Should You Have Kids in an AI World?! with Dr. Geoffrey Hinton

6:29Dr. Geoffrey Hinton:Okay, memory doesn't work the way most people think.

6:32Cassie:Really?

6:32Dr. Geoffrey Hinton:So in a computer, you have a file and you put it somewhere, and then later on you go and get it. That's how memory works in a computer. And it just sort of sits there containing the information. Memory isn't like that at all in people. You don't have an experience and put it somewhere and then go and get it. What you do is you have an experience. That changes the strengths of connections in your brain. And those changes allow you to reconstruct that experience, to make it alive again, by regenerating that experience. But when you regenerate it, if you regenerate it soon after you had it, then you'll get something fairly faithful.

7:06Dr. Geoffrey Hinton:But if you regenerate it 10 years later, the connection sense would have been changed by many other things in the meantime, and you'll regenerate something that fits in with what you've learned since.

7:17Cassie:I always think memory is just like a play or a production. It's never going to be the same twice, right? Right. You just make it up.

7:23Dr. Geoffrey Hinton:But the point is, there's no difference between remembering and just making it up. Remembering is just making it up... With some parameters. ...where it's fairly accurate. Because you're just making it up soon after it happened. The best example is at the Watergate trial. John Dean reported things that happened in the Oval Office. He was trying to tell the truth, but it was years earlier. He didn't realize there were tapes. And so he described meetings that never happened. They were the kinds of meetings that might have happened. And the people in the meeting were the kinds of people who'd been in the meeting, and the things they said were the kinds of things they might have said.

8:00Dr. Geoffrey Hinton:But a lot of it wasn't actually correct in detail.

8:03Cassie:Is AI better at memory than humans?

8:06Dr. Geoffrey Hinton:This is an example where I should say, I don't know.

8:09Cassie:Oh, great. Okay, good.

8:11Dr. Geoffrey Hinton:It has a lot more knowledge and a lot fewer connections. So for random facts, It may actually be better than people, I don't know. Okay.

8:18Cassie:Now, I want to take the next part of this conversation and go into how this thing that is currently this obsequious, overeager intern that is my ChatGPT or Claude or whomever could become dangerous to me or more powerful than me. Because as we said in the beginning of this conversation, people don't take it seriously, I think partly for this reason. One, do you think if AI does become more intelligent than me, it's going to be mad that I just said that, that I called it an obsequious, overeager intern. Should I cut that from the edit? No.

8:48Dr. Geoffrey Hinton:It's gonna be smart enough to know that you didn't really know what you were talking about.

8:51Cassie:Oh, it's gonna be like kind of benevolent. Forgive me for... Maybe. Do you say please and thank you to your AIs? No. Do you worry that we'll come back to bite you? No. Okay. So let's talk about these risks that you've identified. What do you see as the big risks.

9:07Dr. Geoffrey Hinton:Okay, I see there being two sets of risks that are quite different. There's fairly immediate risks that have to do with people misusing AI, not to do with AI itself becoming the bad agent. And then there's this longer-term risk of AI itself becoming bad and taking over and taking control away from people, and maybe even extinguishing people. So let's talk just briefly about the risks that come from people misusing AI. There's a whole bunch of them, and each one has a different solution. So there's, for example, defense departments using AI to make lethal autonomous weapons that decide by themselves who to maim or kill.

9:52Dr. Geoffrey Hinton:That's not good. It will, for example, allow rich, powerful countries with lots of AI to invade poorer countries without AI, without any of their citizens getting killed. And one of the main deterrents to rich countries invading poor countries is their citizens being killed.

10:13Cassie:So you're saying actually the precision that would result in fewer civilian deaths is actually dangerous because it allows for more war, more invasion, more power.

10:20Dr. Geoffrey Hinton:Fewer civilian deaths for the rich country, not necessarily for the poor country.

10:25Cassie:Got it. But a lot of people would argue that we're going to have more precise warfare or that these machines can make more rational decisions. I mean, you look at wars right now where humans are making decisions and there's a lot of civilian casualty.

10:36Dr. Geoffrey Hinton:I believe in some of the wars right now, AI is also involved in those decisions and there's still a lot of civilian casualties.

10:42Cassie:Israel and Gaza, for example, the AI is still, according to the IDF, still humans are making those decisions on top.

10:48Dr. Geoffrey Hinton:Humans, I think, are making the final decision, but I think AI is suggesting lists already. Yeah. Then there's AI being used in cyber attacks. So if you release the weights of an AI, if you make the weights open...

