YT YouTube2Text

YouTube2Text → Transcripts → Godfather of AI: They Keep Silencing Me But I’m Trying to Warn Them!

Godfather of AI: They Keep Silencing Me But I’m Trying to Warn Them! — Transcript

by The Diary Of A CEO · 16,545 words · 2,762 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00They call you the Godfather of AI. So,
0:02what would you be saying to people about
0:04their career prospects in a world of
0:06super intelligence? Train to be a
0:07plumber. Really? Yeah.
0:10Okay, I'm going to become a plumber.
0:12Geoffrey Hinton is the Nobel
0:13Prize-winning pioneer whose
0:15groundbreaking work has shaped AI and
0:17the future of humanity. Why do they call
0:20you the Godfather of AI? Because there
0:21weren't many people who believed that we
0:23could model AI on the brain so that it
0:25learned to do complicated things like
0:27recognize objects in images or even do
0:29reasoning. And I pushed that approach
0:30for 50 years. And then Google acquired
0:32that technology. And I worked there for
0:3310 years on something that's now used
0:35all the time in AI. And then you left?
0:37Yeah. Why? So that I could talk freely
0:39at a conference.
0:40What did you want to talk about freely?
0:42How dangerous AI could be.
0:45I realized that these things will one
0:47day get smarter than us. And we've never
0:49had to deal with that. And if you want
0:50to know what life's like when you're not
0:51the apex intelligence, ask a chicken.
0:54So, there's a risks that come from
0:56people misusing AI. And then there's
0:58risks from AI getting super smart and
1:00suddenly it doesn't need us. Is that a
1:01real risk? Yes, it is. But they're not
1:03going to stop it because it's too good
1:04for too many things. What about
1:05regulations? They have some but they're
1:07not designed to deal with those sort of
1:08threats. Like the European regulations
1:10have a clause that say, "None of these
1:12apply to military uses of AI." Really?
1:14Yeah, it's crazy. One of your students
1:16left OpenAI. Yeah. He was probably the
1:19most important person behind the
1:20development of the early versions of
1:22ChatGPT. And I think he left because he
1:24had safety concerns. We should recognize
1:26that this stuff is an existential
1:27threat. And we have to face the
1:29possibility that unless we do something
1:31soon, we're near the end.
1:34So, let's do the risks and what we end
1:36up doing in such a world.
1:39This has always blown my mind a little
1:41bit. 53% of you that listen to this show
1:43regularly haven't yet subscribed to this
1:45show. So, could I ask you for a favor
1:47before we start? If you like this show
1:49and you like what we do here and you
1:50want to support us, the free simple way
1:51that you can do just that is by hitting
1:53the subscribe button. And my commitment
1:55to you is if you do that, then I'll do
1:57everything in my power, me and my team,
1:58to make sure that this show is better
2:00for you every single week. We'll listen
2:02to your feedback. We'll find the guest
2:03that you want me to speak to. And we'll
2:05continue to do what we do. Thank you so
2:07much.
2:11Geoffrey Hinton.
2:13They call you the Godfather of AI.
2:16Uh yes, they do.
2:17Why do they call you that?
2:19There weren't that many people who
2:21believed that we could make neural
2:23networks work, artificial neural
2:24networks. So, for a long time in AI,
2:27from the 1950s onwards,
2:30there were kind of two ideas about how
2:32to do AI.
2:34One idea was that sort of core of human
2:36intelligence was reasoning.
2:38And to do reasoning, you needed to use
2:40some form of logic.
2:42And so, AI had to be based around logic.
2:45And in your head, you must have
2:47something like symbolic expressions that
2:48you manipulated with rules. And that's
2:51how intelligence worked.
2:52And things like learning or reasoning by
2:54analogy, they'd all come later once we
2:56figured out how basic reasoning works.
2:59There was a different approach, which is
3:01to say,
3:02"Let's model AI on the brain cuz
3:05obviously the brain makes us
3:06intelligent. So, simulate a network of
3:10brain cells on a computer and try and
3:12figure out how you would learn strengths
3:14of connections between brain cells so
3:16that it learned to do complicated things
3:19like recognize objects in images or
3:21recognize speech or even do reasoning."
3:24I pushed that approach for like 50
3:25years.
3:27Because so few people believed in it,
3:30there weren't many good universities
3:32that had groups that did that. So, if
3:35you did that, the best young students
3:37who believed in that came and worked
3:38with you. So, I was very fortunate in
3:40getting a whole lot of really good
3:42students.
3:43Some of which have gone on to create and
3:46play an instrumental role in creating
3:48platforms like OpenAI. Yes, so Ilya
3:50Sutskever would be a
3:52a nice example. A whole bunch of them.
3:54Why did you
3:56believe that modeling it off the brain
3:57was a more effective approach?
3:59It wasn't just me believed it. Early on,
4:01von Neumann believed it.
4:04And Turing believed it. And if either of
4:06those had lived, I think AI would have
4:08had a very different history. But they
4:10both died young.
4:11You think AI would have been here
4:12sooner? I think neural net the neural
4:15net approach would have been accepted
4:18much sooner if either of them had lived.
4:20In this season of your life, what
4:23mission are you on?
4:24My main mission now is to warn people
4:28how dangerous AI could be.
4:30Did you know that when you
4:32became the Godfather of AI? No, not
4:35really. I was quite slow to understand
4:38some of the risks. Some of the risks
4:39were always very obvious like people
4:41would use AI to make autonomous lethal
4:43weapons.
4:44That is, things that go around deciding
4:46by themselves who to kill.
4:48Other risks, like the idea that they
4:49would one day get smarter than us
4:52and maybe we'd become irrelevant.
4:55I was slow to recognize that. Other
4:57people recognized it
4:5820 years ago. I only recognized a few
5:01years ago that that was a real risk that
5:03was come might be coming quite soon. How
5:05could you not have foreseen that if if
5:09with everything you know here about
5:10cracking the ability for these computers
5:12to learn similar to how humans learn and
5:15just, you know, introducing any rate of
5:17improvement? It's a very good question.
5:19How could you not have seen that? But
5:22remember neural networks 20, 30 years
5:24ago were very primitive in what they
5:27could do. They were nowhere near as good
5:28as humans but things like vision and
5:31language and speech recognition.
5:33The idea that you have to not worry
5:35about it getting smarter than people,
5:36that seemed silly then.
5:38When did that change? It changed for the
5:40general population when ChatGPT came
5:42out.
5:43It changed for me when I realized that
5:48the kinds of digital intelligences we're
5:50making have something that makes them
5:52far superior to the kind of biological
5:54intelligence we have.
5:55If I want to share information with you,
5:58so I go off and I learn something.
6:00And I'd like to tell you what I learned.
6:02So, I produce some sentences.
6:04This is a rather simplistic model but
6:05roughly right. Your brain is trying to
6:07figure out, "How can I change the
6:08strengths of connections between neurons
6:10so I might have put that word next?" And
6:12so, you'll do a lot of learning when a
6:13very surprising word comes. And not much
6:15learning when if it's a when it's a very
6:17obvious word. If I say fish and chips,
6:19you don't do much learning when I say
6:21chips. But if I say fish and cucumber,
6:23you do a lot more learning. You wonder,
6:25"Why did I say cucumber?"
6:26So, that's roughly what's going on in
6:28your brain. I'm predicting what's coming
6:30next.
6:31That's how we think it's working. Nobody
6:33really knows for sure how the brain
6:34works. And nobody knows how it gets the
6:37information about whether you should
6:39increase the strength of a connection or
6:40decrease the strength of a connection.
6:42That's the crucial thing.
6:44But what we do know now from AI
6:47is that if you could get information
6:49about whether to increase or decrease a
6:51connection strength so as to do better
6:53whatever task you're trying to do,
6:55then we could learn incredible things
6:57cuz that's what we're doing now with
6:59artificial neural nets.
7:01It's just we don't know for real brains
7:03how they get that signal about whether
7:04to increase or decrease.
7:06As we sit here today, what are the big
7:08concerns you have around safety of AI?
7:10If we were to to list the the top couple
7:14that are really front of mind and that
7:15we should be thinking about. Um Can I
7:17have more than a couple? Go ahead. I'll
7:19write them all down and we'll go through
7:20them. Okay, first of all, I want to make
7:22a distinction between two completely
7:25different kinds of risk.
7:27There's risks that come from people
7:29misusing AI. Yeah. And that's most of
7:32the risks
7:34and all of the short-term risks.
7:36And then there's risks that come from AI
7:38getting super smart and suddenly it
7:40doesn't need us.
7:41Is that a real risk?
7:43And I talk mainly about that second risk
7:45because lots of people say, "Is that a
7:47real risk?"
7:48And yes, it is.
7:50Now, we don't know how much of a risk it
7:52is. We've never been in that situation
7:54before. We've never had to deal with
7:55things smarter than us. So, really the
7:58thing about that existential threat is
8:01that we have no idea how to deal with
8:04it. We have no idea what it's going to
8:05look like. And anybody who tells you
8:07they know just what's going to happen
8:08and how to deal with it, they're talking
8:10nonsense. So, we don't know how to
8:11estimate the probability probabilities
8:14it'll replace us.
8:16Um some people say it's like less than
8:171%. My friend Jan LeCun, who was a
8:20postdoc with me, thinks, "No, no, no,
8:22no. We're always going to be We build
8:24these things. We're always going to be
8:25in control.
8:26We'll build them to be obedient."
8:29And
8:30other people,
8:31like Yudkowsky, say, "No, no, no. These
8:34things are going to wipe us out for
8:35sure. If anybody builds it, it's going
8:37to wipe us all out."
8:38And he's confident of that.
8:40I think both of those positions are
8:41extreme.
8:43It's very hard to estimate the
8:44probabilities in between. If you had to
8:46bet
8:47on who was right out of your two
8:48friends,
8:51I simply don't know. So, if I had to
8:53bet, I'd say the probability is in
8:55between.
8:56And I don't know where to estimate it in
8:57between. I often say 10 to 20% chance
9:00they'll wipe us out.
9:01But that's just
9:03gut. Based on the idea that we're we're
9:05still making them and we're pretty
9:07ingenious. And the hope is
9:10that if enough smart people do enough
9:11research with enough resources, we'll
9:14figure out a way to build them so
9:15they'll never want to
9:17harm us.
9:19Sometimes I think if we we talk about
9:20that second um path, sometimes I think
9:22about nuclear bombs and the the
9:23invention of the atomic bomb and how it
9:26compares. Like how is this different
9:28because the atomic bomb came along and I
9:29imagine a lot of people at that time
9:30thought our days are numbered. Oh yes, I
9:33was there. We did. Yeah. But but but
9:35what's what
9:37We're still here.
