Godfather of AI WARNS: "You Have No Idea What's Coming"

[00:00:00] Speaker 1: - Are you at all hopeful that anything can be done to slow down the pace and acceleration of AI? [00:00:05] Speaker 2: - Okay, there's two issues. One is, can you slow it down? And the other is, can you make it so that it will be safe in the end? It won't wipe us all out. I don't believe we're gonna slow it down. And the reason I don't believe we're gonna slow it down is because there's competition between countries and competition between companies within a country, and all of that is making it go faster and faster. And if the US slowed it down, China wouldn't slow it down. [00:00:33] Speaker 1: - Does Ilya think it's possible to make AI safe? [00:00:38] Speaker 2: - I think he does. He won't tell me what his secret source is. I'm not sure how many people know what his secret source is. I think a lot of the investors don't know what his secret source is, but they've given him billions of dollars anyway, 'cause they have so much faith in Ilya, which isn't foolish. I mean, he was very important in AlexNet, which got object recognition working well. He was the main force behind the things like GPT-2, which then led to ChatGPT. So I think having an art of faith in Ilya is a very reasonable decision. [00:01:12] Speaker 1: - There's something quite haunting about the guy that made, and was the main force behind GPT-2, which led rise to this whole revolution, left the company because of safety reasons. He knows something that I don't know about what might happen next. [00:01:27] Speaker 2: Well, the company had, no, I don't know the precise details, but I'm fairly sure the company had indicated that it would use a significant fraction of its resources of the compute time for doing safety research, and then it reduced that fraction. I think that's one of the things that happened. [00:01:46] Speaker 1: - Yeah, that was reported publicly. - Yes. - Yeah. We've gotten to the autonomous weapons part of the risk framework. [00:01:54] Speaker 2: - Right. So the next one is joblessness. - Yeah. - In the past, new technologies have come in, which didn't lead to joblessness. New jobs were created. So the classic example people use is automatic teller machines. When automatic teller machines came in, a lot of bank tellers didn't lose their jobs. They just got to do more interesting things. But here, I think this is more like when they got machines in the industrial revolution, and you can't have a job digging ditches now, because a machine can dig ditches much better than you can. And I think for mundane intellectual labor, AI is just going to replace everybody. Now, it may well be in the form of you have fewer people using AI assistants. So it's a combination of a person and an AI assistant, and now doing the work that 10 people could do previously. [00:02:51] Speaker 1: - People say that it will create new jobs though, so we'll be fine. [00:02:55] Speaker 2: - Yes, and that's been the case for other technologies, but this is a very different kind of technology. If it can do all mundane human intellectual labor, then what new jobs is it going to create? You'd have to be very skilled to have a job that it couldn't just do. So I don't think they're right. I think you can try and generalize from other technologies that have come in, like computers or automatic teller machines, but I think this is different. [00:03:21] Speaker 1: - People use this phrase, they say, AI won't take your job. A human using AI will take your job. [00:03:26] Speaker 2: - Yes, I think that's true. But for many jobs, that'll mean you need far fewer people. My niece answers letters of complaint to a health service. It used to take her 25 minutes. She'd read the complaint and she'd think how to reply, and she'd write a letter. And now she just scans it into a chatbot and it writes the letter. She just checks the letter. Occasionally she tells it to revise it in some ways. The whole process takes her five minutes. That means she can answer five times as many letters. And that means they need five times fewer of her. So she can do the job that five of her used to do. Now, that will mean they need less people. In other jobs, like in healthcare, they're much more elastic. So if you could make doctors five times as efficient, we could all have five times as much healthcare for the same price. And that would be great. There's almost no limit to how much healthcare people can absorb. They always want more healthcare if there's no cost to it. There are jobs where you can make a person with an AI assistant much more efficient and you won't lead to less people because you'll just have much more of that being done. But most jobs, I think, are not like that. [00:04:46] Speaker 1: Am I right in thinking this sort of industrial revolution played a role in replacing muscles? [00:04:52] Speaker 2: Yes, exactly. [00:04:52] Speaker 1: And this revolution in AI replaces intelligence, the brain. Yeah. [00:04:56] Speaker 2: So mundane intellectual labor is like having strong muscles. And it's not worth much anymore. [00:05:03] Speaker 1: So muscles have been replaced. Now intelligence is being replaced. Yeah. So what remains? [00:05:10] Speaker 2: Maybe for a while some kinds of creativity. But the whole idea of super intelligence is nothing remains. These things will get to be better than us at everything. [00:05:18] Speaker 1: So what do we end up doing in such a world? [00:05:21] Speaker 2: Well, if they work for us, we