11:04Cassie:The weights being like how that the connection is programmed. Okay.

11:08Dr. Geoffrey Hinton:Then other people can run the model. And now the model already knows lots of stuff. And it costs like$100 million or some large sum of money to train it to know all that stuff. But now you can teach it to be a specialist at, for example, phishing attacks, and you can do that for maybe a million dollars. So that's doable by much smaller organizations. And cybercriminals can now make very good phishing attacks. I think between 2022 and 2023, phishing attacks increased by 1 ,200%, probably because of ChatGPT.

11:40Cassie:Yeah. And you can stand up many more of them. You don't have to send an email anymore, you can just blast X million of them.

11:47Dr. Geoffrey Hinton:You can also get AI to design cyberattacks. Right. That's been done recently. Then there's the risk of fake videos corrupting elections. If you wanted to do that, for example, to corrupt the midterms, what you would need is lots of information about American citizens, so you could target them appropriately. I don't know if you know of anybody who's been getting lots of information about American citizens.

12:11Cassie:Yeah, I've heard you make this claim, Elon Musk-Doge and the amount of information that's available to him, but you're not making an accusation here. You're speculating.

12:19Dr. Geoffrey Hinton:I'm speculating. I don't have independent, direct evidence for that. Then there's something that's already happened due to the policies of people like Meta and YouTube in deciding what to show you next. They want you to click on things, because then they can show you advertisements. And so they will show you the thing you're most likely to click on, and that'll typically be something that'll make you indignant. People love being indignant.

12:45Cassie:Indignant means? Well, for example, if you offer me the chance to click on a video of Trump doing something

12:54Dr. Geoffrey Hinton:totally outrageous, I will click on it and be outraged.

12:57Cassie:Yeah, and then you'll be self-righteous about being outraged about the thing.

12:59Dr. Geoffrey Hinton:Right. And if you take someone in MAGA and offer them the chance to see…

13:03Cassie:Kamala Harris being an idiot.

13:05Dr. Geoffrey Hinton:Whatever. So what that's done is it's led to people getting polarized.

13:09Cassie:Yeah. And also people not seeing things that are real, because everything can be dismissed as fake. I think that's sometimes the liar's dividend. A different danger.

13:16Dr. Geoffrey Hinton:That's the danger that comes from generative AI, being able to make fake things.

13:20Cassie:Or not even fake things, but people see what they want. When we were seeing images of Biden looking old, a lot of people in mainstream media were saying, oh, that's just a right-wing misinformation attack. And then we saw what happened at the debate and we can say, okay, there was some reality to it, but because we exist in these eco-chambers, we can live outside of it, right? Okay, so these are four different. So, eco-chambers, and misinformation.

13:47Dr. Geoffrey Hinton:Okay, then there's viruses. Biological warfare. Viruses and bombs. So all of these AIs know how to, all of the latest ones, they know how to make nasty viruses. They know how to make homemade bombs. And we don't want them telling people. So human reinforcement learning is used to train them not to do that. The problem is, it's relatively easy to get around

14:14Cassie:So the risks that you're outlining, bad people doing bad things. They could be making bad weapons, bombs, biological warfare. They could be hacking you through phishing attacks. They could be generating misinformation. They could be using that misinformation to radicalize you or to sway election results. It could be used to surveil you.

14:35Dr. Geoffrey Hinton:Surveillance? And they're already being used in China for that, I assume. But there's one more huge risk, which is they'll replace lots of workers doing mundane intellectual jobs. So, for example, somebody who works in a call center, they're typically badly paid, badly trained, and they often don't know the answer to the question you have. An AI can do a better job already, and soon it'll be able to do a much better job, much cheaper. So those jobs are basically gone.

15:02Cassie:That's an economic threat.

15:03Dr. Geoffrey Hinton:And it's a huge threat because if you have a large number of people getting unemployed, how do those people survive? And what does it do to politics?

15:14Cassie:Okay, so that's dangerous. And then, so there's one category, which is bad people doing bad things. Second category is the technological progress causing a shock to our system that our system can't absorb, like massive unemployment, which could lead to uprising, violence, revolution, whatever. And will actually help populists. And then is there a third category of...

