Godfather of AI WARNS: "It's Too Late To Stop This"

0:00 They call you the Godfather of AI.

0:03 Uh yes, they do.

0:04 Why do they call you that?

0:06 There weren't that many people who believed

0:08 that we could make neural networks work, artificial neural networks.

0:12 So, for a long time in AI, from the 1950s onwards,

0:16 there were kind of two ideas about how to do AI.

0:20 One idea was that sort of core of human intelligence was reasoning,

0:25 and to do reasoning you need to use some form of logic.

0:28 And so, AI had to be based around logic.

0:32 And in your head you must have something

0:34 like symbolic expressions that you manipulated with rules,

0:37 and that's how intelligence worked.

0:39 And things like learning or reasoning by analogy,

0:42 they'd all come later once we figured out how basic reasoning works.

0:46 There was a different approach, which is to say,

0:49 let's model AI on the brain, cuz obviously the brain makes us intelligent.

0:54 So, simulate a network of brain cells on a computer,

0:58 and try and figure out how you

1:00 would learn strengths of connections between brain cells,

1:03 so that it learned to do complicated things,

1:06 like recognize objects in images or recognize speech, or even do reasoning.

1:11 I pushed that approach for like 50 years.

1:13 Because so few people believed in it,

1:16 there weren't many good universities that had groups that did that.

1:20 So, if you did that, the best young students

1:24 who believed in that came and worked with you.

1:26 So, I was very fortunate in getting a whole lot of really good students.

1:30 Some of which have gone on to create

1:32 and played an instrumental role in creating platforms like OpenAI.

1:36 Yes, so Ilya Sutskever would be a a nice example.

1:40 A whole bunch of them.

1:41 Why did you believe that modeling it

1:43 off the brain was a more effective approach?

1:46 It wasn't just me who believed it.

1:47 Early on, von Neumann believed it, and Turing believed it.

1:52 And if either of those had lived,

1:54 I think AI would have had a very different history.

1:56 But they both died young.

1:58 You think AI would have been here sooner?

2:00 I think neural net the neural net approach would have

2:03 been accepted much sooner if either of them had lived.

2:07 In this season of your life, what mission are you on?

2:11 My main mission now is to warn people how dangerous AI could be.

2:17 Did you know that when you became the godfather of AI?

2:21 No, not really.

2:22 I was quite slow to understand some of the risks.

2:26 Some of the risks were always very obvious,

2:27 like people would use AI to make autonomous lethal weapons.

2:31 That is things that go around deciding by themselves who to kill.

2:34 Other risks, like the idea that they would one

2:37 day get smarter than us and maybe we'd become irrelevant,

2:42 I was slow to recognize that.

2:43 Other people recognized it 20 years ago.

2:46 I only recognized it a few years ago that that was

2:48 a real risk that was come might be coming quite soon.

2:52 How could you not have foreseen that if if with everything you know here about

2:57 cracking the ability for these computers to learn

3:00 similar to how humans learn and just,

3:02 you know, introducing any rate of improvement?

3:05 It's a very good question.

3:06 How could you not have seen that?

3:08 But remember neural networks 20 30 years ago

3:12 were very primitive in what they could do.

3:14 They were nowhere near as good as humans

3:16 at things like vision and language and speech recognition.

3:20 The idea that you have to now worry about it getting smarter than people,

3:23 that seemed silly then.

3:25 When did that change?

3:26 It changed for the general population when ChatGPT came out.

3:30 It changed for me when I realized

3:33 that the kinds of digital intelligences we're making

3:37 have something that makes them far superior

3:39 to the kind of biological intelligence we have.

3:42 If I want to share information with you, so I go off and I learn something.

3:46 Mhm.

3:47 And I'd like to tell you what I learned.

3:49 So I produce some sentences.

3:51 This is a rather simplistic model, but roughly right.

3:53 Your brain is trying to figure out,

3:55 how can I change the strength of connections between

3:56 neurons so I might have put that word next.

3:59 And so you'll do a lot of learning when a very surprising word comes.

4:02 And not much learning when it is a when it's very obvious word.

4:05 If I say fish and chips, you don't do much learning when I say chips.

4:09 But if I say fish and cucumber, you do a lot more learning.

4:11 You wonder, why did I say cucumber?

4:13 So that's roughly what's going on in your brain.

4:16 I'm predicting what's coming next.

4:18 That's how we think it's working.

4:19 Nobody really knows for sure how the brain works.