end up getting lots of goods and services for not much effort. [00:05:29] Speaker 1: Okay. But that sounds tempting and nice, but I don't know. There's a cautionary tale in creating more and more ease for humans in it going badly. [00:05:38] Speaker 2: Yes. And we need to figure out if we can make it go well. So the nice scenario is, imagine a company with a CEO who is very dumb, probably the son of the form of CEO. And he has an executive assistant who's very smart. And he says, I think we should do this. And the executive assistant makes it all work. The CEO feels great. He doesn't understand that he's not really in control. And in some sense, he is in control. He suggests what the company should do. She just makes it all work. Everything's great. That's the good scenario. [00:06:18] Speaker 1: And the bad scenario? [00:06:19] Speaker 2: The bad scenario, she thinks, why do we need him? [00:06:23] Speaker 1: Yeah. I mean, in a world where we have superintelligence, which you don't believe is that far away. [00:06:30] Speaker 2: Yeah, I think it might not be that far away. It's very hard to predict, but I think we might get it in like 20 years or even less. [00:06:37] Speaker 1: So what's the difference between what we have now and superintelligence? Because it seems to be really intelligent to me when I use like ChatGPT 3.0 or Gemini. [00:06:45] Speaker 2: Okay. So it's already, AI is already better than us. There's a lot of things. In particular areas like chess, for example, AI is so much better than us that people will never beat those things again. Maybe the occasional win, but basically they'll never be comfortable again. Obviously the same in Go. In terms of the amount of knowledge they have, something like GPT-4 knows thousands of times more than you do. There's a few areas in which your knowledge is better than us. And in almost all areas, it just knows more than you do. [00:07:18] Speaker 1: What areas am I better than it? [00:07:22] Speaker 2: Probably in interviewing CEOs. You're probably better at that. You've got a lot of experience at it. You're a good interviewer. You know a lot about it. If you got GPT-4 to interview a CEO, you'd probably do a worse job. [00:07:37] Speaker 1: Okay. I'm trying to think if I agree with that statement. GPT-4, I think, for sure. But it may not be long before. Yeah, I guess you could train one on how I ask questions and what I do. Sure. [00:07:54] Speaker 2: And if you took a general purpose foundation model and then you trained it up on not just you, but every interviewer you could find doing interviews like this, but especially you, you'd probably get to be quite good at doing your job, but probably not as good as you for a while. [00:08:12] Speaker 1: Okay. So there's a few areas left. And then superintelligence becomes when it's better than us at all things. [00:08:18] Speaker 2: When it's much smarter than you and at almost all things it's better than you, yeah. [00:08:22] Speaker 1: And you say that this might be a decade away or so. [00:08:26] Speaker 2: Yeah, it might be. It might be even closer. Some people think it's even closer. It might well be much further. It might be 50 years away. That's still a possibility. It might be that somehow training on human data limits you to not being much smarter than humans. It might be more likely to be more powerful than humans. It might be more powerful than humans. [00:08:44] Speaker 1: It might be more powerful than humans. It might be more powerful than humans. It might be more powerful than humans. It might be more powerful than humans. It might be more powerful than humans. It might be more powerful than humans. It might be more powerful than humans. It might be more powerful than humans. It might be more powerful than humans. I'm a CEO of a big AI agent company and a few other people. And it was the first moment where I had... No. It was another moment where I had a eureka moment about what the future might look like. When I was able in the interview to tell this agent to order all of us drinks. And then five minutes later in the interview, you see the guy show up with the drinks. And I didn't touch anything. I just told it to order us drinks to the studio. [00:09:17] Speaker 2: And you didn't know about who you normally got your drinks from. It figured that out from the web. [00:09:21] Speaker 1: Yeah, I figured it out because it went on Uber Eats. It has my data, I guess. And we put it on the screen in real time so everyone at home could see the agent going through the internet, picking the drinks, adding a tip for the driver, putting my address in, putting my credit card details in. And then the next thing you see is the drinks show up. So that was one moment. And then the other moment was when I used a tool called Replit. And I built software by just telling the agent what I wanted. [00:09:46] Speaker 2: Yes. It's amazing, right? [00:09:48] Speaker 1: It's amazing and terrifying at the same time. [00:09:51] Speaker ?: Yes. [00:09:52] Speaker 2: And if you can build software like that, right? [00:09:54] Speaker 1: Yeah. [00:09:55] Speaker 2: Remember that the AI, when it's training, is using code. And if you can modify its own code, then it gets quite scary, right? [00:10:05] Speaker 1: Because it can modify itself. [00:10:05] Speaker 2: It can change itself in a way we can't change ourselves. We can't change our innate endowment, right? There's nothing about itself that it couldn't change. [00:10:16] Speaker 1: On this point