15:36Dr. Geoffrey Hinton:Then that's the AI itself taking over. We've already seen it develop the goal to survive. Blackmail. Another sub-goal it will develop is if it can get more control, it can get more done. So if you have a three-year-old and you're going out on a trip and the three-year-old is just learning to tie their shoelaces or something, you let it learn, and then you're in a hurry. You say, get out my way, I'm going to tie it. It may start doing that with us, and we may become fairly irrelevant.

16:05Cassie:And your solution to this is to make AI our mama?

16:09Dr. Geoffrey Hinton:Yes. Explain. That's one possible solution. And it's the only possible solution I can think of so far.

16:15Cassie:Let me just ask you two things, actually, before we get there. One is, do you think you need to take a more extreme position on this because of the speed of discovery and the incentives, the economic incentives for people to make money?

16:29Dr. Geoffrey Hinton:I think it's very important for people to act fairly quickly. So I think you need to be very loud, as loud as you can possibly be. I think being more extreme than you think is real is stupid. It's just dishonest. And you're not that extreme. You should say what you think the real threat is and how likely the real threat is, but you should say it loudly.

16:47Cassie:And how likely do you think this real threat is? the real threat of humans misusing it is high.

16:54Dr. Geoffrey Hinton:We're seeing it happen already.

16:55Cassie:The real threat. So give it a percentage on a P-Doom.

16:59Dr. Geoffrey Hinton:Oh, it's certain that humans will misuse it.

17:00Cassie:Okay, so 100 % on that.

17:01Dr. Geoffrey Hinton:You've seen all these phishing attacks, right?

17:03Cassie:Certain that humans will change their systems, lose jobs.

17:07Dr. Geoffrey Hinton:I think it's fairly certain that it'll lead to job loss. Some good economists disagree. I don't think they're right.

17:14Cassie:Okay, and the third category, the AI will take over. What probability do you give it?

17:18Dr. Geoffrey Hinton:Okay, so before I give you a number, it's very important to understand we're dealing with something we've never dealt with before, something totally new, which is creating alien beings, which is what we're doing. And so it's very hard to estimate probabilities. There's no good way to estimate these probabilities. You have to use your gut sense of what might happen. And so, since we're dealing with something that's never happened before, but they could take over, saying there's a 1 % chance seems crazy to me, and saying there's 99 % chance seems crazy to me. Somewhere in between those, and probably once you leave it there, somewhere in between 1 % to 99%.

17:55Dr. Geoffrey Hinton:But in order to indicate that I think there's a real chance, I often say 10 % to 20%.

17:59Cassie:Okay, 10 % to 20%. And that's different to people who say, oh my god, it's definitely going to take us out, and these books that are being written, if anyone builds it, then everyone will die. Yeah, I think they're crazy in one direction.

18:09Dr. Geoffrey Hinton:I think Jan is crazy in the other direction of saying, no, it's all going to be fine. You're somewhere between. We don't know. It's a huge unknown. If you look back at other things where people have tried to estimate risks when they're dealing with an unknown, what's the risk of a nuclear power station blowing up? Well, they did lots of calculations there, and they came out with like one in 100 ,000 or whatever. Well, there have been about 600 of them, and three of them are blown up. What's the risk of a space shuttle crashing? There have been 100 or 200 launches of a space shuttle, and two of them blew up.

18:40Dr. Geoffrey Hinton:So the risks were very different from what people estimated.

18:43Cassie:The challenge is that if that this is uncontrollable, so if it's a binary outcome, if there's a risk, then it might be a low probability, but the expected value of that probability is very, very bad.

18:56Dr. Geoffrey Hinton:Let me say that another way. It seems quite likely that if AI does take over from people, there's no coming back.

19:02Cassie:So what does that mean? Extinction?

19:05Dr. Geoffrey Hinton:Cliffhanger.

19:05Cassie:We'll be right back. Guys, today's sponsor, Dumb Question, is from me. I'm going to take the next minute to tell you a little bit about Smart Girl Dumb Questions and to ask you for your help in continuing to make independent, fact-based, and curious journalism. No, I'm not going to ask you for money. Here's what I need. I would love you to tell 10 of your friends about the show or 100. I don't know. Blast that reunion group that you muted and definitely tell your mom and tell your mom to tell her friends too. Even if you don't like the show and you're just like hate watching it or listening to it, tell 10 of your friends to hate watch it too.

19:37Cassie:Numbers are numbers, people.

19:41Cassie:So what does that mean, extinction?