9:38We're still here, yes. So, the atomic
9:41bomb was really only good for one thing.
9:43And it was very obvious how it worked.
9:46Even if you hadn't had the pictures of
9:47Hiroshima and Nagasaki, it was obvious
9:50that it was a very big bomb
9:53that was very dangerous.
9:54With AI,
9:56it's good for many, many things. It's
10:00going to be magnificent in healthcare
10:01and education and more or less any
10:03industry that needs to
10:06use its data is going to be able to use
10:08it better with AI.
10:09So, we're not going to stop the
10:11development.
10:13You know, people say, "Well, why don't
10:14we just stop it now?" We're not going to
10:17stop it cuz it's too good for too many
10:19things.
10:20Also, we're not going to stop it cuz
10:21it's good for battle robots and none of
10:23the countries that sell weapons are
10:25going to want to stop it. Like the
10:28European regulations,
10:30they have some regulations about AI and
10:31it's good they have some regulations,
10:33but they're not designed to deal with
10:34most of the threats. And in particular,
10:37the European regulations have a clause
10:40in them that say, "None of these
10:41regulations apply to military uses of
10:43AI."
10:45So, governments are willing to regulate
10:47regulate
10:49companies and people, but they're not
10:50willing to regulate themselves.
10:53It seems pretty crazy to me that they
10:55I go back and forth, but if Europe has a
10:58regulation, but the rest of the world
10:59doesn't, Yeah, it puts them at a
11:01competitive disadvantage.
11:02Yeah. And we're seeing this already. I
11:04don't think people realize that when
11:05OpenAI release a new model or a new
11:07piece of software in America,
11:09they can't release it to the to Europe
11:11yet because of regulations here. So, Sam
11:13Altman tweeted saying, "Our new AI agent
11:15thing is available to everybody, but it
11:17can't come to Europe yet because there's
11:18regulations."
11:20Yes.
11:20What does that do? Does that give us a
11:21productive disadvantage? Productivity
11:23disadvantage? Right. What we need is I
11:26mean, at this point in history, when
11:28we're about to produce things more
11:29intelligent than ourselves, what we
11:32really need is a kind of world
11:34government that works run by
11:36intelligent, thoughtful people. And
11:38that's not what we got.
11:40So, free for all.
11:42Well, that what we've got is
11:45sort of
11:47we've got capitalism, which is done very
11:49nicely by us. It has produced lots of
11:51goods goods and services for us,
11:53but
11:54these big companies,
11:56they're legally required to try maximize
11:59profits.
12:00And that's not what you want from the
12:02people developing this stuff.
12:05So, let's do the risks then. You talked
12:06about there's human risks and then
12:08there's
12:08So, I've distinguished these two kinds
12:09of risk. Let's talk about all the risks
12:11from bad human actors using AI.
12:15There's cyber attacks.
12:18So, between 2023 and 2024,
12:22they increased by about a factor of 12,
12:241,200%.
12:27And that's probably because these large
12:29language models make it much easier to
12:31do phishing attacks.
12:33And a phishing attack for anyone that
12:34doesn't know is
12:35It's they send you something saying, uh,
12:39"Hi, I'm your friend John and I'm stuck
12:41in El Salvador. Could you just wire this
12:43money?" That's one kind of attack. But
12:45the phishing attacks are really trying
12:47to get your login credentials. And now
12:49with AI, they can clone my voice, my
12:51image.
12:52all that. I'm struggling at the moment
12:53because there's a bunch of AI scams on X
12:56and also Meta. And there's one in
12:57particular on Meta, so Instagram,
12:59Facebook at the moment, which is a paid
13:00advert where they've taken my voice from
13:03the podcast. They've taken the my
13:04mannerisms and they've made a new video
13:06of me encouraging people to go and take
13:08part in this crypto Ponzi scam or
13:11whatever. And we've been you know, we
13:12spent weeks and weeks and weeks and
13:14weeks and end emailing Meta telling,
13:15"Please take this down." They take it
13:17down, another one pops up. They take
13:18that one down, another one pops up. So,
13:20it's like whack-a-mole. Yeah, that's
13:21very annoying. The the heartbreaking
13:23part is you get the messages from people
13:24that have fallen for the scam.
13:25And they've lost 500 pounds or 500
13:27dollars or something.
13:28with you cuz you recommended it.
13:29And I'm I'm like I'm sad for them. It's
13:31very annoying. I have a a smaller
13:34version of that, which is peo- some
13:35people now publish papers
13:38with me as one of the authors. Mhm.
13:40And it looks like it's in order that
13:42they can get lots of citations to
13:43themselves. Ah.
13:46So, cyber attacks are a very real
13:47threat. There's been an explosion of
13:48those.
13:48And these already, obviously AI is very
13:52patient, so they can go through 100
13:54million lines of code looking for known
13:56ways of attacking them.
13:58That's easy to do, but they're going to
14:00get more creative and they may
14:02some people believe, and I
14:06some people who know a lot believe that
14:08maybe by 2030,
14:10they'll be creating new kinds of cyber
14:12attacks
14:13which no person ever thought of.
14:16So, that's very worrisome. Because they
14:18can think for themselves and discover
14:20new ways to attack.
14:21They can draw new conclusions from much
14:23more data than a person ever saw.
14:26Is there anything you're doing
14:28to protect yourself from cyber attacks
14:29at all? Yes. It's one of the few places
14:32where I changed what I do radically
14:35because I'm scared of cyber attacks.
14:37Canadian banks are extremely safe. In
14:402008, no Canadian banks came anywhere
14:43near going bust.
14:44So, they're very safe banks cuz they're
14:46well regulated, fairly well regulated.
14:49Nevertheless, I think a cyber attack
14:51might be able to bring down a bank.
14:53Now,
14:54if you have all my savings are in shares
14:57in banks,
14:58held by banks.
15:00So, if the bank
15:01gets attacked and it holds your shares,
15:04they're still your shares.
15:06And so, I think you'd be okay unless the
15:10attacker sells the shares cuz the bank
15:12can sell the shares.
15:13If the attacker sells your shares, I
15:16think you're screwed.
15:18I don't know I mean, maybe the bank
15:20would have to try and reimburse you, but
15:21the bank's bust by now, right? So,
15:24So, I'm worried about a Canadian bank
15:26being taken down by a cyber attack and
15:29the attacker selling selling shares that
15:31it holds.
15:32So, I spread my money my children's
15:34money between three banks
15:37in the belief that if a cyber attack
15:39takes down one Canadian bank,
15:41the other Canadian banks will very
15:43quickly get very careful.
15:46And do you have a phone that's not
15:47connected to the internet? Do you have
15:49any like you know, I'm thinking about
15:50storing data and stuff like that. Do you
15:52think it's wise to consider having cold
15:54storage? I have a little disk drive and
15:57I back up my laptop on this hard drive.
16:00So, I actually have everything on my
16:02laptop on a hard drive.
16:04At least, you know, if the whole
16:05internet went down, I had the sense I
16:07still got it on my laptop and I still
16:09got
16:10my information. Okay.
16:12Then the next thing is using AIs to
16:15create nasty viruses.
16:18Okay.
16:19And the problem with that is
16:21that just requires one crazy guy with a
16:25grudge. One guy who knows a little bit
16:27of molecular biology, knows a lot about
16:29AI,
16:30and just wants to destroy the world.
16:33You can now create
16:35new viruses relatively cheaply using AI.
16:39And you don't have to be a very skilled
16:41molecular biologist to do it. And that's
16:43very scary. So, you could have a small
16:44cult, for example.
16:47A small cult might be able to raise a
16:49few million dollars.
16:50For a few million dollars, they might be
16:52able to design a whole bunch of viruses.
16:54Well, I'm thinking about some of our
16:55foreign adversaries doing
16:57government-funded programs. I mean,
16:58there was lots of talk around COVID and
17:00the Wuhan laboratory and what they were
17:01doing in gain-of-function research, but
17:03I'm wondering if in, you know, a China
17:05or a Russia or an Iran or something,
17:08the government could fund a a program
17:10for a small group of scientists to make
17:11a virus that they could, you know,
17:13I think they could, yes. Now, they'd be
17:16worried about retaliation. They'd be
17:18worried about other governments doing
17:19the same to them. Hopefully, that would
17:20help keep it under control. They might
17:22also be worried about the virus
17:23spreading to their country. Okay.
17:26Then there's, um,
17:27corrupting elections.
17:30Okay.
17:31So, if you wanted to use AI to corrupt
17:33elections,
17:35a very effective thing is to be able to
17:37do targeted political advertisements
17:40where you know a lot about the person.
17:44So,
17:45anybody wanting to use AI for corrupting
17:47elections would try and get as much data
17:50as they could about everybody in the
17:52electorate. With that in mind, it's a
17:55bit worrying what Musk is doing at
17:57present in the States going in and
17:59insisting on getting access to all these
18:01things that were very carefully siloed.
18:03The claim is it's to make things more
18:05efficient, but it's exactly what you
18:07would want if you intended to corrupt
18:09the next election.
18:10How do you mean? Could you get all this
18:12data on the people?
18:12all this data on people. You know how
18:14much they make, where they live, you
18:15know everything about them. Once you
18:17know that, it's very easy to manipulate
18:19them.
18:20Because you can make an AI that You can
18:23send messages, um, that they'll find
18:25very convincing telling them not to
18:27vote, for example.
18:29So, I have no no
18:31reason other than common sense to think
18:33this, but I wouldn't be surprised if
18:36part of the motivation of getting all
18:38this data from American government
18:40sources
18:41is to corrupt elections. Another part
18:44might be that it's very nice training
18:46data for a big model.
18:48But he would have to be taking that data
18:50from the government and feeding it into
18:51his Yes. And what they've done is turned
18:54off lots of the security controls, got
18:56rid of the
18:58some of the organization to protect
18:59against that.
19:01Um, so that's corrupting elections.
19:03Okay. Then there's, um, creating these
19:07two echo chambers
19:09by organizations like YouTube
19:12and Facebook
19:15showing people things that will make
19:16them indignant. People love to be
19:19indignant.
19:20Indignant as in angry?
19:22Or what does indignant mean?
19:23Feeling I'm
19:25sort of angry, but feeling righteous.
19:27Okay. So, for example, if you were to
19:30show me something that said, "Trump did
19:33this crazy thing. Here's a video of
19:34Trump doing this completely crazy
19:36thing." I would immediately click on it.
19:40Okay, so putting us in echo chambers and
19:42dividing us. Yes. And that's, um, the
19:45policy that YouTube and Facebook and
19:48others
19:49use for deciding what to show you next
19:52is causing that.