4:22 And nobody knows how it gets the information about whether you should

4:25 increase the strength of a connection or decrease the strength of a connection.

4:29 That's the crucial thing.

4:30 But what we do know now from AI

4:33 is that if you could get information about whether

4:37 to increase or decrease the connection strength so

4:39 as to do better whatever task you're trying to do,

4:42 then we could learn incredible things cuz that's

4:45 what we're doing now with artificial neural nets.

4:48 It's just we don't know for real brains how

4:50 they get that signal about whether to increase or decrease.

4:53 As we sit here today, what are the big concerns you have around safety of AI?

4:57 If we were to to list the the top couple that are

5:01 really front of mind and that we should be thinking about.

5:03 Um Can I have more than a couple?

5:05 Go ahead.

5:05 I'll write them all down and we'll go through them.

5:08 Okay, first of all, I want to make

5:09 a distinction between two completely different kinds of risk.

5:14 There's risks that come from people misusing AI.

5:18 Yeah.

5:18 And that's most of the risks and all of the short-term risks.

5:23 And then there's risks that come from AI

5:25 getting super smart and suddenly it doesn't need us.

5:28 Is that a real risk?

5:30 And I talk mainly about that second risk because lots of people say,

5:34 is that a real risk?

5:35 And yes, it is.

5:37 Now, we don't know how much of a risk it is.

5:39 We've never been in that situation before.

5:41 We've never had to deal with things smarter than us.

5:43 So really the thing about that existential threat is

5:48 that we have no idea how to deal with it.

5:51 We have no idea what it's going to look like.

5:53 And anybody who tells you they know just what's

5:55 going to happen and how to deal with it, they're talking nonsense.

5:57 So, we don't know how to estimate

5:59 the probability probabilities it'll replace us.

6:02 Um some people say it's like less than 1%.

6:05 My friend Jan LeCun, who was a postdoc with me, thinks, "No, no, no, no.

6:09 We're always going to be We build these things.

6:11 We're always going to be in control." We'll build them to be obedient.

6:15 And other people, like Yudkowsky, say, "No, no, no.

6:21 These things are going to wipe us out for sure.

6:22 If anybody builds it,

6:23 it's going to wipe us all out." And he's confident of that.

6:27 I think both of those positions are extreme.

6:30 It's very hard to estimate the probabilities in between.

6:32 If you had to bet on who was right out of your two friends, I simply don't know.

6:40 So, if I had to bet, I'd say the probability is in between,

6:43 and I don't know where to estimate it in between.

6:45 I often say 10 to 20% will wipe us out, but that's just gut,

6:51 based on the idea that we we're still making them, and we're pretty ingenious.

6:55 And the hope is that if enough

6:57 smart people do enough research with enough resources,

7:01 we'll figure out a way to build them so they'll never want to harm us.

7:05 Sometimes I think if we we talk about that second um path,

7:08 sometimes I think about nuclear bombs and the the invention

7:11 of the atomic bomb and how it compares.

7:13 Like how is this different because the atomic bomb came along

7:16 and I imagine a lot of people at that time thought,

7:18 "Our days are numbered." Uh yes, I was there.

7:20 We did.

7:21 Yeah.

7:22 But but but what's what We're still here.

7:25 We're still here, yes.

7:26 So, the atomic bomb was really only good for one thing,

7:30 and it was very obvious how it worked.

7:32 Even if you hadn't had the pictures of Hiroshima and Nagasaki,

7:36 it was obvious that it was a very big bomb that was very dangerous.

7:41 With AI, it's good for many, many things.

7:47 It's going to be magnificent in healthcare

7:48 and education and more or less any industry

7:51 that needs to use its data is going to be able to use it better with AI.

7:56 So, we're not going to stop the development.

8:00 You know, people say, "Well, why don't we just stop it now?" We're not going

8:03 to stop it cuz it's too good for too many things.

8:07 Also, we're not going to stop it cuz it's good for battle robots

8:09 and none of the countries that sell weapons are going to want to stop it.

8:14 Like the European regulations,

8:16 they have some regulations about AI and it's good they have some regulations,

8:20 but they're not designed to deal with most of the threats.

8:23 And in particular, the European regulations have a clause in them that say,

8:27 "None of these regulations apply to military uses of AI." So,

8:33 governments are willing to regulate regulate companies and people,

8:37 but they're not willing to regulate themselves.