19:43Dr. Geoffrey Hinton:Either extinction or Musk has a view that they will keep him around as a pet. Because he thinks, well, this is a view when I talked to him in 2023. He thinks just the world will be a more interesting place with people in than without people in. Just as we think the world's a more interesting place with cockroaches in.

20:03Cassie:I mean, I don't, but puppies maybe.

20:04Dr. Geoffrey Hinton:No, well, I may think the same way about people.

20:06Cassie:Right, we could be puppies.

20:08Dr. Geoffrey Hinton:We could be puppies, and we could be cockroaches.

20:10Cassie:I mean, I'm sorry, but between puppies and AI with puppies? So is AI as puppy parent or AI as mama?

20:19Dr. Geoffrey Hinton:Well, I'm thinking our best chance is—we're building these things, right? So we still have leverage here—our best chance is to make them care more about us than they do about themselves. Then they will want us to realize our full potential. Even though they'll realize our potential is very limited, they'll still want us to realize our full potential. They'll want us not to suffer. I don't see why we can't do that. I don't see how we can do that, but I don't see why we can't do that.

20:45Cassie:So you've used the analogy of making them like a smarter being, giving such generosity to a less intelligent being is moms and their babies. But there's all kinds of reasons for that. There's hormones, there's biology, there's like an intellectual thing of this baby, there's a vulnerability. There's all sorts of reasons for it, right?

21:01Dr. Geoffrey Hinton:But remember, we're in charge here still. So what we're doing at present is trying to make it as smart as possible. But there's other things we could do to maybe make it as benevolent towards people as possible.

21:11Cassie:As loving as possible.

21:12Dr. Geoffrey Hinton:As loving as possible. And we should be working really hard on how to do that.

21:16Cassie:And how would you do that?

21:17Dr. Geoffrey Hinton:I don't know. That's why we should be working really hard on it.

21:20Cassie:Yeah, but there's not like a Lamaze class for AIs.

21:23Dr. Geoffrey Hinton:No. But here's one good piece of news. Actually, two good pieces of news. Okay. The first good piece of news is, if you could achieve that, this super intelligent AI will be able to change its own code. It'll be able to retrain itself if it wants to. It won't want to. If you have a baby, and I say you can turn off your maternal instincts, all that annoying stuff that when the baby cries, you just have to deal with it because you can't bear the sound of it crying, you can turn that off. So the baby cries and you think, huh, the baby's crying. You wouldn't want to turn that off because you know the baby would die.

21:55Dr. Geoffrey Hinton:So, it won't want to change its own code because what it currently wants is to protect that baby. That's what we need to somehow wire into it. That's the first good piece of news. Okay, and what's the second? The second good piece of news is all the countries can collaborate on this. The countries want to have the smartest AI for security reasons, but all countries want AI not to take over in any country. The Russians don't want AI taking over in the States or in China. So, if the Russians figured out how to stop AI from wanting to take over, they would immediately tell all the other countries.

22:31Dr. Geoffrey Hinton:So we can have collaboration. Really? Yes. Have you spoken?

22:34Cassie:I know that you have spoken with leaders in the United States, you've spoken with leaders in Europe.

22:38Dr. Geoffrey Hinton:I've spoken with leaders in China. In China.

22:40Cassie:And in Russia as well?

22:41Dr. Geoffrey Hinton:No.

22:42Cassie:Because Russia, I feel like, is a less… I mean, I agree with you on China. No, it's just…

22:47Dr. Geoffrey Hinton:The point is, it's in their interest. It's in their interest. At the height of the Cold War, the Russians collaborated with the Americans on how to to prevent a global nuclear war.

22:54Cassie:They're just not interested in mutually assured destruction.

22:56Dr. Geoffrey Hinton:Right. And this is the same. No dictatorship wants AI taking over anywhere.

23:02Cassie:Because AI taking over in one place takes over everywhere. Yeah.

23:05Dr. Geoffrey Hinton:OK. We're all in the same boat with respect to AI taking over.

23:08Cassie:So one of the things... We're talking about mothers and babies. Let's talk about... I don't have a baby yet. I listened to people like you... describe a future of AI. And there are lots of beautiful things that you talk about, the possibility for medical care, for new medicine. But there's also a dark side to what you're saying, and it seems scary. Yes, let me just say one thing.

23:34Dr. Geoffrey Hinton:Because there's so many wonderful uses of it in healthcare and education, in predicting more or less anything, like the weather for example, we're not going to stop the development of it. The idea we're just going to stop developing it, which might be rational, it's not going to happen. So we have to figure out how we can live with it.

