Full interview: "Godfather of artificial intelligence" talks impact and potential of AI

How would you describe this current moment in AI machine learning whatever we want to call it? I think it's a pivotal moment where GPT has shown that these great language models can do amazing things and the general public has suddenly realized, yes, we've done this because Microsoft released something and suddenly they realize things that people in big companies have been aware of for the last five years. Yeah, what did you think the first time you used GPT chat? Okay, I've used a lot of things that came before the chat. gpg which were quite similar, so chat GPD itself didn't really surprise me with gpt2, which was one of the earlier languages.

The brands surprised me and a Google model amazed me because it could actually explain why a joke was funny, oh really, yes, in natural language. I'll tell you, yeah, you tell a joke, not for all jokes, but for some of them he can tell you why it's funny, okay, and it seems really hard to say, he doesn't understand when he can tell you why a joke is funny, so if the GPT chat wasn't that surprising or impressive, were you surprised by the public reaction because the reaction was big? Yeah, I think everyone was a little surprised by how big the reaction was, which was the fastest kind of growth.

Yes, maybe we shouldn't have been surprised, but people researchers had gotten used to the fact that these things really worked. Yeah, you were famously like half a century ahead of the curve on this AI thing. Go ahead and correct me. go ahead not really not because there were two schools of thought in AI um there was conventional AI and then there were years but it was all about friends yeah they thought it was about reasoning and logic and then there were neural networks which won't. Call AI then, um, who thought it would be better to study biology because those were the only things that really worked and so conventional AI based their theories on reasoning and logic and we based our theories on the idea of that the connections between neurons change and that's how you learn and It turned out that in the long run, um, we came out triumphant, um, but in the short term it seemed a little hopeless.

Well, looking back, knowing what you know now, do you think there's anything you could have said then that would have convinced people that you could have said it? So, but it wouldn't have convinced people and what I might have said then is that the only reason neural networks didn't work very well in the 1980s was because the computers weren't fast enough and the data sets weren't fast enough. They weren't big. enough, but in the 80s the big problem was whether you could expect a large neural network with many neurons, computing nodes and connections between them that would learn simply by changing the strengths of the connections.

Could you expect that to just look at the data and with no kind of innate prior knowledge learn how to do things and people in mainstream AI thought it was completely ridiculous it sounds a little ridiculous it's a little ridiculous but it works and how did you know or why What did you intuit that would work because the The brain works because you have to explain how it is that we can do things and how it is that we can do things as if we did not evolve, like reading, reading is too recent for us to have had a significant evolutionary contribution, but we can learn to do it.

That and Math we can learn that, so there must be a way to learn in these neural networks yesterday. Nick Frost, who used to work with you, told us that you're not really that interested in creating AI, your main interest is simply understanding how the brain works. It works, yeah, I'd really like to understand how the brain works, obviously, if your failed theories about how the brain works lead to good technology, you cash in on that and get grants and stuff like that, but I'd really like to know how the brain works. . It works and I think there is currently a divergence between the

artificial

neural networks that are the basis of all this new AI and how the brain actually works.

I think they're taking different paths now, so we're still not doing it the right way. What I think is my personal opinion, but all the big models now use a technique called backpropagation that you helped popularize, popularize in the '80s, very good, um and I don't think that's what the brain does, explain why. , okay, there is a fundamental factor. difference between two different, there are two different paths to

intelligence

, so one path is a biological path where you have hardware that is a little unstable and an analogue, so what we have to do is communicate using natural language and also showing people how to do these imitations and stuff. like that, but instead of being able to communicate a hundred trillion numbers, we can only communicate what could be said in a sentence other than that many bits per second, yeah, and that's why we're really bad at communicating compared to these current computer models that They run on digital computers is almost infinite, they can do it, it is a band of communication with this huge yes, because they are exactly the same model, they are clones of the same model running on different computers and because of that they can see large amounts of data. because different computers can see different data and then they can combine what they learned more than any person could understand, much more than any person could ever be, and yet somehow we are smarter than them, still Okay, so they're like wise idiots, right?

