detect little combinations of bright pixels. So pixels, a little piece of a digital image and the first stage when you're recognizing objects just detects whether there's little pieces of edge at particular locations in the image. You might think you could take the pixels. So each pixel has a brightness and that's what goes into the computer, the brightness of each pixel. You might think you could connect those directly to other neurons that say whether there's a bird there. So suppose you have a very bright pixel. Should that say there's a bird or should it say there's not a bird? Well, the problem is that there's white birds and there's black birds. If all birds were white and everything else was black, when you get a bright pixel, you should say that makes it more likely there's a bird there. But birds can be white or black. So the brightness of a pixel doesn't really tell you much about whether there's a bird, but the brightness of a pixel can tell you about whether there's a little piece of edge there. If you imagine a row of... I need to wave my right hand otherwise I'm in danger of waving the microphone, but that won't work. No. So imagine a little row of pixels like this. Sorry, a little column of pixels like this and a little column of pixels next to it like this. If all of these are bright and all of these are dim, then you've got an edge there. And you can make a brain cell that detects that, a neuron that detects that. It can have big positive connection strengths coming from the three pixels here and big negative connection strengths coming from there. So if these are bright and these are bright, it gets positive input from here and it gets negative input from there and they cancel out. Nothing happens. But if these are bright, it gets positive input from there. And if these ones which would have given it negative input are dim, it doesn't get any input from there because they're dim. And so it gets very excited and it says, 'Hey, there's an edge here.' So it's quite easy to see how you could wire up a brain cell to detect a little piece of edge in the image. And so that's a good start. We have these neurons that look all over the image, lots and lots of neurons, gazillions of them, detecting little bits of edge in different positions, in different orientations, and also different scales because you can have a big fuzzy edge of a cloud, for example. That's the first layer. The next layer is now going to find little combinations of edges. So you might, for example, have a brain cell that looks for three edges like this and this and this and three more edges like this and this and this. And if it sees these three edges and these three edges all active, it says, 'Hey, I found what I like.' And it does that by having big positive connection strengths coming from these three neurons that detect those three edges and these three neurons that detect those three edges. When it sees that combination, it says, 'Hey, I found what I liked.' And obviously what it likes is a potential beak. It could be lots of other things. It could be an arrow head. It could be all sorts of things, but it's a candidate for a beak. And you might have something else in that layer that detects a little circle of edges. That's a candidate for an eye. Then the next layer up, you might have something that looks at a beak. It finds that there's a beak somewhere and it finds in the right relative location there's an eye. So maybe there's a sort of beak here and there's an eye here and a neuron in the layer higher up sees the beak and it sees the eye and it says, 'Hey, there might be a bird's head there.' And so now you can see how I might make a neuron that would detect a bird's head. You'd need many different neurons for different shapes of bird's head and so on. And then at the very next layer up, maybe you got bored by now. So you hook those things up, things that detect a bird's head or the talons of a bird or the wing of a bird. You hook them all up to a neuron that says if I find a bunch of those things, I'll say there's a bird there. Okay, so I just hand-designed a little neural net that would be okay at detecting birds in images if I was very patient and made lots and lots of these detectors. Little detectors for edge in the first layer, detectors for little combinations of edge in the next layer, detectors for combinations of the combinations in the next layer, and finally an object detector. Now, the real system's got more layers than that, but that gives you a sort of rough idea how it works, and you can see how you might wire it all up by hand, but that would take much too long. So, here's another thing you could do. You could start with your layers, and you could put in random weights. Instead of saying for my edge detector, I want three big positive weights coming from these three and three big negative weights from these three. And then I'll be able to detect an edge that's brighter here than there, I'll just put in random weights, small positive and negative numbers. And so my neural net will just sort of do random things. It won't do anything sensible. And I'll have these layers and it'll have some output for bird, an output for cat, an output for dog, and so on. And when I put in an image of a bird, it'll say, well, it's a little bit a bird and it's a little bit a cat and it's a little bit of dog and I could ask the following question. Could I change the strengths of just one of the connections in the network so that when I put in that particular bird, it's a little bit more sure it's a bird and a little bit less sure it's a cat. So I could go around fiddling with all the connection strengths and I could change the connection strength to say what did that do to whether it thinks it's a bird. And if I had astronomical time, if I had billions of years, I could fiddle with each connection strength, see if it made it more likely to get the right answer. And if I make a change and it gets more likely to get the right answer, I keep that change. If I make a change and it gets less likely to get the right answer, I probably make the opposite change. That's a kind of mutation method. That's a bit like random evolution. It would work in the end, but it would take forever. Because these networks have billions and billions of these connection strengths and you have to change each connection strength many times and you have to show it a lot of images to know whether increasing the weight or decreasing the weight, that's the strength of the connection helps or hurts. So that would take forever. Do I have five more minutes?

Okay. So, oh, in that case, I've got 80 minutes. So, what I've shown you is what a network might look like for detecting a bird and how you might get an evolutionary approach to build that network starting from random weights and just changing the weights by making little random changes and seeing if they help. And if you go on like forever, that'll work, but it'll take much too long. What you'd like to do is put in an image of a bird. It'll say it's sort of 10% a bird, 10% a cat, 10% a dog. And you want it to say 11% a bird, and 9% a cat. And what you'd like to know is is there a way of figuring out for every single weight in the network, all 100 billion weights, whether I should increase them a little bit or decrease them a little bit to make it work better. And there is a way of doing that. It's called back propagation. It was invented by many different people. A colleague of mine invented it in San Diego in 1981. And we did a lot of work on it and showed that it really worked very well. And the way it works is you look at the difference between what you got, which is 10% it was a bird, and what you wanted, which was 100% it was a bird. And you take that difference, and think of that difference as a piece of elastic. So, think of the output as a height. 10% is a bird, and you want it to be 100% is a bird. So, think of a little piece of elastic attaching the 10% to the 100%. And that's trying to pull the 10% up. But the 10% can't move because you put in the image of the bird and the weights determined the activations of all the feature detectors in all the layers and that determined that it was 10% a bird and to make the 10% move you'd have to change the weights. So you take this piece of elastic that would like to pull the 10% up to 100% and you say what if I took the force pulling, transmitted it backwards through the network to the weights that go into that bird unit. So if I want it to go up and there's a weight going into the bird decision unit and it's coming from a highly active neuron, I know that if I make that weight a bit stronger, the 10% will go up to 11%. So when I take this force pulling it upwards, I can transmit that back to the weights and it'll change the weights so as to make it more likely to say bird. But it'll also do something else. I can transmit the force backwards not just to the weights coming into the bird unit. I can transmit it back to the feature detectors in the layer before. Remember we had a feature detector that detected the head of a bird. Well suppose it is saying sort of maybe there's a head of a bird and you want it to be and it is a bird and you want the bird output to be stronger. Well if you made that feature detector more active it would be more confident it was a bird. So we transmit this force from the elastic back through the connection to the activity of this feature detector for the head of a bird and say you should be more active and that's like a piece of elastic pulling on that feature detector. So we can take the elastic pulling on the outputs and convert it into a piece of elastic pulling on the feature detectors. And that solved the fundamental problem of neural networks which is how do you decide for all these neurons in the intermediate layers whether they should be more active or less active? Well, you do it by taking the forces pulling on the output transmitting them backwards and seeing how they pull on these intermediate feature detectors. That's called back propagation. Now, if you still did calculus in grade 13 in Ontario, you'd know that this was the chain rule of differentiation, but unfortunately, you have to go to university to learn that. But that's how these things learn.

Yes. I'll try and justify that now.