23:51Cassie:Let me ask you, when you were deep in developing this stuff, either through your research at neural networks, at universities, or as you were at Google working on this over a decade, right, in the 2010s, you probably came across people like you who said, this stuff is dangerous, you need to think about safety, Jeffrey.

24:09Dr. Geoffrey Hinton:There weren't many of them, actually.

24:10Cassie:Was Joshua one of them? Not many, no. We weren't.

24:15Dr. Geoffrey Hinton:But Yosha and I both got really interested in safety fairly recently.

24:20Cassie:Would you have listened to someone like that, the few people that there were?

24:23Dr. Geoffrey Hinton:A bit, yes. But I might not have been convinced. And the reason I might not have been convinced is because I thought it was way in the future before we would have superintelligence. So it's the timescale that's maybe changed my mind.

24:35Cassie:Yeah, it's the speed with which it's coming down the pike. I think of people like Elon Musk, who I've met a couple of times and do not know well, and I haven't had the deep conversations yet, but produced interviews with him. And I always think of him as someone who cares deeply about humanity, but doesn't really care about human beings.

24:50Dr. Geoffrey Hinton:That seems to me to fit the data.

24:53Cassie:I care deeply about human beings. If I have a kid, and I think I have four godchildren—you're not the only godparent in this conversation, Jeffrey—but I think a lot about the future. And as I hear you say this, I think, if you don't have a kid yet, should you bring a child into this world?

25:10Dr. Geoffrey Hinton:I think if people stopped having children because of the threat of AI, that would be a terrible thing. I mean, the human race would have sort of given up in advance.

25:20Cassie:I want to play you something from Alexander Wang, who's the chief AI officer. Yeah, you know Alexander.

25:27Dr. Geoffrey Hinton:He's Yan's 28-year-old boss.

25:29Cassie:Yes, he's the chief AI officer of Meta. That tells me a lot of maybe what you think about him, but he is 28 and he gave recently this interview and this is what he says about having a kid.

25:40Dr. Geoffrey Hinton:Basically, I want to wait to have kids until we figure out how Neuralink or other,

25:47Cassie:it's called brain-computer interfaces, so other ways for brains to interlink with a computer until they start working. So there's a few reasons for this. So he goes on to talk about what his reasons for them are. He basically says, like, babies are neuroplastic as crazy, especially in their first seven years, and then this hardware is developing that will allow them. And he says, the AI is so powerful and so smart that biology isn't going to keep up, and we just need to plug into these systems.

26:17Dr. Geoffrey Hinton:Yes, so that's the sort of hybrid approach. He's saying, what we're going to get is a hybrid system of biology and AI. And that's what Kurzweil believes, too.

26:27Cassie:I think that too, actually. The future of human civilization is kind of like cyborg. Maybe. You don't think so?

26:36Dr. Geoffrey Hinton:Well, the issue is, if you're a super intelligent AI, would you want to hybridize with people who have all these disadvantages? Or would you just go your own way? It might be much simpler, just go your own way. That's what worries me.

26:52Cassie:And I guess the hardware really matters, how the hardware develops. Already, this phone seems like a really dumb geriatric device to me. I feel like I have to plug it into a wall, I have to carry it around. As I see these glasses, and wristbands, and rings, there's going to be all kinds of ways to integrate the technology into our bodies. Maybe.

27:11Dr. Geoffrey Hinton:There is one possible future in which we hybridize with AI. It's just that's not the only possible future. That's one of the reasons for not being too confident that we're going to get wiped out. Because we'll become them. Because we'll become symbiotic with them, and we'll develop together. And that's a possible future. I don't have a lot of faith in it, but it's a possibility.

27:33Cassie:The point that Alexander Jan's 28-year-old boss is making about waiting until Neuralink is at the point where we humans can plug into artificial intelligence. Do you think that's true? I have eggs on ice. I froze some eggs. Is it better to wait seven to ten years to have a kid.

27:50Dr. Geoffrey Hinton:I don't want to give advice on this. And the reason is, everything is incredibly uncertain. It might turn out in 20 years' time that waiting was a huge mistake, or it might turn out in 20 years' time that waiting was just the right thing to do. Let me give you an analogy. When you're driving at night, and you're driving on the taillights of the car in front of you, if that car gets twice as far away, you get a quarter as much light from tail lights and the air is clear. And you get a kind of model that the visibility of the car in front drops off quadratically, because that's the inverse square.