GPT knows so much more than anyone, if you had a competition you know how much you know, I would just eliminate anyone, he was amazing at bar trivia, yeah he would be amazing, I could use him and he can do everything you can. Write poems, can you tell? um, they're not as good at reasoning, we're better at reasoning, we have to extract our knowledge from a lot less data, so we have a hundred trillion connections, most of which we learn, but we only live for a billion seconds, which it's not a lot of time, whereas things like GPT chat have been running for a lot longer to absorb all this data, but on a lot of different computers 1986 you post something in the wild, that's the idea that we're going to have a phrase of words and it will predict the last word yes that was the first language model that's basically what we're doing now yes and no 1986 was a long time ago why didn't people still say oh okay I think you're right oh because back then , if you were asking me how much data I trained that model on, I had a small, um simple, world of just family relationships, there were 112 possible sentences and I trained it on 104 of them and checked to see if the last eight were correct, okay?

As I would do it? He hit most of the last eight. Well, it worked better than the token AI. So it's just that computers weren't powerful enough at that time. The computers we have now are millions of times faster. They are parallel, but they can. do millions of times more proficiency, so I did a little calculation if I had taken the computer I had in 1986 and started learning something on it, it would still be running now and I wouldn't have gotten there, huh, and that's something that you now take a few seconds to know, did you know that's what was holding you back?

I didn't know. I think that might be what was holding us back, but people scoffed at the idea that the statement that well you know if I just had a much bigger computer and a lot more data it would all work and the reason why it wouldn't works now is because we don't have enough data in enough computing, which looks like some kind of lame excuse for the fact that yours doesn't work. The work was difficult in the 90s to do this work in the 90s computers were getting better, but yes, there were other learning techniques that were on small data sets that worked at least as well as neural networks and were easier to do. explain and they had much more sophisticated mathematics. behind them, so people within computer science lost interest in neural networks within psychology, they didn't because within psychology they are interested in how people could actually learn and these other techniques they seemed even less plausible than backpropagation, which is an interesting part.

Because of your background, you came to this not necessarily because you were interested in computers, but because you were interested in the brain. Yes, I sort of decided that I was originally interested in Psychology, then I decided that we would never understand how people work without understanding the brain. The idea that you could do it without worrying about your brain was kind of a hot idea in the '70s, but I decided that wasn't the case, you had to understand how the brain worked, so fast forward to the 2000s. Wait a minute. key thing you think about is a turning point where okay, our side is going to prevail.

Around 2006 we started doing what we call deep learning. Before that, it had been difficult to come up with neural networks with many representation layers. to learn complicated things and we found better ways to do it, better ways to initialize the networks called pre-training and the p in gbt chat means pre-training okay and the t is Transformer and G is generative and they were actually models generative. provided this better way to pre-train neural ads, so the seeds were there in 2006. By 2009, we had already produced something that was better than the best speech recognizers and recognized what phoneme you were saying using different technology than everyone else.

Speech recognizers were then the standard approach that has been tuned for 20 to 30 years. There were other people using neural networks but not using deep neural ads and then something big happened in 2012. Yeah, actually, two big ones. The good thing is that the research we did in 2009, done by two of my students over a summer, led to better speech recognition that spread to all the big speech recognition labs including Microsoft, IBM and Google, and in 2012 , Google was the first to introduce a product and suddenly voice recognition on Android became as good as Siri, if not better, so it was a deployment of deep neural networks applied to voice recognition three years earlier, at the same time that happened a few months later.

It so happened that two other students of mine developed an object recognition system that looked at images and told you what the object was, and it worked much better than previous systems. How did this system work? Well, there was someone named Faith a Lee among his collaborators who created a huge database of images, like a million images from a thousand different categories, you would have to look at an image and give your best guess as to what the main object in the image was. image, so images would normally have an object in the middle, yeah, and I.

I didn't have to say things like bullet train or Husky or and the other systems were getting like 25 errors and we were getting like 15 errors. Within a few years, 15 were down to three percent, which was about human level and can you explain that to us? Somehow people would understand the difference between the way they were doing it and the way their team did it. I can try, that's all we can hope for. Well, let's say you want to recognize a bird in an image. Well, the image itself, let's assume it is. a 200 by 200 image that has 200 x 200 pixels and each pixel has three values ​​for the three RGB colors, so you have 200 by 200 by 3 numbers on the computer, they're just numbers on the computer, right? and the job is to take those numbers on the computer and convert them into a string that says bird, so how would that be done?