8:40 It seems pretty crazy to me that they I go back and forth,

8:43 but if Europe has a regulation, but the rest of the world doesn't,

8:47 Yeah, put someone at a competitive disadvantage.

8:50 Yeah.

8:50 We're seeing this already.

8:50 I don't think people realize that when OpenAI release

8:52 a new model or a new piece of software in America,

8:56 they can't release it to the to Europe yet because of regulations here.

8:59 So, Sam Altman tweeted saying,

9:01 "Our new AI agent thing is available to everybody,

9:03 but it can't come to Europe yet because there's regulations." Yes.

9:07 What does that do that gives us a productive disadvantage?

9:10 Productive disadvantage?

9:11 What we need is I mean, at this point in history,

9:15 when we're about to produce things more intelligent than ourselves,

9:18 what we really need is a kind of world government that works run by intelligent,

9:23 thoughtful people, and that's not what we got.

9:27 So, free for all.

9:29 Well, that what we've got is sort of we've got capitalism,

9:35 which is done very nicely by us.

9:37 It has produced lots of goods goods and services for us.

9:40 But, these big companies, they're legally required to try and maximize profits.

9:47 And that's not what you want from the people developing this stuff.

9:51 So, let's do the risks then.

9:52 You talked about there's human risks and then there's

9:55 So, I've distinguished these two kinds of risk.

9:57 Let's talk about all the risks from bad human actors using AI.

10:01 There's cyber attacks.

10:05 So, between 2023 and 2024, they increased by about a factor of 12, 1,200%.

10:14 And that's probably because these large language models

10:17 make it much easier to do phishing attacks.

10:20 And for phishing attack for anyone that doesn't

10:21 know is It's they send you something saying,

10:25 uh, "Hi, I'm your friend John and I'm stuck in El Salvador.

10:29 Could you just wire this money?" That's one kind of attack.

10:32 But, the phishing attacks are really trying to get your login credentials.

10:36 And now with AI, they can clone my voice, my image.

10:38 do all that.

10:39 I'm struggling at the moment because there's a bunch

10:41 of AI scams on X and also Meta.

10:44 And there's one in particular on Meta, so Instagram, Facebook at the moment,

10:47 which is a paid advert, where they've taken my voice from the podcast,

10:50 they've taken the my mannerisms,

10:52 and they've made a new video of me encouraging people

10:54 to go and take part in this crypto Ponzi scam or whatever.

10:58 And we've been we know we spent weeks and weeks and weeks

11:00 and weeks and and emailing Meta telling them please take this down.

11:03 They take it down, another one pops up.

11:05 They take that one pops up.

11:06 So, it's like whack-a-mole.

11:07 It's very annoying.

11:09 The the heartbreaking part is you get the messages

11:10 from people that have fallen for the scam.

11:12 And they've lost 500 pounds or 500 dollars.

11:14 cross with you cuz you recommended it.

11:16 And I'm I'm like I'm as sad for them.

11:18 It's very annoying.

11:19 Yeah.

11:19 I have a a smaller version of that which is peo-

11:22 some people now publish papers with me as one of the authors.

11:26 Mhm.

11:27 And it looks like it's in order

11:28 that they can get lots of citations to themselves.

11:31 Ah.

11:32 So, cyber attacks are very real threat.

11:34 There's been an explosion of those.

11:35 And these already, obviously, AI is very patient,

11:39 so they can go through 100 million lines

11:41 of code looking for known ways of attacking them.

11:45 That's easy to do, but they're going to get more creative,

11:48 and they may, some people believe,

11:51 and I some people who know a lot believe that maybe by 2030,

11:57 they'll be creating new kinds of cyber attacks.

12:00 Which no person ever thought of.

12:03 So, it knows it will essentially know everything that humans know,

12:07 but more, because it will learn new things.

12:10 It will learn new things.

12:11 It would also see all sorts of analogies that people probably never saw.

12:16 So, for example, at the point when GPT-4 couldn't look on the web,

12:21 I asked it, "Why is a compost heap like an atom bomb?" Off you go.

12:28 I have no idea.

12:29 Exactly.

12:29 Excellent.

12:30 Most That's exactly what most people would say.

12:32 It said, "Well, the time scales are very different,

12:35 and the energy scales are very different." But then

12:38 it went on to talk about how a compost heap,

12:40 as it gets hotter, generates heat faster.

12:43 And an atom bomb, as it produces more neutrons, generates neutrons faster.