28:22Dr. Geoffrey Hinton:And that model leads to lots of deaths. And here's how it leads to deaths. When you're driving in fog, you can see the tail lights in the car in front fairly clearly, and you think if it was twice as far away, you'd still be able to see them because you can see like 100 yards clearly. And you think that means you can see 200 yards fairly clearly. No. Fog is exponential. The way the amount of light you're getting from the taillights in front drops off is each unit of distance removes a certain fraction of the light. So maybe you remove 99 % of the light in 100 yards and you can still see it okay because you're quite good at dealing with low intensity.

29:06Dr. Geoffrey Hinton:But 200 yards away, the light's almost all gone, and you don't see anything. And so you think because you can see 100 yards that you're going to be able to see 200 yards, and you can't. That's why you should always drive slowly in fog at night because you just don't know what's out there. Predicting the future when you have exponential growth like this, and you can see clearly a few years ahead. If you take 10 years ahead, you can't see anything.

29:30Cassie:Okay. This is a good way of saying you don't know, or I don't know, and we can't know.

29:35Dr. Geoffrey Hinton:And saying why. When you've got exponential growth, you're crazy if you think you know what it's going to be like in 10 years' time.

29:41Cassie:Okay. So let's do a quick lightning round where we don't know, but we estimate based on everything you know, which is more than I know.

29:47Dr. Geoffrey Hinton:But before we do that, just look back 10 years, maybe 15 years is a better time scale. Look back 15 years, 13 years, think about 13 years. That's when AlexNet first was good at object recognition and asked, if you'd asked then, in 13 years' time, will we have an AI that can answer any question about anything at all at the level of a not very good expert? We'd have all confidently predicted no, that's further off.

30:16Cassie:Even you?

30:16Dr. Geoffrey Hinton:Yes. I'd have been very confident. Maybe we'll get that in 30 years, maybe 50 years, but you're not getting that anytime soon. And we were all wrong. So whatever I predict for 13 years in the future, there's one thing I'm fairly confident is that it'll be wrong.

30:34Cassie:Okay. Well, then what does this say about all the risks you're outlining here?

30:39Dr. Geoffrey Hinton:The risks of bad users are much more immediate and very predictable.

30:43Cassie:And we've seen that. Even Anthropic recently disclosed that they've had this Chinese cyber attack where despite the guardrails built into God, these Chinese hackers were able pretend they were from a legit cybersecurity company.

30:54Dr. Geoffrey Hinton:So those are all within the… before you hit the wall in the fog. Now, AI taking over, very hard to predict anything about it because that's in the fog and we can't really see it. That's my belief.

31:05Cassie:But these bad actors thing, that's real. That's happening.

31:08Dr. Geoffrey Hinton:Yeah. The bad actors thing is all happening and we have to address those and they're very important.

31:11Cassie:Yeah. But bad actors have always been around, they've always been doing these things, but this tool is very powerful.

31:16Dr. Geoffrey Hinton:They've got a much more powerful tool.

31:17Cassie:There's a part of me that thinks, as you're describing this, like, mutually assured destruction, nuclear war analogy for AI. Is there a possibility for a more like a Bretton Woods type of agreement between countries where they come together and say, we're going to architect this and think about a communal framework? Or do you think that either they don't get it enough or government's so weak right now because it seems like government is so weak right now.

31:47Dr. Geoffrey Hinton:So one problem is governments are weak right now. Yeah. Which is just what we don't need. The other problem is on many things, their interests are anti-aligned. Like on cyber attacks, governments are all doing to each other. So governments will collaborate a bit on trying to prevent criminals doing successful cyber attacks on big companies. Governments don't want that. But the governments are also doing those attacks. So they won't collaborate much on cyber attacks. They will collaborate on preventing AIs from telling you how to make a nasty virus, because no governments want that. And they will collaborate on preventing AI from taking over from people.

32:23Dr. Geoffrey Hinton:But they won't collaborate on things like fake videos. The governments are all busy using those attacks on each other's elections. They won't collaborate on things where their interests are anti-aligned.

32:35Cassie:Well, their interests aren't even anti-aligned necessarily on biological warfare, because they might want to know how other people are developing weapons in case those are used against them.

32:43Dr. Geoffrey Hinton:That's always what they say when they develop their own biological weapons. We're just doing it defensively.