19:55If they had a policy of showing you
19:57balanced things, they wouldn't get so
19:59many clicks and they wouldn't be able to
20:00sell so many advertisements.
20:02And so it's basically the profit motive
20:04is saying
20:05show them whatever will make them click.
20:07And what will make them click is
20:10things that are more and more extreme.
20:12And that confirm my existing bias. They
20:14confirm my existing bias. So you're
20:15getting your biases confirmed all the
20:17time. Further and further and further
20:19and further. Means you're you're driving
20:21away
20:21now there's in the states there's two
20:22communities that don't hardly talk to
20:24each other. I'm not sure people realize
20:26that this is actually happening every
20:27time they open an app. But if you go on
20:28a TikTok or a YouTube or one of these
20:30big social networks,
20:32the algorithm as you you said is
20:33designed to show you more of the things
20:36that you had interest in last time. So
20:38if you just play that out over 10 years,
20:40it's going to drive you further and
20:41further and further into whatever
20:43ideology or belief you have and further
20:45away from nuance and common sense and
20:48um parity, which is a pretty remarkable
20:51thing. That I like people don't know
20:52it's happening. They just open their
20:53phones and experience something and
20:56think this is the news or the experience
20:59everyone else is having.
21:00Right. So basically, if you have a
21:03newspaper and everybody gets the same
21:04newspaper, Yeah. you get to see all
21:06sorts of things you weren't looking for
21:08and you get a sense that if it's in the
21:10newspaper, it's an important thing or
21:12significant thing. But if you have your
21:13own news feed, my news feed on my
21:16iPhone, three quarters of the stories
21:19are about AI.
21:20And I find it very hard to know if the
21:23whole world's talking about AI all the
21:24time or if it's just my news feed.
21:28Okay, so driving me into my echo
21:30chambers, um which is going to continue
21:32to divide us further and further. I'm
21:34actually noticing that the algorithms
21:35are becoming even more
21:38what's the word?
21:40Tailored. And people might go that's
21:42great, but what it means is they're
21:43becoming even more personalized which
21:45was is means that my reality is becoming
21:47even further from your reality. Yeah,
21:49it's crazy. We don't have a shared
21:51reality anymore.
21:53I share reality with other people who
21:55watch the BBC and other BBC news and
21:58other people who read the Guardian and
21:59other people who read the New York
22:00Times.
22:02I have almost no shared reality with
22:04people who watch Fox News.
22:08It's pretty it's pretty um
22:09I I I
22:10It's worrisome. Yeah.
22:12Behind all this is the idea that these
22:14companies just want to make profit and
22:16they'll do whatever it takes to make
22:17more profit. Because they have to.
22:20They're legally obliged to that.
22:23So we almost can't blame the company,
22:25can we? If they're if that's
22:27Well,
22:28capitalism's done very well for us. It's
22:29produced lots of goodies. Yeah. But you
22:31need to have it very well regulated.
22:34So what you really want
22:36is to have rules so that when some
22:39company is trying to make as much profit
22:41as possible,
22:43in order to make that profit, they have
22:44to do things that are good for people in
22:46general, not things that are bad for
22:48people in general. So once you get to a
22:50situation where in order to make more
22:52profit, the company starts doing things
22:54that are very bad for society,
22:56like showing you things that are more
22:58and more extreme,
22:59that's what regulations are for.
23:01So you need regulations with capitalism.
23:04Now companies will always say
23:06regulations get in the way, make us less
23:09efficient, and that's true. The whole
23:11point of regulations is to stop them
23:12doing things to make profit that hurts
23:14society.
23:16And we need strong regulation. Who's
23:18going to decide whether it has society
23:19or not? Because, you know, That's the
23:21job of politicians. Unfortunately, if
23:24the politicians are owned by the
23:25companies, that's not so good. And also
23:27the politicians might not understand the
23:28technology. We you've probably seen the
23:30Senate hearings where they wheel out,
23:31you know, Mark Zuckerberg and these big
23:32tech CEOs. And it is quite embarrassing
23:34because they're asking the wrong
23:35questions.
23:37Well, I've seen the video of the US
23:40education secretary talking about how
23:42they're going to get AI in the
23:44classrooms, except she thought it was
23:46called A1.
23:48She's actually there saying we're going
23:49to have all the kids interacting with
23:51A1.
23:53There is a school system that's going to
23:54start um making sure that first graders
23:57or even pre-K's have A1 teaching, you
24:01know, every year starting, you know,
24:02that far down in the grades. And that's
24:04just a that's a wonderful thing.
24:10And these are what these are the people
24:11that These are the people in charge.
24:14Ultimately, the tech companies are in
24:15charge because they will outsmart
24:17the tech companies in the states now,
24:20at least a few weeks ago when I was
24:22there,
24:23they were running an advertisement about
24:26how it was very important not to
24:28regulate AI cuz it would hurt us in the
24:30competition with China. Yeah.
24:32And that's a that's a plausible
24:33argument, no? Yes, it will.
24:35But you have to decide.
24:37Do you want to compete with China
24:40by doing things that will
24:42do
24:43a lot of harm to your society?
24:46And you probably don't.
24:49I guess they would say that it's not
24:51just China, it's Denmark and Australia
24:53and Canada and
24:55Yeah, they're not they're not so worried
24:56about and Germany. But if they kneecap
24:58themselves with regulation, if they slow
24:59themselves down, then the founders, the
25:01entrepreneurs, the investors are going
25:02to go I think calling it kneecapping is
25:04uh taking a particular point of view.
25:07It's tak- taking the point of view that
25:08regulations are sort of very harmful.
25:10What you need to do is just constrain
25:13the big companies so that in order to
25:14make profit,
25:16they have to do things that are socially
25:17useful. Like Google search is a great
25:20example. That didn't need regulation
25:22because it just made information
25:24available to people. It was great.
25:26But then if you take YouTube which
25:28starts
25:29showing you adverts and showing you more
25:31and more extreme things, that needs
25:33regulation.
25:35But we don't have the people to regulate
25:36it.
25:37As we've identified.
25:38I think people know pretty well
25:40um that particular problem of showing
25:43you more and more extreme things. That's
25:44a well- known problem that the
25:45politicians understand.
25:47They just um need to get on and regulate
25:49it.
25:50So that was the the next point which was
25:52that the algorithms are going to drive
25:53us further into our echo chambers.
25:55Right.
25:56What's next? Lethal autonomous weapons.
25:59Lethal autonomous weapons.
26:03That means things that can kill you and
26:05make their own decision about whether to
26:07kill you.
26:08Which is the great dream, I guess, of
26:10the military-industrial complex. Being
26:13able to create such weapons.
26:15the worst thing about them is big
26:18powerful countries always have the
26:20ability to invade smaller poorer
26:23countries.
26:24They're just more powerful.
26:26But if you do that using actual
26:28soldiers,
26:29you get bodies coming back in bags
26:32and the relatives of the soldiers who
26:34were killed don't like it.
26:36So you get something like Vietnam.
26:39In the end there's a lot of protest at
26:40home.
26:41If instead of bodies coming back in
26:44bags, it was dead robots,
26:47there'd be much less protest and the
26:49military-industrial complex would like
26:51it much more cuz robots are expensive.
26:54And suppose you had something that could
26:56get killed and
26:58was expensive to replace, that would be
27:00just great.
27:01Big countries can invade small countries
27:03much more easily because they don't have
27:05their soldiers being killed.
27:07And the risk here is that
27:11these robots will
27:12malfunction or they'll just be more
27:14No, no. That's even if the robots do
27:16exactly what the people who built the
27:17robots want them to do,
27:19the risk is that it's going to make big
27:21countries invade small countries more
27:22often.
27:22More often because they can. And it's
27:24not a nice thing to do. So it brings
27:25down the friction of war. It brings down
27:27the cost of doing an invasion.
27:30And these machines will be smarter at
27:32warfare as well. So they'll be
27:34Well, even when the machines aren't
27:35smarter. So the lethal autonomous
27:37weapons, they can make them now.
27:40And they I think all the big defense
27:42firms are busy making them.
27:44Even if they're not smarter than people,
27:46they're still very nasty, scary things.
27:48Cuz I'm thinking that, you know, they
27:49could show just a picture, go get this
27:52guy.
27:53Yeah. And go take out anyone he's been
27:55texting.
27:56And this little wasp So two days ago, I
27:59was visiting a friend of mine in Sussex
28:01who had a drone that cost less than
28:03£200.
28:05And
28:07the drone went up, it took a good look
28:09at me,
28:10and then it could follow me through the
28:11woods.
28:13And it follow- it was very spooky having
28:14this drone. It was about 2 m behind me.
28:17It was looking at me.
28:18If I moved over there, it moved over
28:20there. It could just track me.
28:22For £200. But it was already quite
28:24spooky.
28:26Yeah, and I imagine there's as you say a
28:27race going on as we speak to who can
28:29build the most complex autonomous
28:31autonomous weapons.
28:33There is a a risk I often hear that some
28:35of these things will combine and the
28:38cyber attack will release weapons.
28:41Sure. Um you can you can get
28:44combinatorially many risks by combining
28:46these other risks.
28:48So I mean, for example, you could get a
28:50superintelligent AI
28:53that decides to get rid of people.
28:55And the obvious way to do that is just
28:56to make one of these nasty viruses.
28:58If you made a virus that was
29:01very contagious, very lethal, and very
29:04slow,
29:06everybody would have it before they
29:07realized what was happening.
29:09I mean, I think if a superintelligence
29:10wanted to get rid of us,
29:12it would probably go for something
29:13biological like that that wouldn't
29:14affect it. Do you not think it could
29:16just very quickly turn us against each
29:17other? For example, it could send a
29:19warning on the nuclear systems in
29:21America that there's a nuclear bomb
29:23coming from Russia
29:24or vice versa and one retaliates.
29:26Yeah. I mean, my basic view is there's
29:29so many ways in which a
29:30superintelligence could get rid of us.
29:32It's not worth speculating about.
29:35What what is What you have to do is
29:38prevent it ever wanting to. That's what
29:40we should be doing research on.
29:42There's no way we're going to prevent it
29:44from it's smarter than us, right?
29:46There's no way we're going to prevent it
29:47getting rid of us if it wants to.
29:50We're not used to thinking about things
29:51smarter than us.
29:53If you want to know what life's like
29:55when you're not the apex intelligence,
29:58ask a chicken.
30:03Yeah, I was thinking about my dog Pablo,
30:04my French bulldog, this morning as I
30:05left home.
30:07He has no idea where I'm going. He has
30:09no idea what I do. Right.
30:10I can't even talk to him.