12:48 Mhm.

12:49 And so, they're both chain reactions,

12:51 but at very different time and energy scales.

12:53 And I believe GPT-4 had seen that during its training.

12:57 It had understood the analogy between a compost heap and an atom bomb.

13:00 And the reason I believe that is if you've only got a trillion connections,

13:04 remember you have 100 trillion, Mhm.

13:07 and you need to have thousands of times more knowledge than a person,

13:10 you need to compress information into those connections.

13:14 And to compress information, you need to see analogies between different things.

13:18 In other words, it needs to see all

13:20 the things that are chain reactions and understand the basic

13:23 idea of a chain reaction and code

13:24 that, and then code the ways in which they're different.

13:27 And just a more efficient way of coding

13:28 things than coding each of them separately.

13:32 So, it's seen many, many analogies.

13:34 Probably many analogies that people have never seen.

13:37 That's why I also think that people who say these things will never be creative,

13:41 they're going to be much more creative than us.

13:43 Because they're going to see all sorts of analogies we never saw.

13:46 And a lot of creativity is about seeing strange analogies.

13:50 People are somewhat romantic about the specialness of what it is to be human.

13:53 And you hear lots of people saying that it's very, very different.

13:55 It's a it's computer.

13:56 We are, you know, we're conscious.

13:58 We are creatives.

14:00 We we have these sort of innate

14:02 unique abilities that the computers will never have.

14:05 What do you say to those people?

14:06 I'd argue a bit with the innate.

14:08 Um So the first thing I say is we

14:14 have a long history of believing people are special.

14:17 And we should have learned by now.

14:19 We thought we were at the center of the universe.

14:21 We thought we were made in the image of God.

14:24 White people thought they were very special.

14:26 We just tend to want to think we're special.

14:30 My belief is that more or less everyone has

14:35 a completely wrong model of what the mind is.

14:37 Let's suppose I drink a lot or I drop some acid.

14:40 And not recommended.

14:42 And I say to you I have the subjective

14:46 experience of little pink elephants floating in front of me.

14:50 Most people interpret that as there's some

14:55 kind of inner theater called the mind.

14:59 And only I can see what's in my mind.

15:02 And in this inner theater there's a little pink elephants floating around.

15:07 So, in other words, what's happened is my perceptual system's gone wrong.

15:11 And I'm trying to indicate to you how it's

15:14 gone wrong and what it's trying to tell me.

15:16 And the way I do that is by telling you what would have

15:19 to be out there in the real world for it to be telling the truth.

15:26 And so these little pink elephants, they're not in some inner theater.

15:30 These little pink elephants are hypothetical things in the real world.

15:34 And that's my way of telling you how my perceptual system's telling me fibs.

15:39 So now let us do that with a chatbot.

15:41 Yeah.

15:42 Cuz I believe that current multimodal chatbots have subjective experiences.

15:47 And very few people believe that.

15:49 But I'll try and make you believe it.

15:51 So suppose I have a multimodal chatbot.

15:53 It's got a robot arm so it can point and it's got a camera so it can see things.

15:58 And I put an object in front of it and I say point at the object.

16:02 It goes like this.

16:03 No problem.

16:04 Then I put a prism in front of its lens.

16:07 And so then I put an object in front of it

16:09 and I say point at the object and it goes there.

16:12 Good.

16:12 And I say no, that's not where the object is.

16:15 The object's actually straight in front of you,

16:17 but I put a prism in front of your lens.

16:20 And the chatbot says, "Oh, I see.

16:22 The prism bent the light rays.

16:24 So um the object's actually there,

16:26 but I had the subjective experience that it was there." Mhm.

16:30 Now if the chatbot says that, it's using

16:32 the word subjective experience exactly the way people use them.

16:35 It's an alternative view of what's going on.

16:38 They're hypothetical states of the world which if they

16:41 were true would mean my perceptual system wasn't lying.

16:44 And that's the best way I can tell you

16:45 what my perceptual system's doing when it's lying to me.

16:48 Mhm.

16:48 Now we need to go further to do

16:51 with sentience and consciousness and feelings and emotions,

16:54 but I think in the end they're all going to be dealt with in a similar way.

16:57 There's no reason machines can't have them all.

16:59 But people say machines can't have feelings.

17:01 And people are curiously confident about that.

17:05 I've no idea why.

17:06 Suppose I make a battle robot and it's a little battle robot.

17:10 And it sees a big battle robot that's much more powerful than it.