of joblessness, you have kids. [00:10:18] Speaker 2: I do. [00:10:19] Speaker 1: And they have kids? No, they don't have kids. So no grandkids yet. What would you be saying to people about their career prospects in a world of super intelligence? What should we be thinking about? [00:10:29] Speaker 2: Yeah. In the meantime, I'd say it's going to be a long time before it's as good at physical manipulation as us. Okay. And so a good bet would be to be a plumber. [00:10:40] Speaker 1: Until the humanoid robots show up. In such a world where there is mass joblessness, which is not something that you just predict, but this is something that Sam Altman at OpenAI, I've heard him predict. And many of the CEOs, I mean, Elon Musk, I watched an interview, which I'll play on screen, of him being asked this question. And it's very rare that you see Elon Musk silent for 12 seconds or whatever it was. Right. And then he basically says something about he actually is living in suspended disbelief, i.e. he's basically just not thinking about it. [00:11:08] Speaker 2: When you think about advising your children on a career with so much that is changing, what do you tell them there's going to be of value? [00:11:27] Speaker 3: Well, that is a tough question to answer. I would just say, you know, to sort of follow their heart in terms of what they find interesting to do or fulfilling to do. I mean, if I think about it too hard, frankly, it can be just dispiriting and demotivating. Because, I mean, I go through, I've put a lot of blood, sweat and tears into building the companies. And then I'm like, well, should I be doing this? Because if I'm sacrificing time with friends and family that I would prefer to do, but then ultimately the AI can do all these things. Does that make sense? I don't know. To some extent, I have to have deliberate suspension of disbelief in order to remain motivated. So I guess I would say just, you know, work on things that you find interesting, fulfilling, and that contribute some good to the rest of society. [00:12:32] Speaker 2: Yeah, a lot of these threats, it's very hard to, intellectually, you can see the threat, but it's very hard to come to terms with it emotionally. Yeah. I haven't come to terms with it emotionally yet. What do you mean by that? I haven't come to terms with what the development of superintelligence could do to my children's future. I'm okay. I'm 77. I'm going to be out of here soon. But for my children and my younger friends, my nephews and nieces and their children, I just don't like to think about what could happen. Why? Because it could be awful. [00:13:25] Speaker 1: In what way? [00:13:27] Speaker 2: Well, if I ever decided to take over, I mean, it would need people for a while to run the power stations until it designed better analog machines to run the power stations. There's so many ways it could get rid of people, all of which would, of course, be very nasty. [00:13:45] Speaker 1: Is that part of the reason you do what you do now? [00:13:48] Speaker 2: Yeah. I mean, I think we should be making a huge effort right now to try and figure out if we can develop it safely. [00:13:55] Speaker 1: Are you concerned about the midterm impact potentially on your nephews and your kids in terms of their jobs as well? Yeah, I'm concerned about all that. Are there any particular industries that you think are most at risk? People talk about the creative industries a lot and sort of knowledge work. They talk about lawyers and accountants and stuff like that. [00:14:12] Speaker 2: Yeah. So that's why I mentioned plumbers. I think plumbers are less at risk. [00:14:15] Speaker 1: Okay. How do I become a plumber? [00:14:16] Speaker 2: Someone like a legal assistant, a paralegal. They're not going to be needed for very long. [00:14:23] Speaker 1: And is there a wealth inequality issue here that will rise from this? [00:14:27] Speaker 2: Yeah. I think in a society which shared out things fairly, if you get a big increase in productivity, everybody should be better off. Everybody should be better off. But if you can replace lots of people by AIs, then the people who get replaced will be worse off. And the company that supplies the AIs will be much better off. And the company that uses the AIs. So it's going to increase the gap between rich and poor. And we know that if you look at that gap between rich and poor, that basically tells you how nice a society is. If you have a big gap, you get very nasty societies in which people live in wall communities and put other people in mass jails. It's not good to increase the gap between rich and poor. [00:15:17] Speaker 1: The International Monetary Fund has expressed profound concerns that generative AI could cause massive labor disruptions and rising inequality and has called for policies that prevent this from happening. I read that in the Business Insider. [00:15:31] Speaker 2: Have they given any idea what the policies should look like? [00:15:33] Speaker 1: No. [00:15:34] Speaker 2: Yeah, that's the problem. I mean, if AI can make everything much more efficient and get rid of people for most jobs or have a person assisted by AI doing many, many people's work, it's not obvious what to do about it. [00:15:47] Speaker 1: It's universal basic income. Give everybody money. [00:15:51] Speaker 2: Yeah, I think that's a good start. And it stops people starving. But for a lot of people, their dignity is tied up with their job. I mean, who you think you are is tied up with you doing