30:12Yeah. And the get the intelligence gap
30:14will be like that. So, you're telling me
30:16that if I'm Pablo, my French bulldog,
30:18I need to figure out a way to make
30:21my owner
30:22not wipe me out.
30:24Yeah.
30:25So, we have one example of that, which
30:27is mothers and babies.
30:29Evolution put a lot of work into that.
30:31Mothers are smarter than babies, but
30:32babies are in control.
30:34And they're in control cuz the mother
30:35just can't bear Lots of hormones and
30:37things, but the baby The mother just
30:40can't bear the sound of the baby crying.
30:42Not all mothers. Not all mothers. And
30:44then the baby's not in control, and then
30:46bad things happen.
30:48We somehow need
30:50to figure out how to make them not want
30:52to take over. The analogy I often use is
30:55forget about intelligence, think about
30:57physical strength. Suppose you have a
30:59nice little tiger cub.
31:00It's sort of a bit bigger than a cat.
31:02It's really cute.
31:04It's very cuddly, very interesting to
31:06watch, except that you better be sure
31:08that when it grows up, it never wants to
31:10kill you, cuz if it ever wanted to kill
31:12you,
31:12you'd be dead in a few seconds.
31:15And you're saying that AI we have now is
31:16the tiger cub. Yep.
31:18And it's growing up. Yep.
31:21So, we need to train it as it's when
31:23it's a baby.
31:23a tiger has lots of innate stuff built
31:25in, so you know when it grows up, it's
31:27not a safe thing to have around. But
31:29lions, people that have lions as pets,
31:31Yes. sometimes the lion is affectionate
31:33to its creator, but not to others. Yes.
31:36And we don't know whether these AIs
31:40We We simply don't know whether we can
31:42make them not want to take over and not
31:43want to hurt us. Do you think we can? Do
31:45you think it's possible to train
31:47superintelligence?
31:48don't think it's clear that we can.
31:50So, I think it might be hopeless.
31:52But I also think
31:54we might be able to.
31:56And it'd be sort of crazy if people went
31:58extinct cuz we couldn't be bothered to
32:00try.
32:01If that's even a possibility, how do you
32:03feel about your life's work? Because you
32:05were
32:06Yeah.
32:07Um it's sort of takes the edge off it,
32:09doesn't it?
32:11I mean, the AI is going to be wonderful
32:12in healthcare, and wonderful in
32:13education,
32:15and wonderful I mean, it's going to make
32:16call centers much more efficient. Though
32:18one worries a bit about what the people
32:20who are doing that job now do. It makes
32:22me sad. I don't feel particularly guilty
32:25about developing AI like
32:2740 years ago,
32:29because
32:30at that time we had no idea that this
32:32stuff was going to happen this fast. We
32:34thought we had plenty of time to worry
32:36about things like that. They When you
32:38When you can't get the AI to do much,
32:40you want to get it to do a little bit
32:41more, you don't worry about
32:43this stupid little thing is going to
32:44take over from people. You just want it
32:46to be able to do a little bit more of
32:47the things people can do.
32:49It's not like I knowingly did something
32:53thinking, "This might wipe us all out,
32:55but I'm going to do it anyway." Mhm.
32:58But it is a bit sad that it's not just
33:00going to be something for good.
33:03So, I feel I have a duty now to talk
33:05about the risks.
33:07And if you could play it forward, and
33:08you could go forward 30, 50 years, and
33:09you found out that it led to the
33:10extinction of humanity,
33:13and if that does end up being the
33:17being the outcome,
33:21Well, if you played it forward and
33:22it led to the extinction of humanity,
33:25I would use that to tell
33:27people to tell their governments that we
33:29really have to work on how we're going
33:31to keep this stuff under control.
33:34I think we need people to tell
33:35governments that governments have to
33:37force the companies to use their
33:39resources to work on safety.
33:41And they're not doing much of that,
33:42because you don't make profits that way.
33:45One of your your students we talked
33:46about earlier, um Ilya? Yep. Ilya left
33:51OpenAI. Yep. And there was lots of
33:53conversation around the fact that he
33:55left because he had safety concerns.
33:57Yes. And he's gone on to set set up a AI
34:00safety company.
34:02Yes.
34:03Why do you think he left?
34:06I think he left cuz he had safety
34:07concerns. Really?
34:09Um I still have lunch with him from time
34:11to time. Oh, okay. His parents live in
34:13Toronto, and when he comes to Toronto,
34:14we have lunch together. He doesn't talk
34:16to me about what went on at OpenAI, so I
34:18have no inside information about that,
34:20but I know Ilya very well.
34:22And he is genuinely concerned with
34:23safety. So, I think that's why he left.
34:26Because he was one of the top people. I
34:27mean, he was He was probably the most
34:29important person behind the development
34:31of
34:32um ChatGPT.
34:34The The early versions like GPT-2, he
34:36was very important in the development of
34:37that. You know him personally, so you
34:39know his character.
34:41Yes. He has a good moral compass. He's
34:43not like someone like Musk who has no
34:45moral compass.
34:47Does Sam Altman have a good moral
34:48compass?
34:50We'll see.
34:53I don't know Sam, so I don't want to
34:56comment on that.
34:57But from what you've seen,
34:59are you concerned about the actions that
35:01they've taken?
35:02Cuz if you know Ilya, and Ilya's a good
35:04guy, and he's left,
35:06that would give you some insight, yes.
35:08It would give you some reason to believe
35:10that there's a problem there. And if you
35:12look at Sam's statements
35:15some years ago,
35:17he sort of happily said in one
35:20interview, "Um this stuff will probably
35:21kill us all." That's not exactly what he
35:23said, but that's what it amounted to.
35:25Now he's saying you don't need to worry
35:26too much about it.
35:28And I suspect that's not driven by
35:32seeking after the truth. That's driven
35:34by seeking after money.
35:36Is it money, or is it power?
35:39Yeah, I shouldn't have said money. It's
35:41It's some some combination of this, yes.
35:43Okay, I guess money's a proxy for power,
35:44but
35:45I I've got a friend who's a billionaire,
35:47and he is in those circles.
35:51And when I went to his house and had
35:53lunch with him one day, he knows lots of
35:54people in AI building the biggest AI
35:56companies in the world, and he gave me a
35:58cautionary warning across the across his
36:00kitchen table in London, where he gave
36:02me an insight into the private
36:03conversations these people have, not the
36:05media interviews they do where they talk
36:07about safety and all these things, but
36:09actually what some of these individuals
36:10think is going to happen.
36:12And what do they think's going to
36:13happen?
36:14It's not what they say publicly.
36:16You know, one one person who I should
36:18probably shouldn't name, who is the who
36:20is leading one of the biggest AI
36:21companies in the world, he told me that
36:23he knows this person very well, and he
36:24privately thinks that we're heading
36:26towards this kind of dystopian world
36:28where we have just huge amounts of free
36:30time, we don't work anymore,
36:32and this person doesn't really give a
36:33[ __ ] about the harm that it's going to
36:35have on the world. And this person who
36:36I'm referring to is building one of the
36:38biggest AI companies in the world.
36:39And I then watch this person's
36:41interviews online,
36:42I'm trying to figure out which of the
36:42three people it is.
36:43Yeah, well, it's one of those three
36:44people. Okay. And I watch this person's
36:46interviews online, and I I reflect on
36:47the conversation that my billionaire
36:49friend had with me, who knows him, and I
36:51go, "Fucking hell, this guy's lying
36:52publicly. Like, he's not telling the the
36:54truth to the world." And that's haunted
36:56me a little bit. It's part of the reason
36:57I have so many conversations around AI
36:59on this podcast, because I'm like, I
37:00don't know if they're
37:02I think they're a lit Some of them are a
37:04little bit sadistic about power.
37:06I think they they like the idea that
37:08they will change the world. That they
37:11will be the one that fundamentally
37:14shifts the world. I think Musk is
37:15clearly like that, right?
37:19He's such a complex character that I
37:21don't I don't really know how to place
37:22Musk. Um He's done some really good
37:24things like um pushing electric cars.
37:28That was a really good thing to do.
37:29Yeah. Some of the things he said about
37:31self-driving were a bit exaggerated, but
37:33he
37:34That was a really useful thing he did.
37:36Giving the Ukrainians communication
37:38during the war with Russia. Starlink,
37:40yeah.
37:41That was a really good thing he did.
37:43There's a bunch of things like that.
37:45Mhm. Um but he's also done some very bad
37:46things.
37:49So, coming back to this point of
37:53the possibility of
37:55destruction,
37:57and the motives of these big companies,
38:01are you at all hopeful that anything can
38:03be done to slow down the pace and
38:05acceleration of AI? Okay, there's two
38:07issues. One is, can you slow it down?
38:10Yeah. And the other is, can you make it
38:12so of it will be safe in the end? It
38:15won't wipe us all out.
38:17I don't believe we're going to slow it
38:18down.
38:20Yeah.
38:20And the reason I don't believe we're
38:21going to slow it down is because there's
38:22competition between countries, and
38:24competition between companies within a
38:26country,
38:27and all of that is making it go faster
38:29and faster.
38:30And if the US slowed it down, China
38:32wouldn't slow it down.
38:34Does
38:35Ilya think it's possible to make AI
38:37safe?
38:40I think he does. He won't tell me what
38:42his secret source is.
38:44I don't I'm not sure how many people
38:46know what his secret source is. I think
38:47a lot of the investors don't know what
38:48his secret source is, but they've given
38:50him billions of dollars anyway, cuz they
38:52have so much faith in Ilya, which isn't
38:54foolish. I mean,
38:56he was very important in AlexNet, which
38:59got object recognition working well. He
39:01was the main
39:03the main force behind the things like
39:05GPT-2,
39:07which then led to
39:08ChatGPT.
39:10So, I think having a lot of faith in
39:12Ilya is a very reasonable decision.
39:14There's something quite haunting about
39:15the guy that made and was the main force
39:18behind GPT-2, which led rise to this
39:20whole revolution, left the company
39:23because of safety reasons.
39:25He knows something that I don't know.
39:28About what might happen next.
39:29Well,
39:30the company had
39:32No, I don't know the precise details. Um
39:34but I'm fairly sure the company had
39:36indicated that would it would use a
39:38significant fraction of its resources
39:40of the compute time for doing safety
39:42research, and then it kept then it
39:45reduced that fraction. I think that's
39:47one of the things that happened. Yeah,
39:48that was reported publicly. Yes. Yeah.
39:51We've gotten to the autonomous weapons
39:54part of the risk framework. Right. So,
39:57the next one is joblessness. Yeah. In
40:00the past, new technologies have come in
40:03which didn't lead to joblessness. New
40:05jobs were created.