For 50 years, people in standard AI tried to do that and couldn't get a bunch of numbers on a label that said bird, so here's a way to do it at the first level of features: you can create feature detectors that take small combinations of pixels, so you could create a feature detector that says "look if all these pixels are dark and all these pixels are bright". I'm going to turn it on, so that feature detector would represent an edge here, okay, a vertical edge, you might have another one that says if all of these pixels are bright and all of these selections are dark, I'm going to turn it on.

If each detector that represents on a horizontal edge is fine and you can have others for the edges of different bodies, we had a lot of work to do, all we have done is create a correct frame, so we must have many features. configurations like that and that's what you actually have in your brain, okay, so if you look at the cortex of a cattle monkey, it has feature detectors like that, so at the next level, you could say that if you were worried by hand , you would create all these little feature detectors at the next level you would say, um, okay,Suppose I have two edge detectors that join at a fine angle which could just be a peak, so the next level will have a feature detector that detects two of the lower level detectors that join at a fine angle, well , we might also notice a bunch of edges that form a kind of circle, we might have a detector for that, then at the next level we might have a detector that says, "Hey, I found this beak-shaped thing and I find this thing." circular". in about the right spatial relationship to make a bird's eye and beak, so at the next level you would have a bird detector that says if I see those two there, I think it might be a bird, okay? and you can imagine it. plugging all that in by hand, okay, so the idea of ​​backpropagation is just putting in random weights to begin with and now the textures presented would just be garbage, either garbage, okay, okay, but look to see what it predicts and if it happened predict bird it wouldn't, but if you leave the weights alone, you got that right, the connection strings, but if it predicts cat, then what you do is go back through the network and ask the next question and you can do this. with a branch of mathematics called calculus, but you just need to think about the question and the question is how should I change the strength of this connection so that I'm less likely to say cat?

I'm more likely to say bird that's called. the ER, the error, the discrepancy, okay and you figure out, for each connection strength, how do I change it a little bit so that I'm more likely to say bird and less likely to say bird? say cat and a person notices that or the algorithm is set to work. one person has said that this is a bird, then one person looked at the image and said that it is a bird, it is not a cat, it is a bird, then that is a tag provided by a person, but then the backpropagation algorithm it's just a way to figure out how each connection strength changes so that it's more likely to say load less likely to say cat, just keep trying, keep turning, keep doing that and now if you showed enough birds and enough cats when you showed a bird , it will say load when you showed a cat, it will say cat and it turns out that it works much better than trying to connect everything by hand and that's what your students did on this image database, that's why they did it on the image verification device, yes, and they made it work. very good now they were very smart students in fact one of them ilya sutskova is also one of the main people who buy chat gbt so that was a great moment on Ai and chat gbt was another great moment and actually He was involved in both, yes. yeah, I don't know, maybe it's cold in the room, you reached the end of the story.

It gives me chills the thought of you making this little dial and it says bird, it feels like an incredible breakthrough, yes it was mainly because the other people in computer vision thought it was cool, so these neural networks work for simple things like recognizing a handwritten digit, but that's not a really complicated image with a kind of natural background with things that will never work for these big complicated images and suddenly They did it, and you have to give credit to the people who have been staunch critics of neural networks and said these things will never work.

When they worked, they did something that scientists don't normally do and she said, "Oh, it worked, we'll do it." people see it as a big change, yes, it was quite impressive that they changed very quickly because they saw that it worked better than what they were doing. Yes, you point out that when people think about both their machines and ourselves the way that I think we think that language within language must be a language in between, yes, and this is a major misunderstanding, yes, Can you just explain that I think that's complete nonsense? Yeah, so if that were true and it was just a language in the middle, you would have thought that that approach called symbolic AI, yeah, it would have been really good for doing things like machine translation, which is just taking English and producing French art. , or something like that.