17:15 It would be really useful if it got scared.

17:18 Mhm.

17:19 Now, when I get scared,

17:22 um various physiological things happen that we don't need

17:25 to go into, and those won't happen with the robot.

17:28 But all the cognitive things, like I better get the hell out of here,

17:31 and I better sort of change my way of thinking

17:35 so I focus and focus and focus and don't get distracted,

17:38 all of that will happen with robots, too.

17:42 People will build in things so that they when the circumstances

17:46 are such they should get the hell out of there, they get scared and run away.

17:49 They'll have emotions, then.

17:51 They won't have the physiological aspects,

17:53 but they will have all the cognitive aspects.

17:56 And I think it would be odd to say they're just simulating emotions.

17:59 No, they're really having those emotions.

18:00 The little robot got scared and ran away.

18:03 It's not running away because of adrenaline,

18:04 it's running away because of a sequence of sort

18:07 of neurological in its neural net processes happened, which

18:11 Which have the equivalent effect to adrenaline.

18:14 So, do you do And it's not just adrenaline, right?

18:16 There's a lot of cognitive stuff goes on when you get scared.

18:20 Yeah.

18:19 So, do you think that there is conscious AI?

18:24 And when I say conscious,

18:25 I mean that represents the same properties of consciousness that a human has.

18:30 There's two issues here.

18:31 There's a sort of empirical one and a philosophical one.

18:33 I don't think there's anything in principle

18:36 that stops machines from being conscious.

18:39 I'll give you a little demonstration of that before we carry on.

18:42 Suppose I take your brain, and I take one brain cell in your brain,

18:47 and I replace it by this a bit Black Mirror-like.

18:49 I replace it by a little piece of nanotechnology that's just the same size,

18:55 that behaves in exactly the same way when it gets pings from other neurons,

18:58 it sends out pings just as the brain cell would have.

19:01 So, the other neurons don't know anything's changed.

19:04 Okay, I've just replaced one of your brain

19:06 cells with this little piece of nanotechnology.

19:09 Would you still be conscious?

19:11 Yeah.

19:12 Now, you can see where this argument's going.

19:14 Yeah.

19:14 So, if you replaced all of them, as I replace them all,

19:17 at what point do you stop being conscious?

19:20 Well, people think of consciousness as this like

19:22 ethereal thing that exists maybe beyond the brain cells.

19:26 Yeah, well, people have a lot of crazy ideas.

19:30 Um people don't know what consciousness is,

19:32 and they often don't know what they mean by it.

19:35 Mhm.

19:35 And then they fall back on saying, well,

19:37 I know it cuz I've got it, and I can see that I've got it.

19:40 And they fall back on this theater model of the mind, which I think is nonsense.

19:45 What do you think of consciousness as if you had to try and define it?

19:47 Is it cuz I think of it is just like the awareness of myself, I don't know.

19:51 I think it's a term we'll stop using.

19:54 Suppose you want to understand how a car works.

19:57 Well, you know, some cars have a lot of oomph,

19:59 and other cars have a lot less oomph.

20:01 Like, an Aston Martin's got lots of oomph.

20:04 Mhm.

20:04 And a little Toyota Corolla doesn't have much oomph.

20:07 But oomph isn't a very good concept for understanding cars.

20:12 Um if you want to understand cars, you need to understand about electric engines

20:15 or petrol engines and how they work.

20:17 Mhm.

20:17 And it gives rise to oomph.

20:20 But oomph isn't a very useful explanatory concept.

20:22 It's a kind of essence of a car.

20:23 It's the essence of an Aston Martin.

20:25 Mhm.

20:25 But it doesn't explain much.

20:27 I think consciousness is like that.

20:29 And I think we'll stop using that term.

20:31 But I don't think there's anything any reason why a machine shouldn't have it.

20:35 If your view of consciousness is that it intrinsically involves self-awareness,

20:41 then the machine's got to have self-awareness.

20:43 It's got to have cognition about its own cognition and stuff.

20:46 But I'm a materialist through and through, Mhm.

20:51 and I don't think there's any reason why a machine shouldn't have consciousness.

20:55 If you love The Diary of a CEO brand,

20:57 and you watch this channel, please do me a huge favor,

21:00 become part of the 15% of the viewers

21:03 on this channel that have hit the subscribe button.

21:05 It helps us tremendously,

21:06 and the bigger the channel gets, the bigger the guests.