40:06So, the classic example people use is
40:08automatic teller machines. When
40:10automatic teller machines came in,
40:13a lot of bank tellers didn't lose their
40:14jobs. They just got to do more
40:16interesting things.
40:17But here,
40:19I think this is more like when they got
40:21machines in the Industrial Revolution,
40:24and
40:26you can't have a job digging ditches now
40:28because a machine can dig ditches much
40:30better than you can.
40:32And I think for mundane intellectual
40:34labor,
40:35AI is just going to replace everybody.
40:39Now, it will may well be in the form of
40:42you have fewer people using AI
40:45assistants. So, it's a combination of a
40:46person and an AI assistant, and they're
40:49doing the work that 10 people could do
40:51previously.
40:52People say that it will create new jobs,
40:54though. So, we'll be fine.
40:56Yes, and that's been the case for other
40:58technologies, but this is a very
40:59different kind of technology. If it can
41:01do all mundane human intellectual labor,
41:05then what new jobs is it going to
41:06create? You'd have You'd have to be very
41:09skilled to have a job that it couldn't
41:11just do.
41:12So, I don't I don't think they're right.
41:14I think you can try and generalize from
41:17other technologies that come in like
41:18computers or automatic teller machines,
41:21but I think this is different. People
41:23use this phrase. They say, AI won't take
41:25your job, a human using AI will take
41:27your job. Yes, I think that's true. But
41:29for many jobs,
41:31that will mean you need far fewer
41:32people.
41:33My niece answers letters of complaint to
41:36a health service.
41:38It used to take her 25 minutes. She'd
41:40read the complaint, and she'd think how
41:41to reply, and she'd write a letter, and
41:44now she just scans it into
41:47um a chatbot,
41:48and
41:50it writes the letter. She just checks
41:52the letter. Occasionally, she tells it
41:53to
41:54revise it in some ways.
41:56The whole process takes her 5 minutes.
41:59That means she can answer five times as
42:00many letters.
42:02And that means they need five times
42:04fewer of her.
42:06So, she can do the job that five of her
42:07used to do.
42:09Now,
42:10that will mean they need less people. In
42:13other jobs, like in health care,
42:16they're much more elastic. So, if you
42:19could make doctors five times as
42:20efficient, we could all have five times
42:22as much health care for the same price,
42:24and that would be great. There's There's
42:27almost no limit to how much health care
42:28people can absorb.
42:30They always want more health care if
42:32there's no cost to it.
42:34There are jobs where you can make a
42:36person with an AI assistant much more
42:38efficient, and you won't need to less
42:40people because you'll just have much
42:42more of that being done. But most jobs I
42:45think are not like that.
42:47Am I right in thinking this sort of
42:48Industrial Revolution
42:50would play a role in replacing muscles?
42:53Yes, exactly. And this revolution in AI
42:55replaces intelligence, the brain.
42:57Yeah. So, So, mundane intellectual labor
42:59is like having strong muscles, and
43:02it's not worth much anymore.
43:05So, muscles have been replaced. Now, we
43:06intelligence is being replaced.
43:08Yeah.
43:09So, what remains?
43:11Maybe for a while some kinds of
43:13creativity. But the whole idea of
43:15superintelligence is nothing remains.
43:17Um these things will get to be better
43:19than us at everything. So, what what do
43:20we end up doing in such a world?
43:22Well, if they work for us,
43:25we end up getting lots of goods and
43:27services for not much effort.
43:30Okay. But that sounds tempting and nice,
43:33but I don't know. There's a cautionary
43:35tale in creating more and more ease for
43:37humans in in it going badly. Yes, and
43:42we need to figure out if we can make it
43:44go well.
43:45So, the the nice scenario is imagine a
43:47company with a CEO
43:50who is very dumb,
43:52probably the son of the former CEO,
43:55and he has an executive assistant who's
43:57very smart,
43:59and he says,
44:01I think we should do this.
44:03And the executive assistant makes it all
44:04work.
44:05The CEO feels great. He doesn't
44:07understand that he's not really in
44:09control. And in In some sense, he is in
44:11control. He suggests what the company
44:13should do. She just makes it all work.
44:15Everything's great.
44:17That's the good scenario.
44:19And the bad scenario? The bad scenario
44:21is she thinks, why do we need him?
44:24Yeah.
44:26I mean, in a world where we have
44:28superintelligence, which you don't
44:29believe is that far away.
44:31Yeah, I think it might not be that far
44:33away. It's very hard to predict, but I
44:34think we might get it in like 20 years
44:37or even less.
44:38I made the biggest investment I've ever
44:40made in a company because of my
44:42girlfriend. I came home one night, and
44:44my lovely girlfriend was up at 1:00 a.m.
44:46in the morning pulling her hair out as
44:49she tried to piece together her own
44:51online store for her business. And in
44:54that moment, I remembered an email I'd
44:56had from a guy called John, the founder
44:58of Stan Store, our new sponsor, and a
45:00company I've invested incredibly heavily
45:02in. And Stan Store helps creators to
45:04sell digital products, courses,
45:06coaching, and memberships all through a
45:08simple customizable link-in-bio system.
45:11And it handles everything: payments,
45:12bookings, emails, community engagement,
45:14and even links with Shopify. And I
45:16believe in it so much that I'm going to
45:18launch a Stan Challenge. And as part of
45:22this challenge, I'm going to give away
45:24$100,000 to one of you. If you want to
45:26take part in this challenge, if you want
45:28to monetize the knowledge that you have,
45:30visit stevenbartlett.stan.store
45:33to sign up. And you'll also get an
45:35extended 30-day free trial of Stan Store
45:38if you use that link. Your next move
45:40could quite frankly change everything.
45:42Because I talked about ketosis on this
45:43podcast and ketones, a brand called
45:45Ketone-IQ sent me their little product
45:48here. It was on my desk when I got to
45:50the office. I picked it up. It sat on my
45:51desk for a couple of weeks. Then one
45:53day, I tried it.
45:55And honestly, I have not looked back
45:57ever since. I now have this everywhere I
46:00go. When I travel all around the world,
46:02it's in my hotel room. My team will put
46:03it there. Before I did the podcast
46:05recording today that I've just finished,
46:07I had a shot of Ketone-IQ. And as is
46:09always the case when I fall in love with
46:10a product, I called the CEO and asked if
46:12I could invest a couple of million quid
46:14into their company. So, I'm now an
46:16investor in the company as well as them
46:17being a brand sponsor. I find it so easy
46:20to drop into deep focused work when I've
46:22had one of these. I would love you to
46:24try one and see the impact it has on
46:26you, your focus, your productivity, and
46:27your endurance. So, if you want to try
46:29it today, visit ketone.com/steven for
46:3130% off your subscription. Plus, you'll
46:33receive a free gift with your second
46:35shipment. That's ketone.com/steven.
46:39I'm excited for you.
46:41I am.
46:42So, what's the difference between what
46:43we have now and superintelligence?
46:44Because it seems to be really
46:45intelligent to me when I use like
46:47ChatGPT-3 or Gemini or Okay. So, it's
46:50already AI is already better than us at
46:53a lot of things in particular areas
46:56like chess, for example. Yeah. AI is so
46:59much better than us that people will
47:01never beat those things again. Maybe the
47:03occasional win, but basically, they'll
47:05never be comparable again.
47:07Obviously, same in Go. In terms of the
47:09amount of knowledge they have,
47:12um something like GPT-4 knows thousands
47:14of times more than you do.
47:16There's a few areas in which your
47:18knowledge is better than it's.
47:20And
47:21almost all areas, it just knows more
47:22than you do. What areas am I better than
47:25it?
47:27Probably in interviewing CEOs. You're
47:31probably better at that.
47:33You've got a lot of experience at it.
47:34You're a good interviewer.
47:36You know a lot about it.
47:37If you tried If you got GPT-4 to
47:40interview a CEO, probably do a worse
47:42job.
47:43Okay.
47:46I'm trying to think if that if I agree
47:47with that statement. Uh GPT-4, I think,
47:50for sure. Yeah. Um but I But I guess you
47:52could train one on the how I ask Yeah, I
47:54guess you could train one on this how I
47:56ask questions and what I do and Sure.
47:58And if you took a general-purpose sort
48:01of foundation model, and then you
48:02trained it up on
48:04not just you, but every every interview
48:07you could find doing interviews like
48:09this,
48:10but especially you, it would probably
48:11get to be quite good at doing your job,
48:13but probably not as good as you for a
48:14while.
48:17Okay. So, there's a few areas left, and
48:19then superintelligence becomes when it's
48:22better than us at all things. When it's
48:23much smarter than you in almost all
48:25things, it's better than you. Yeah. And
48:27you you you say that this might be a
48:28decade away or so.
48:30Yeah, it might be. It might be even
48:32closer. Some people think it's even
48:34closer.
48:35It might well be much further. It might
48:36be 50 years away. That's still a
48:38possibility.
48:39It might be that somehow
48:42training on human data limits you to not
48:44be much smarter than humans. My guess is
48:46between 10 and 20 years we'll have
48:48superintelligence.
48:50On this point of joblessness, it is
48:51something I've been thinking a lot about
48:53in particular because I started messing
48:54around with AI agents, and we released
48:56an episode on the podcast actually this
48:57morning where we had a debate about AI
48:59agents with some CEO of a big AI agent
49:02company and a few other people.
49:04And it was the first moment where I had
49:06No, it was another moment where I had a
49:08eureka moment about what the future
49:10might look like. When I was able in the
49:12interview to tell this agent to order
49:14all of us drinks, and then 5 minutes
49:16later in the interview, you see the guy
49:17show up with the drinks, and I didn't
49:19touch anything. I just told it to order
49:21us drinks to the studio.
49:22And it didn't know about who you
49:24normally got your drinks from. It
49:25figured that out from the web. Yeah,
49:26figured it out cuz it went on Uber Eats.
49:28It has my my my data, I guess. And it I
49:31we put it on the screen in real time so
49:32everyone at home could see the agent
49:34going through the internet, picking the
49:35drinks, adding a tip for the driver,
49:38putting my address in, putting my credit
49:39card details in, and then the next thing
49:40you see is the drinks show up. So, that
49:43was one moment, and then the other
49:44moment was when
49:45I used a tool called Replit
49:48and I built software by just telling the
49:50agent what I wanted. Yes. It's amazing,
49:52right?
49:53It's amazing and terrifying at the same
49:55time. Yes. Because
49:57it can build software like that, right?
49:59Yeah. Remember that the AI, when it's
50:02training, is using code.
50:05And if it can modify its own code,
50:08then it gets quite scary, right? Cuz it
50:10can modify its own code.
50:11itself in a way we can't change
50:12ourselves.