You thought manipulating symbols was the right approach for that, but actually neural networks work a lot better than Google Translate when it went from doing that kind of approach to actually using its alerts a lot better, what I think it has in the middle is which has millions of neurons and some of them are active and some of them are not, and that's what's there, the only place you'll find the symbols is at the input and at the output. We are not exactly at the University of Toronto, we are close to the University of Toronto, in universities here and around the world we are teaching. so many people to code, does it still make sense to teach so many people to code?

I don't know the answer to that, around 2015 I said there was no point in teaching radiologists to recognize things in images. and because within the next five years computers will be better at that, yes, we are all about to be radiologists, although well, then the Coopers are not better. I was wrong, it will take 10 years, not five. I was not wrong in spirit, I just received. I think two computers are now comparable to radiologists in many medical images, yes they are still not much better in all of them but they will improve so I think there will be a time when it will still be worth having coders and I don't know how long It will take, but we will need less of them, yes, maybe, or we will need the same number and they will be able to achieve much more.

I was talking about consistency. We reviewed and I visited them yesterday, you are an investor, uh, in them, maybe, maybe the question is how they convinced you, what was the argument that convinced you. I want to invest in this, so they are good people, and I have worked with several of them. Yes, and they were one of the first companies to realize that there is a need to take these great language models that are being developed to places like Google and other places, open up Ai and make them available to companies, so it will be enormously valuable . for companies to be able to use these big language models, um, and that's what they've been doing and they have a significant advantage in that, which is why I think they're going to be successful, another thing you've said.

I find it fascinating, so I want you to talk about the idea that there will be some kind of new type of computer that will be sent to this problem. What is that idea? So there is the biological route to

intelligence

where every brain is different and we have to communicate knowledge from one to another through the use of language and there is the current AI version of neural networks where you have identical models running on different computers and they can actually share the strength of connection to be able to share billions of numbers. how we make a bird, yes, so that they can share all the connection strengths to recognize a bird and one can learn to recognize cats and the other can learn to recognize birds and they can share their connection strengths and now each of them can do both fine and that's what's happening in these big language models that you're sharing, but that only works on digital computers because they have to be able to do identical things and you can't make different biological brains behave identically, so so you can't share connections, yes. but why don't we stick with digital computers?

Due to power consumption, a lot of energy is needed. There are fewer and fewer as chips improve, but it takes a lot of energy to do this, to run a digital computer, you have to run it. such a high power that it pays off exactly the right way, whereas if you're willing to run at a much lower power like the brain is, then you'll allow a little bit of noise and so on, but that particular system will suit the type of noise. on that particular system and everything will work even though you're not running it at such a high power that it pays exactly as you expected and the difference is that the brain runs on 30 watts, a big AI system needs like a megawatt, so we're training with 30 watts and these big systems they're using because they have many copies of the same thing that they're using like a megawatt so you know you're talking about a factor on the order of a thousand in the power requirements and So I think there will be a phase where we train on digital computers, but once something is trained, we'll run it on very low power systems, so if you want your toaster to be able to have a conversation with you and you want to include that.

It only costs a couple of dollars but you can chat with GBT, it better be a low power animal chip. What are the next things you think this technology will do that will

impact

people's lives? It's hard to choose one thing. I think it's going to be everywhere, right, it's already reaching me everywhere. Chat GPT just made a lot of people realize that it's going to be everywhere, but you know, when Google searches, it uses big neural networks to help decide which one it is. The best thing to show you is that we are now at a transition point where Chat GBT is kind of an idiot savant and doesn't really understand the truth either.

It's being trained with a lot of inconsistent data and trying to predict what someone will say next. the web, yes, and people have different opinions and you have to have some kind of combination of all these opinions so that you can model what anyone might say, it's very different from a person trying to have a coherent view of the world, yes, especially if you want to act in the world it's good to have a consistent worldview and I think the good thing that's going to happen is that we're going to move towards systems that can understand different worldviews and can understand that, it's okay if you have this world. point of view, then this is the answer and if you have this other world view then that is the answer, we have our own truths, that is the problem, because what you and I probably believe, unless you are an extreme relativist, is that it's actually a truth for the certainly matters on many topics on many topics or even most topics, yeah, like the Earth isn't actually flat, it just looks flat, yeah, so do we really want a model that says well for some people like we don't know that's going to be a big problem and we don't know how to deal with others present, yeah, and I don't think Microsoft knows how to deal with that, neither do they and it seems to be a big challenge to governance who makes these Decisions are very complicated things, you don't want a big for-profit company deciding what is true, but they are controlling how we fire neurons.