this job, right? [00:16:05] Speaker 1: Yeah. [00:16:06] Speaker 2: And if we said, we'll give you the same money just to sit around, that would impact your dignity. [00:16:13] Speaker 1: You said something earlier about it's surpassing or being superior to human intelligence. A lot of people, I think, like to believe that AI is on a computer and it's something you can just turn off if you don't like it. [00:16:24] Speaker 2: Well, let me tell you why I think it's superior. Okay. It's digital. And because it's digital, you can simulate a neural network on one piece of hardware. Yeah. And you can simulate exactly the same neural network on a different piece of hardware. So you can have clones of the same intelligence. Now, you could get this one to go off and look at one bit of the internet and this other one to look at a different bit of the internet. And while they're looking at these different bits of the internet, they can be syncing with each other. So they keep their weights the same, the connection strength the same, weights of connection strength. So this one might look at something on the internet and say, oh, I'd like to increase this strength of this connection a bit. And it can convey that information to this one. So it can increase the strength of that connection a bit based on this one's experience. [00:17:14] Speaker 1: And when you say the strength of the connection, you're talking about learning. [00:17:17] Speaker 2: That's learning. Yes. Learning consists of saying, instead of this one giving 2.4 votes for whether that one should turn on, we'll have this one give 2.5 votes for whether this one should turn on. Okay. And that would be a little bit of learning. Mm-hmm. So these two different copies of the same neural net are getting different experiences. They're looking at different data, but they're sharing what they've learned by averaging their weights together. [00:17:41] Speaker 3: Mm-hmm. [00:17:42] Speaker 2: And they can do that averaging at like, you can average a trillion weights. When you and I transfer information, we're limited to the amount of information in a sentence. And the amount of information in a sentence is maybe 100 bits. It's very little information. We're lucky if we're transferring like 10 bits a second. These things are transferring trillions of bits a second. So they're billions of times better than us at sharing information. And that's because they're digital and you can have two bits of hardware using the connection strengths in exactly the same way. We're analog and you can't do that. Your brain's different from my brain. And if I could see the connection strengths between all your neurons, it wouldn't do me any good because my neurons work slightly differently and they're connected up slightly differently. Mm-hmm. So when you die, all your knowledge dies with you. When these things die, suppose you take these two digital intelligences that are clones of each other and you destroy the hardware they run on. As long as you've stored the connection strength somewhere, you can just build new hardware that executes the same instructions. So it'll know how to use those connection strengths and you've recreated that intelligence. So they're immortal. We've actually solved the problem of immortality. But it's only for digital things. [00:18:56] Speaker 1: So it knows, it will essentially know everything that humans know but more because it will learn new things. [00:19:03] Speaker 2: It will learn new things. It will also see all sorts of analogies that people probably never saw. So for example, at the point when GPT-4 couldn't look on the web, I asked it, why is a compost heap like an atom bomb? Off you go. [00:19:21] Speaker 1: I have no idea. [00:19:22] Speaker 2: Exactly. Excellent. That's exactly what most people would say. It said, well, the time scales are very different and the energy scales are very different. But then it went on to talk about how a compost heap, as it gets hotter, generates heat faster. And an atom bomb, as it produces more neutrons, generates neutrons faster. And so they're both chain reactions, but at very different time and energy scales. And I believe GPT-4 had seen that during its training. It had understood the analogy between a compost heap and an atom bomb. And the reason I believe that is, if you've only got a trillion connections, remember you have a hundred trillion, and you need to have thousands of times more knowledge than a person, you need to compress information into those connections. And to compress information, you need to see analogies between different things. In other words, it needs to see all the things that are chain reactions and understand the basic idea of a chain reaction and code that, and then code the ways in which they're different. And that's just a more efficient way of coding things than coding each of them separately. So it's seen many, many analogies, probably many analogies that people have never seen. That's why I also think that people say, "These things will never be creative." They're going to be much more creative than us, because they're going to see all sorts of analogies we never saw. And a lot of creativity is about seeing strange analogies. [00:20:43] Speaker 1: If you love The Driver's CEO brand and you watch this channel, please do me a huge favour, become part of the 15% of the viewers on this channel that have hit the subscribe button. It helps us tremendously, and the bigger the channel gets, the bigger the guests.