50:14We can't change our innate endowment,
50:16right?
50:17There's nothing about itself that it
50:19couldn't change.
50:21On this point of joblessness, you have
50:22kids.
50:23I do. And they have kids?
50:25No, they don't have kids. No grandkids
50:27yet. What would you be saying to people
50:28about their career prospects in a world
50:31of super intelligence? What should we we
50:32be thinking about? Um in the meantime,
50:35I'd say it's going to be a long time
50:37before it's as good at physical
50:39manipulation as us. Okay. And so,
50:42a good bet would be to be a plumber.
50:47Until the humanoid robots show up.
50:49In such a world where there is mass
50:51joblessness, which is not something that
50:53you just predict, but this is something
50:54that Sam Altman at OpenAI, I've heard
50:56him predict and many of the CEOs and
50:58Elon Musk, I watched an interview which
51:00I'll play on screen of him being asked
51:01this question, and it's very rare that
51:03you see Elon Musk silent for 12 seconds
51:05or whatever it was. And then he
51:07basically says something about he
51:09actually is living in suspended
51:10disbelief. I he's basically just not
51:12thinking about it.
51:13When you think about advising your
51:14children on a career with so much that
51:16is changing,
51:18what do you tell them that's going to be
51:19of value?
51:33Well,
51:35that is a tough question to answer.
51:37I would just say, you know, to to sort
51:39of follow their heart in terms of what
51:41they they find um interesting to do or
51:43fulfilling to do.
51:45I mean, if I think about it too hard, it
51:46frankly can be uh just just
51:48disheartening and uh demotivating.
51:51Um
51:53because
51:54I mean, I I go through I I know I
51:57I've
51:58put a lot of blood, sweat, and tears
51:59into building the companies and then it
52:02and then I'm like, wait, well, like,
52:03should I be doing this? Because
52:06if I'm sacrificing time with friends and
52:08family that I would prefer to to to
52:11But but then, ultimately, the AI can do
52:13all these things.
52:14Does that make sense? I I don't know.
52:17Um
52:18to some extent, I have to have
52:19deliberate suspension of disbelief in
52:21order to be to remain motivated. Um
52:25so I I I guess I would say just, you
52:27know,
52:31work on things that you find
52:32interesting, fulfilling, and um
52:34and and that contribute uh some good to
52:36the rest of society. Yeah, a lot of
52:37these threats, it's very hard to
52:41intellectually, you can see the threat,
52:44but it's very hard to come to terms with
52:45it emotionally.
52:47Yeah.
52:48I I haven't come to terms with it
52:49emotionally yet.
52:50What do you mean by that?
52:53I haven't come to terms with
52:55what the development of super
52:57intelligence could do to my children's
52:59future.
53:01I'm okay. I'm 77.
53:04I'm going to be out of it
53:05Yeah, soon.
53:06But for my children and my my younger
53:09friends,
53:11my nephews and nieces,
53:13and their children,
53:14um
53:17I just don't like to think about what
53:19could happen.
53:23Why?
53:25Cuz it could be awful.
53:29In in what way?
53:32Well, if AI ever decided to take over,
53:35I mean, it would need people for a while
53:37to run the power stations
53:39until it
53:40designed better analog machines to run
53:41the power stations.
53:43There's so many ways it could get rid of
53:46people,
53:47all of which would, of course, be very
53:48nasty.
53:50Is that part of the reason you do what
53:51you do now?
53:53Yeah. I I mean, I think we should be
53:54making a huge effort right now
53:57to try and figure out if we can develop
53:59it safely.
54:00Are you concerned about the mid-term
54:02impact potentially on your nephews and
54:04your your kids in terms of their jobs as
54:06well? Yeah, I'm concerned about all
54:07that. Are there any particular
54:09industries that you think are most at
54:10risk? People talk about the creative
54:11industries a lot, and it's sort of
54:13knowledge work. They talk about lawyers
54:15and accountants and stuff like that.
54:17Yeah, so that's why I mentioned
54:18plumbers. I think plumbers are less at
54:20risk. Okay, I'm going to become a
54:21plumber.
54:21Someone like a legal assistant,
54:24a paralegal, Mhm. um they're not going
54:27to be needed for very long. And is there
54:29a wealth inequality issue here that will
54:31will
54:32arise from this?
54:33I think in a society which shared out
54:35things fairly,
54:37if you get a big increase in
54:39productivity,
54:40everybody should be better off. Mhm.
54:43But if you can replace lots of people by
54:46AIs,
54:48then the people who get replaced will be
54:50worse off
54:52and the company that supplies the AIs
54:55will be much better off
54:58and the company that uses the AIs.
55:01So, it's going to increase the gap
55:02between rich and poor. And we know that
55:05if you look at that gap between rich and
55:07poor, that basically tells you how nice
55:09a society is. If you have a big gap, you
55:12get very nasty societies in which people
55:14live in walled communities and put
55:16other people in mass jails.
55:20It's not good to increase the gap
55:21between rich and poor. The International
55:23Monetary Fund has expressed profound
55:25concerns that generative AI could cause
55:27massive labor disruptions and rising
55:29inequality and has called for policies
55:31that prevent this from happening.
55:33I read that in the Business Insider.
55:35Have they given any idea of what the
55:37policy should look like?
55:38No.
55:39Yeah, that's the problem. I mean, if AI
55:41can make everything much more efficient
55:42and get rid of people for most jobs
55:45or have a person assisted by AI doing
55:48many, many
55:49people's work, it's not obvious what to
55:52do about it. Universal basic income?
55:55Give everybody money? Yeah, I I I think
55:57that's a good start.
55:59And
56:01it stops people starving,
56:03but for a lot of people, their dignity
56:04is tied up with their job. I mean, who
56:06you think you are is tied up with you
56:08doing this job, right? Yeah.
56:10And
56:12if we said, we'll give you the same
56:13money just to sit around,
56:15that would impact your dignity.
56:18You said something earlier about it's
56:20surpassing or being superior to human
56:22intelligence. A lot of people, I think,
56:24like to believe that AI is is on a
56:27computer and it's something you can just
56:28turn off if you don't like it. Well, let
56:30me tell you why I think it's superior.
56:32Okay. Um it's digital.
56:35And because it's digital,
56:37you can have you can simulate a neural
56:39network on one piece of hardware. Yeah.
56:42And you can simulate exactly the same
56:43neural network on a different piece of
56:45hardware.
56:46Mhm. So, you can have clones of the same
56:48intelligence.
56:49Now, you could get this one to go off
56:52and look at one bit of the internet
56:54and this other one to look at a
56:55different bit of the internet. And while
56:57they're looking at these different bits
56:58of the internet,
57:00they can be syncing with each other, so
57:02they keep their weights the same. The
57:04connection strengths the same. Weights
57:05the connection strengths. Mhm. So, this
57:07one might look at something on the
57:08internet and say, oh, I'd like to
57:09increase this strength of this
57:11connection a bit.
57:12And it can convey that information to
57:14this one, so it can increase the
57:16strength of that connection a bit based
57:17on this one's experience. And when you
57:19say the strength of the connection,
57:21you're talking about learning. That's
57:23learning, yes. Learning consists of
57:24saying, instead of this one giving 2.4
57:27votes for whether that one should turn
57:28on, we'll have this one give 2.5 votes
57:31for whether this one should turn on.
57:33And that would be a little bit of
57:34learning. Mhm. So, these two different
57:36copies of the same neural net
57:39are getting different experiences.
57:41They're looking at different data, but
57:43they're sharing what they've learned by
57:44averaging their weights together. Mhm.
57:47And they can do that averaging at like a
57:49you can average a trillion weights.
57:51When you and I transfer information,
57:54we're limited to the amount of
57:55information in a sentence. And the
57:57amount of information in a sentence is
57:58maybe 100 bits. It's very little
58:00information. We're lucky if we're
58:02transferring like 10 bits a second. Mhm.
58:04These things are transferring trillions
58:06of bits a second. So, they're billions
58:08of times better than us at sharing
58:10information.
58:12And that's because they're digital and
58:14you can have two bits of hardware using
58:16the connection strengths in exactly the
58:17same way. We're analog and you can't do
58:20that. Your brain's different from my
58:21brain.
58:22And if I could see the connection
58:24strengths between all your neurons, it
58:26wouldn't do me any good cuz my neurons
58:28work slightly differently and they're
58:29connected up slightly differently. Mhm.
58:31So, when you die,
58:33all your knowledge dies with you.
58:35When these things die, suppose you take
58:37these two digital intelligences that are
58:39clones of each other,
58:40and you destroy the hardware they run
58:42on.
58:43As long as you've stored the connection
58:44strengths somewhere, you can just build
58:46new hardware
58:48that executes the same instructions, so
58:50it'll know how to use those connection
58:52strengths, and you've recreated that
58:54intelligence. So, they're immortal.
58:56We've actually solved the problem of
58:58immortality,
58:59but it's only for digital things.
59:02So, it knows
59:03it will essentially know everything that
59:06humans know, but more, because it will
59:07learn new things.
59:09It will learn new things. It will also
59:11see all sorts of analogies that people
59:13probably never saw.
59:15So, for example,
59:17at the point when GPT-4 couldn't look on
59:19the web,
59:20I asked it, why is a compost heap like
59:23an atom bomb?
59:25Off you go. I have no idea.
59:28Exactly. Excellent. Most That's exactly
59:30what most people would say. It said,
59:32"Well, the time scales are very
59:33different and the energy scales are very
59:35different.
59:37But then it went on to talk about how a
59:38compost heap, as it gets hotter,
59:40generates heat faster.
59:42And an atom bomb, as it produces more
59:44neutrons, generates neutrons faster.
59:47Mhm. And so they're both chain
59:48reactions, but at very different time
59:50and energy scales.
59:52And I believe GPT-4 had seen that during
59:54its training.
59:56It had understood the analogy between a
59:58compost heap and an atom bomb. And the
1:00:00reason I believe that is, if you've only
1:00:02got a trillion connections, remember you
1:00:04have 100 trillion, Mhm. and you need to
1:00:06have thousands of times more knowledge
1:00:08than a person,
1:00:09you need to compress information into
1:00:11those connections.
1:00:13And to compress information, you need to
1:00:15see analogies between different things.
1:00:17In other words, it needs to see all the
1:00:19things that are chain reactions and
1:00:21understand the basic idea of a chain
1:00:22reaction and code that, and then code
1:00:24the ways in which they're different. And
1:00:26that's just a more efficient way of
1:00:27coding things than coding each of them
1:00:29separately. Mhm.
1:00:31So, it's seen many, many analogies,
1:00:33probably many analogies that people have
1:00:35never seen.