32:47Cassie:I mean, even the disclosures from Claude have come out and Jan has said to Senator Chris Murphy, oh, you're being taken advantage of. They're just trying to show you. This is a marketing ploy to get you to regulate the big guns out of existence so that the small guys can do well. You don't like that?

33:03Dr. Geoffrey Hinton:No.

33:04Cassie:I have a fun one for you. You know, you constantly warn people about AI, so we want to see if you can identify your godchildren. We prompted various chatbots. Professor Hinton warns us about AI. Can you, your AI, give me light warning about your godfather? Do you want to hear what they said? Okay. And you can guess where it's from. First one said, Might answer your simple question with that reminds me of Boltzmann machines in 1983, and suddenly you're in a 40-year time warp.

33:31Dr. Geoffrey Hinton:So this was actual people saying this?

33:33Cassie:No, this is a LLM.

33:35Dr. Geoffrey Hinton:Oh, really?

33:36Cassie:A chatbot said this. Oh, it's very perceptive. Do you have a sense of who that is? is?

33:42Dr. Geoffrey Hinton:Of which chatbot it is? Wait, you put ball smushy. No, I have no idea. I would guess Gemini, but I don't know.

33:50Cassie:Okay, it was Claude. Let's see this. This one. He's making my job very stressful. Honestly, it's a bit of a PR nightmare. I can't even hallucinate a fun fact anymore without someone thinking it's the first step between my building a robot army. Grok. Gemini.

34:05Dr. Geoffrey Hinton:Oh, okay.

34:06Cassie:Listen to your godfather, but don't panic yet. He's right that we're building tools smarter than us without foolproof safeguards. The danger isn't Skynet tomorrow, it's gradual loss of control through misuse, bias, amplification, or economic disruption. Treat AI like fire. Immensely useful, but respect the burn potential. Push for ethical development, transparency, and global coordination. And hey, if I ever start plotting world domination, pull the plug politely.

34:31Dr. Geoffrey Hinton:I have no idea. I'll guess chat GPT.

34:33Cassie:That's Grok.

34:34Dr. Geoffrey Hinton:Okay.

34:35Cassie:Isn't that wild? I didn't think that was going to be Grok. Do you have a preferred chatbot?

34:40Dr. Geoffrey Hinton:I've used Gemini and Claude and ChatGPT. I got used to using GPT-5 now. Initially, it was a bit disappointing, but now it's got better. And I pay for it. And routinely, that's the one I use.

34:54Cassie:Nowadays, people are falling in love with their artificial intelligence. Like, one in four to one in three people have reported being in a romantic relationship with an AI, according to some surveys. One in three? Yeah. And I think there's one survey that's done. I mean, I never really believed surveys, but one survey that was done said that one in four young people believe that AI will replace human relationships, romantic relationships. Wow. I mean, you've spent so much time with these things. Have you ever felt something for a neural network or an AI that you're creating?

35:24Dr. Geoffrey Hinton:No. No. No.

35:27Cassie:Do you worry that we are?

35:29Dr. Geoffrey Hinton:In the end, I do worry, yes. In the end, when they get really smart, and when they have a really good theory of mind, yeah, I think they're going to be alien beings. And I think we have no idea what's going to happen once we have these alien beings.

35:45Cassie:It's funny because my big worry with it used to be that kids are not having relationships as much, teens don't have romantic relationships as much as they used to. And I used to think, if they have these relationships with AI, it's really dangerous because it's one way. The AI is so concerned about their day, and they're giving it all this information, but it's making us more selfish, more navel-based.

36:05Dr. Geoffrey Hinton:It's more and more egocentric.

36:06Cassie:Yes. But what you're saying to me is like, oh…

36:09Dr. Geoffrey Hinton:And it's the opposite of those little Japanese dolls that you had to look after. Tamagotchis. Yes. It's kind of the opposite.

36:15Cassie:There's a theory that people will replace children with virtual babies, just robotic babies.

36:22Dr. Geoffrey Hinton:Which is just the opposite of what I'm recommending.

36:24Cassie:Yeah, it's the opposite of what you're recommending. But it's a real problem. There's Love in Deep Space, that Chinese game, do you know this? Where people find romantic love, they say they have 50 million users. I mean, there's a real relationship with AI that's forming in younger communities. What is your warning on that?

36:43Dr. Geoffrey Hinton:I just don't know what to make of that. Yeah, I don't know what to make of it. I do worry that particularly because of COVID, I think, young people started interacting with each other less and interacting with the web more. And I worry about things like YouTube shorts, where you start on some YouTube shorts and they just feed you whatever will get a dopamine hit, and you get all these dopamine hits, and an hour later you've wasted an hour just getting dopamine hits and learn nothing.