Google is very careful not to do that right now. What Google will do is refer you to relevant documents. that they will have all kinds of opinions, well, they haven't launched their chat product at least as we speak, but we have seen that at least the people who have launched chat products feel that there are certain things that they don't want. it has to be said by his voice, so they go in there and interfere so he doesn't say offensive things, yes, but there's a limit to what you can do that way, there will always be things you didn't think about, yes. so I think Google will be much more careful than Microsoft when they release the chatbot, yes, and it will probably come with a lot of warnings, this is just a chatbot and don't necessarily believe what it says, be careful with the labeling. or be careful how they meddle with him so that he doesn't do terrible things, all those things, careful how they present him as a product and careful how they train him, yeah, um, and they do a lot of work to keep him from saying things bad and good, who decides what is bad?

Some bad things are pretty obvious, but a lot of the big ones aren't, so that's a big open topic right now. I think Microsoft was extremely brave in launching GPT chat, yes. you see this as a bigger thing, some people see it as a bigger social thing, we need regulation or big public debates about how we handle these issues well when it comes to what is true. I mean, do you want the government to decide what's true? certain problem, yes, you don't want the government to do it either. I'm sure you've thought deeply about this question for a long time.

How do we navigate the line between you just sending it out into the world and us finding ways? to cure it like I say, I don't know the answer and I don't think anyone really knows how to handle these problems, we're going to have to learn pretty quickly how to handle these problems because it's a big problem with the president, but Yeah, I don't know how it's going to be do, but I suspect that as a first step, at least these big language models will have to understand that there are different points of view and that endings are relative to a point of view, some people are worried that this could take off very quickly and we may not be prepared for that.

Does that worry you? He does it a little until recently. I thought it was going to be 20 to 50 years, we would probably have general purpose AI, yes. and now I think it may be 20 years or less, so it's okay, some people think thatAI programming was originally established in the 1980s, which is what brought me to Canada, which was into token AI, but you came. I was a weirdo, okay, I was a little weird because I did these things. everyone else thought it was nonsense, they recognized that I was good at this kind of nonsense, so if anyone is going to do that nonsense, it might as well be him.

One of my letters of recommendation said, "You know, I don't believe in this." things, but if you want someone to do it, Jeff's motors work fine, um and then after the program ended, I went back to Britain for a few years and then when I came back to Canada, they decided to fund a deep learning program. , essentially sensitivity. I think you have complaints even with how you define that, yes, when it comes to sensitivity, I'm surprised that people can confidently pronounce that these things are not sensitive and when you ask them what they mean by sensitive, they say well, no.

Not really. I know, so how can you be sure they aren't sentient if you don't know what sentient means? So maybe they already are, who knows. I think whether they are sensitive or not depends on what you mean by sensitive, so you'd better. Define what you mean by sensitive before attempting to answer the question: Are they sensitive? Does it matter what we think or does it only matter if we actually act like we are sentient? That's a very good question Matt and what's his answer. Don't know. have one for sure, it's okay because if you are not sensitive but decide for whatever reason that you believe you are and need to achieve some goal that is contrary to our interests but you believe in your interests, does it really matter if in any human being I believe that a Good , a good context to think about this is a lethal autonomous weapon, yes, okay, so it's all very well to say it's not sentient, but when it's chasing you to shoot you, um, yeah, you're going to start thinking it's sentient, in actually we are not.

We no longer care about an important standard, the kind of intelligence we are developing is very different from our intelligence, so it is this kind of idiot wise intelligence, yes, it is very possible that, as if it were a tool center, it was essentially in a slightly different way than us, but your goal is to make it more like us and you think we'll get there and my goal is to understand each other oh okay no, but yeah, and I think the way you understand us is by building things Like us, okay, that's what I mean. physics is called Richard Feynman said you can't, you can't, you can't understand things unless you can build them, that's the true test of whether you understand it and that's why you've been building, so I've been building, yeah.