1:00:36That's why I also think that people who
1:00:38say these things will never be creative,
1:00:39they're going to be much more creative
1:00:41than us.
1:00:42Because they're going to see all sorts
1:00:43of analogies we never saw. And a lot of
1:00:45creativity is about seeing strange
1:00:47analogies.
1:00:49People are somewhat romantic about the
1:00:50specialness of what it is to be human.
1:00:52And you hear lots of people saying, "Oh,
1:00:53it's very, very different. It's a it's a
1:00:55computer. We are, you know, we're
1:00:56conscious. We are creative. We we have
1:00:59these sort of innate, unique abilities
1:01:02that the computers will never have."
1:01:04What do you say to those people? I'd
1:01:05argue a bit with the innate. Um
1:01:09So,
1:01:11the first thing I say is we have a long
1:01:13history of believing people are special.
1:01:16And we should have learned by now. We
1:01:18thought we were at the center of the
1:01:19universe. We thought we were made in the
1:01:21image of God.
1:01:23White people thought they were very
1:01:24special. Mhm. We just tend to want to
1:01:27think we're special. Mhm.
1:01:29My belief is
1:01:31that more or less everyone
1:01:33has a completely wrong model of what the
1:01:35mind is.
1:01:36Let's suppose I drink a lot or I drop
1:01:38some acid, Mhm. and not recommended,
1:01:41and I
1:01:43say to you,
1:01:44"I have the subjective experience of
1:01:46little pink elephants floating in front
1:01:47of me." Mhm.
1:01:49Most people
1:01:51interpret that as
1:01:53there's some kind of inner theater
1:01:55called the mind,
1:01:58and only I can see what's in my mind.
1:02:01And in this inner theater,
1:02:03there's a little pink elephants floating
1:02:04around. Mhm.
1:02:06So, in other words, what's happened is
1:02:07my perceptual system's gone wrong,
1:02:10and I'm trying to indicate to you how
1:02:12it's gone wrong and what it's trying to
1:02:14tell me.
1:02:15And the way I do that is by telling you
1:02:17what would have to be out there in the
1:02:19real world
1:02:21for it to be telling the truth.
1:02:24And so these little pink elephants,
1:02:26they're not in some inner theater.
1:02:29These little pink elephants are
1:02:30hypothetical things in the real world.
1:02:33And that's my way of telling you how my
1:02:35perceptual system's telling me fibs.
1:02:38So, now I must do that with a chatbot.
1:02:39Yeah.
1:02:41Cuz I believe that current multimodal
1:02:43chatbots have subjective experiences.
1:02:46And very few people believe that.
1:02:48But I'll try and make you believe it.
1:02:50So, suppose I have a multimodal chatbot.
1:02:52It's got a robot arm, so it can point,
1:02:54and it's got a camera, so it can see
1:02:55things.
1:02:57And I put an object in front of it,
1:02:59and I say, "Point at the object."
1:03:00It goes like this. No problem.
1:03:03Then I put a prism in front of its lens.
1:03:06And so then I put an object in front of
1:03:07it, and I say, "Point at the object."
1:03:09And it goes there.
1:03:11Good.
1:03:11And I say, "No, that's not where the
1:03:13object is. The object's actually
1:03:15straight in front of you, but I put a
1:03:17prism in front of your lens."
1:03:19And the chatbot says, "Oh, I see. The
1:03:21prism bent the light rays. So, um the
1:03:24object's actually there, but I had the
1:03:26subjective experience that it was
1:03:27there."
1:03:28Mhm. Now, if the chatbot says that, it's
1:03:31using the word subjective experience
1:03:32exactly the way people use them. It's an
1:03:35alternative view of what's going on.
1:03:37They're hypothetical states of the
1:03:38world,
1:03:40which if they were true would mean my
1:03:41perceptual system wasn't lying. And
1:03:43that's the best way I can tell you what
1:03:44my perceptual system's doing when it's
1:03:46lying to me. Mhm. Now,
1:03:49we need to go further to deal with
1:03:50sentience and consciousness and feelings
1:03:51and emotions, but I think in the end
1:03:53they're all going to be dealt with in a
1:03:54similar way. There's no reason machines
1:03:56can't have them all.
1:03:58But people say machines can't have
1:03:59feelings.
1:04:00And people are curiously confident about
1:04:03that. I've no idea why. Suppose I make a
1:04:05battle robot, and it's a little battle
1:04:08robot,
1:04:09and it sees a big battle robot
1:04:11that's much more powerful than it.
1:04:14It would be really useful if it got
1:04:15scared.
1:04:17Mhm.
1:04:18Now,
1:04:19when I get scared, um various
1:04:22physiological things happen that we
1:04:23don't need to go into, and those won't
1:04:25happen with the robot.
1:04:27But all the cognitive things, like I
1:04:28better get the hell out of here, Yeah.
1:04:30Mhm. and I better sort of
1:04:32change my way of thinking, so I focus
1:04:35and focus and focus and I get
1:04:36distracted,
1:04:37all of that will happen with robots,
1:04:39too.
1:04:41People will build in things so that
1:04:43they, when it the circumstance is such
1:04:45they should get the hell out of there,
1:04:46they get scared and run away.
1:04:48They'll have emotions then.
1:04:50They won't have the physiological
1:04:51aspects, but they will have all the
1:04:53cognitive aspects.
1:04:55And I think it would be odd to say
1:04:56they're just simulating emotions. No,
1:04:58they're really having those emotions.
1:04:59The little robot got scared and ran
1:05:00away.
1:05:02It's not running away because of
1:05:03adrenaline, it's running away because of
1:05:05a sequence of sort of neurological in
1:05:07its neural net processes happened, which
1:05:10means which have the equivalent effect
1:05:11to adrenaline.
1:05:13So, do you do you think
1:05:14just adrenaline, right? There's a lot of
1:05:15cognitive stuff goes on when you get
1:05:16scared. Yeah.
1:05:18So, do you think that
1:05:21there is conscious AI?
1:05:23And when I say conscious, I mean
1:05:25that represents the same properties of
1:05:27consciousness that a human has.
1:05:29There's two issues here. There's a sort
1:05:30of empirical one and a philosophical
1:05:32one. I don't think there's anything in
1:05:34principle that stops machines from being
1:05:36conscious.
1:05:38I'll give you a little demonstration of
1:05:39that before we carry on. Mhm. Suppose I
1:05:41take your brain,
1:05:43and I take one brain cell in your brain,
1:05:46and I replace it by, it's a bit Black
1:05:48Mirror-like, I replace it by a little
1:05:50piece of nanotechnology that's just the
1:05:52same size,
1:05:54that behaves in exactly the same way
1:05:56when it gets pings from other neurons.
1:05:57It sends out pings just as the brain
1:05:59cell would have.
1:06:00So, the other neurons don't know
1:06:01anything's changed.
1:06:03Okay. I've just replaced one of your
1:06:05brain cells with this little piece of
1:06:06nanotechnology. Would you still be
1:06:08conscious?
1:06:10Yeah.
1:06:11Now you can see where this argument's
1:06:12going. Yeah. So, if you replaced all of
1:06:14them,
1:06:15as I replace them all, at what point do
1:06:16you stop being conscious?
1:06:19Well, people think of consciousness as
1:06:20this like ethereal thing that exists
1:06:23maybe beyond the brain cells. Yeah,
1:06:25well, people have a lot of crazy ideas.
1:06:29Um People don't know what consciousness
1:06:31is, and they often don't know what they
1:06:32mean by it. Mhm. And then they fall back
1:06:35on saying, "Well,
1:06:36I know it cuz I've got it, and I can see
1:06:38that I've got it." And they fall back on
1:06:40this theater model of the mind, which I
1:06:42think is nonsense.
1:06:43What do you think of consciousness as if
1:06:45you had to try and define it? Is it cuz
1:06:46I think of it as just like the awareness
1:06:48of myself? I don't know.
1:06:50I think it's a term we'll stop using.
1:06:53Suppose you want to understand how a car
1:06:54works.
1:06:56Well, you know some cars have a lot of
1:06:57oomph, and other cars have a lot less
1:06:59oomph. Like an Aston Martin's got lots
1:07:01of oomph. Mhm. And a little Toyota
1:07:04Corolla doesn't have much oomph.
1:07:06But oomph isn't a very good concept for
1:07:08understanding cars.
1:07:10Um if you want to understand cars, you
1:07:12need to understand about electric
1:07:13engines or petrol engines and how they
1:07:15work.
1:07:16And it gives rise to oomph.
1:07:18But oomph isn't a very useful
1:07:20explanatory concept. It's a kind of
1:07:21essence of a car. It's the essence of an
1:07:23Aston Martin.
1:07:24But it doesn't explain much. I think
1:07:26consciousness is like that.
1:07:28And I think we'll stop using that term.
1:07:30But I don't think there's anything any
1:07:32reason why a machine shouldn't have it.
1:07:34If
1:07:36your view of consciousness is that it
1:07:37intrinsically involves self-awareness,
1:07:40then the machine's got to have
1:07:41self-awareness. It's got to have
1:07:42cognition about its own cognition and
1:07:44stuff.
1:07:45But
1:07:47I'm a materialist through and through,
1:07:50and I don't think there's any reason why
1:07:52a machine shouldn't have consciousness.
1:07:54Do you think they do then have the same
1:07:56consciousness that we think of ourselves
1:07:58as being uniquely uh
1:08:01given as a gift when we're born?
1:08:03I'm ambivalent about that at present.
1:08:06So,
1:08:08I don't think there's this hard line. I
1:08:10think as soon as you have a machine that
1:08:12has some self-awareness,
1:08:14it's got some consciousness.
1:08:16Um I think it's an emergent property of
1:08:19a complex system.
1:08:21It's not a sort of essence that's
1:08:24throughout the universe. It's you make
1:08:26this really complicated system that's
1:08:27complicated enough to have a model of
1:08:29itself,
1:08:30and it does perception.
1:08:32And I think
1:08:34then you're beginning to get a conscious
1:08:36machine. So, I don't think there's any
1:08:37sharp distinction between what we've got
1:08:39now and conscious machines. I don't
1:08:41think it's going to one day we're going
1:08:42to wake up and say, "Hey, if you put
1:08:45this special chemical in, it becomes
1:08:46conscious." It's not going to be like
1:08:48that.
1:08:49I think we all wonder if these computers
1:08:50are like thinking like we are
1:08:53on their own when we're not there, and
1:08:55if they're experiencing emotions, if
1:08:56they're contending with I we I think we
1:08:58probably, you know, we think about
1:08:59things like love and things that feel
1:09:01unique to biological species.
1:09:04Um are they sat there thinking?