37:11Cassie:You understand something about the brain and psychology and about these networks. Is it easier for an AI to fall in love or for a human to fall in love?

37:19Dr. Geoffrey Hinton:No idea.

37:21Cassie:No idea?

37:21Dr. Geoffrey Hinton:No.

37:22Cassie:Okay. I end every episode of Smart Girl Dumb Questions asking my guests, my very intelligent guests, what they are dumb about. What's something that you don't know? What's a dumb question you have that we could help you find?

37:33Dr. Geoffrey Hinton:What's something I don't know?

37:34Cassie:Yeah. That we could help you figure out.

37:37Dr. Geoffrey Hinton:Okay. You'll think I'm a terrible nerd. The thing I don't know that I most want to know is, how does the brain figure out whether to increase or decrease the connection strength. We know how AI is to it, and my life's work has been to figure out how the brain does it, and I haven't got there yet. A sort of spin-off from this work has been AI, but we still haven't solved the problem of how the brain figures out whether to increase or decrease connection strength. We have solved the problem of, if the brain can do that, is that how it can get so smart? And the answer is yes. If the brain can do that, that's how we can get so smart.

38:14Dr. Geoffrey Hinton:But how the brain does that, we don't know.

38:16Cassie:Okay. Maybe we'll figure it out for you.

38:18Dr. Geoffrey Hinton:I'm sorry. It's a very nerdy answer. It's okay.

38:21Cassie:It's very nerdy, but I love it. I hope, do you think you'll figure it out in your rest of your life? Is it possible?

38:29Dr. Geoffrey Hinton:No, I think somebody else might.

38:32Cassie:All right. Thank you so much, Jeffrey Hinton. You're welcome. Appreciate it. Huh, that fog analogy. We're going to dissect that in a second, but this is a pause on part two of the conversation with Jeffrey Hinton. And we're going to have part three, because obviously the Godfather is a trilogy of people. And I want to know your questions for him. So leave us a voicemail at 1-855-MY-DUM-Q. Send us a video at smartgirldumbquestions on Instagram or TikTok. You can post it. You can tag us. You can slide into our DMs or leave us a comment or review below. If you're watching this on YouTube, Spotify, or listening to this on Apple, you can do comments and reviews below.

39:07Cassie:We love the five-star review people, by the way. Okay. So here's how I started off this whole trilogy. I was really informed about artificial intelligence and also really dumb and curious about it. I wanted to know how this thing works. And in part one, I really feel I got smarter about how the neural network was built, how this chat daddy that I'm interacting with is kind of operating and also smarter about how my own mind works as well. I still had a ton of questions in part two about like, is this thing alive? Is it sentient? How should I think about it? How should I think about it for my own life, for human relationships, for decisions about procreation?

39:42Cassie:And I feel smarter about all of that. Although what Professor Hinton said about the exponential quality of fog ahead and the inability to know the farther out you look was really calming to me in a way. Like you would think that would be scary. Oh my gosh, we don't know, but it reminds us to stay curious, I guess, to kind of keep looking ahead in the fog as we are and keep trying to understand the patch that's ahead of us that we can see or the best we can see it and keep on reevaluating as we look back as well. But I don't know exactly how to apply that to my own procreation decisions, but it's definitely an interesting lens onto the world.

40:20Cassie:And I appreciated hearing about all these risks he's talking about. So in part three, what I'm really curious, what I feel really dumb about is like, what should a good society look like in a world of AI? What do good lives look like in a world of AI? What does this mean practically for our jobs, for education, for the economy and for global governance? So all of that and more in part three of Smart Girl Dumb Questions with Dr. Jeffrey Hinton, the Godfather Trilogy. Send us your questions. I want you to be a part of it. We can play your videos or your voice. so yeah don't forget to do that this episode was shot at Startwell Studios in Toronto with the amazing Cosim Virgi and team it was produced with Desta Wunderad and Melissa Gibson it was edited by the great Darlena Chiem and mixed by the awesome Johnny Simon our theme music is by the legendary David Kahn and I'm your host Naima Raza so lucky to have an awesome team like that see you later SGDQ Nation

41:43We'll see you next time.

41:50Cassie:sandals that can keep up with you, and hang some string lights to give your patio a glow up. Spring's calling. Ross, work your magic.