1:09:06Are they do they have concerns?
1:09:08I think they really are thinking.
1:09:10And I think as soon as you make AI
1:09:11agents, they will have concerns. If you
1:09:14want to make an effective AI agent,
1:09:16suppose you let's take a call center.
1:09:18Mhm. In a call center, you have people
1:09:20at present.
1:09:22They have all sorts of emotions and
1:09:23feelings,
1:09:24which are kind of useful. So, suppose I
1:09:27call up the call center
1:09:30and I'm actually lonely and I don't
1:09:32actually want to know the answer to why
1:09:35my computer isn't working. I just want
1:09:36somebody to talk to.
1:09:38After a while, the person in the call
1:09:41center
1:09:42will either get bored or get annoyed
1:09:44with me
1:09:45and will terminate it.
1:09:47Well, you replace them by an AI agent.
1:09:50The AI agent needs to have the same kind
1:09:52of responses. If someone's just called
1:09:54up cuz they just want to talk to the AI
1:09:55agent and we're happy to talk for whole
1:09:57the whole day to the AI agent, that's
1:09:59not good for business and you want an AI
1:10:02agent that either gets bored or gets
1:10:03irritated and says, "I'm sorry, but I
1:10:05don't have time for this." Then
1:10:07once it does that, I think it's got
1:10:09emotions.
1:10:11Now,
1:10:12like I say, emotions have two aspects to
1:10:15them. There's the cognitive aspect and
1:10:17the behavioral aspect and then there's a
1:10:19physiological aspect and these go
1:10:21together with us
1:10:23and if the AI agent gets embarrassed, he
1:10:26won't go red. Yeah. Um So, there's no
1:10:28physiological
1:10:29won't start sweating. Yeah. But it might
1:10:31have all the same behavior and in that
1:10:32case I'd say, "Yeah, it's having emotion
1:10:34It's got an emotion." So, it's going to
1:10:36have the same sort of cognitive thought
1:10:38and then it's going to act upon that
1:10:39cognitive thought.
1:10:40way, but without the physiological
1:10:42responses. And does that matter that it
1:10:45doesn't go red in the face and it's just
1:10:46a different I mean, that's a response to
1:10:48the
1:10:48it somewhat different from us. Yeah. For
1:10:50some things, the physiological aspects
1:10:53are very important like love.
1:10:55They're a long way from having love the
1:10:56same way we do.
1:10:58But I don't see why they shouldn't have
1:11:00emotions.
1:11:01So, I think what's happened is people
1:11:04have a model of how the mind works and
1:11:07what feelings are and what emotions are
1:11:09and their model is just wrong.
1:11:12What um what brought you to Google?
1:11:15You You worked at Google for about a
1:11:16decade, right? Yeah. What brought you
1:11:18there?
1:11:19I have a
1:11:21son who has learning difficulties
1:11:23and in order to be sure he would never
1:11:26be out on the street
1:11:28I needed to get several million dollars
1:11:32and I wasn't going to get that as an
1:11:33academic.
1:11:34I tried. So, I taught a Coursera course
1:11:37in the hope that I'd make lots of money
1:11:39that way, but there was no money in
1:11:40that.
1:11:41So, I figured out, well,
1:11:43the only way to get millions of dollars
1:11:46is to sell myself to a big company.
1:11:51And so, when I was 65
1:11:54fortunately for me, I had two brilliant
1:11:56students who produced something called
1:11:58AlexNet, which was neural net that was
1:12:01very good at recognizing objects in
1:12:02images.
1:12:04And
1:12:05so,
1:12:07Ilya and Alex and I
1:12:09set up a little company and auctioned
1:12:11it.
1:12:12And we actually set up an auction where
1:12:13we had a number of big companies bidding
1:12:15for us.
1:12:17And that company was called AlexNet. No,
1:12:21the the network that recognized objects
1:12:24was called AlexNet. Company was called
1:12:26DNN Research, deep neural network
1:12:28research.
1:12:29And it was doing things like this. I'll
1:12:30put this graph up on the screen.
1:12:31That's AlexNet. This picture shows eight
1:12:34images and AlexNet's ability, which is
1:12:38your company's ability to spot what was
1:12:40in those images. Yeah.
1:12:42So, it could tell the difference between
1:12:43various kinds of mushroom
1:12:45and about 12% of ImageNet is dogs
1:12:49and to be good at ImageNet, you have to
1:12:51tell the difference between very similar
1:12:53kinds of dog
1:12:54and it would got to be very good at
1:12:56that.
1:12:57And your your company AlexNet won
1:12:59several awards, I believe, for its
1:13:01ability to out outperform its
1:13:03competitors and so Google ultimately
1:13:05ended up acquiring your technology.
1:13:08Google acquired that technology and some
1:13:10other technology.
1:13:12And you went to work at Google at age,
1:13:14what, 66? I went at age 65 to work at
1:13:17Google.
1:13:1865 and you left at age 76? 75.
1:13:2175, okay. I worked there for more or
1:13:23less exactly 10 years. And what were you
1:13:24doing there?
1:13:26Okay, they were very nice to me.
1:13:28They said They said pretty much you can
1:13:29do what you like.
1:13:31I worked on something called
1:13:32distillation that did really work well
1:13:35and that's now used all the time. In AI?
1:13:38In AI and distillation is a way of
1:13:40taking what a big model knows, a big
1:13:42neural net knows, and getting that
1:13:44knowledge into a small neural net. Then
1:13:46at the end, I got very interested in
1:13:48analog computation and whether it would
1:13:50be possible to get these big language
1:13:52models running in analog hardware
1:13:55so they used much less energy.
1:13:57And it was while I was doing that work
1:13:59that I began to really realize how much
1:14:01better digital is for sharing
1:14:03information.
1:14:05Was there a eureka moment?
1:14:08There was a eureka month or two.
1:14:10Um and it was a sort of coupling of
1:14:13ChatGPT coming out. Although Google had
1:14:15very similar things a year earlier. And
1:14:17I'm
1:14:18I'd seen those and that had a big impact
1:14:20effect on me.
1:14:21The closest I had to a eureka moment was
1:14:24when a Google system called Palm was
1:14:28able to say why a joke was funny.
1:14:30And I'd always thought of that as a kind
1:14:32of landmark. If it can say why a joke's
1:14:34funny, it really does understand.
1:14:37And it could say why a joke was funny.
1:14:41And that coupled with realizing why
1:14:43digital is so much better than analog
1:14:45for sharing information
1:14:47suddenly made me
1:14:49very interested in AI safety
1:14:51and that these things were going to get
1:14:53a lot smarter than us.
1:14:55Why did you leave Google?
1:14:57The main reason I left Google was cuz I
1:14:59was 75
1:15:01and I wanted to retire.
1:15:02I've done a very bad job of that.
1:15:05The precise time year when I left Google
1:15:07was so that I could talk freely at a
1:15:09conference at MIT.
1:15:11But I left cuz
1:15:12I was
1:15:14I'm old and I was finding it harder to
1:15:15program. I was making many more mistakes
1:15:17when I programmed, which is very
1:15:18annoying. You wanted to talk freely at a
1:15:21conference at MIT. Yes. I'd MIT
1:15:23organized by MIT Tech Review. What did
1:15:25you want to talk about freely? AI
1:15:26safety. And you couldn't do that while
1:15:28you were at Google? Well, I could have
1:15:31done it while I was at Google and Google
1:15:32encouraged me to stay and work on AI
1:15:33safety. I said I could do whatever I
1:15:35liked on AI safety.
1:15:37You kind of censor yourself. If you work
1:15:39for a big company
1:15:40you don't feel right saying things that
1:15:43will damage the big company.
1:15:45Even if you could get away with it, it
1:15:46just feels wrong to me.
1:15:49I didn't leave cuz I was cross with
1:15:50anything Google was doing. I think
1:15:51Google actually behaved very
1:15:52responsibly. When they had these big
1:15:55chatbots, they didn't release them.
1:15:57Possibly cuz they were worried about
1:15:59their reputation. They had a very good
1:16:01reputation and they didn't want to
1:16:02damage it. So, OpenAI didn't have a
1:16:05reputation and so they could afford to
1:16:07take the gamble. I mean, there's also a
1:16:09big conversation happening around how it
1:16:11will cannibalize their core business in
1:16:12search.
1:16:14There is now, yes.
1:16:15Yeah. Yeah.
1:16:16And it's the old innovator's dilemma to
1:16:18some degree, I guess.
1:16:19Exactly. Yes, it is.
1:16:20Bad skin, I've had it and I'm sure many
1:16:23of you listening have had it, too. Or
1:16:25maybe you have it right now.
1:16:27I know how draining it can be,
1:16:29especially if you're in a job where
1:16:30you're presenting often like I am. So,
1:16:32let me tell you about something that's
1:16:33helped both my partner and me and my
1:16:35sister, which is red light therapy. I
1:16:38only got into this a couple of years
1:16:39ago, but I wish I'd known a little bit
1:16:41sooner. I've been using our show
1:16:43sponsors BonCharge's infrared sauna
1:16:45blanket for a while now, but I just got
1:16:47hold of their red light therapy mask as
1:16:49well. Red light has been proven to have
1:16:51so many benefits for the body. Like any
1:16:53area of your skin that's exposed will
1:16:55see a reduction in scarring, wrinkles
1:16:57and even blemishes.
1:16:58It also helps with complexion. It boosts
1:17:00collagen and it does that by targeting
1:17:02the upper layers of your skin. And
1:17:04BonCharge ships worldwide with easy
1:17:06returns and a year-long warranty on all
1:17:08of their products. So, if you'd like to
1:17:10try it yourself, head over to
1:17:11boncharge.com/diary
1:17:13and use code diary for 25% off any
1:17:16product sitewide. Just make sure you
1:17:18order through this link.
1:17:19boncharge.com/diary
1:17:22with code diary. Make sure you keep what
1:17:24I'm about to say to yourself. I'm
1:17:26inviting 10,000 of you to come even
1:17:29deeper into the Diary of a CEO. Welcome
1:17:31to my inner circle. This is a brand new
1:17:34private community that I'm launching to
1:17:35the world. We have so many incredible
1:17:37things that happen that you are never
1:17:39shown. We have the briefs that are on my
1:17:41iPad when I'm recording the
1:17:42conversation. We have clips we've never
1:17:44released. We have behind-the-scenes
1:17:46conversations with the guest and also
1:17:47the episodes that we've never ever
1:17:50released and so much more. In the
1:17:53circle, you'll have direct access to me.
1:17:55You can tell us what you want this show
1:17:56to be, who you want us to interview and
1:17:58the types of conversations you would
1:17:59love us to have. But remember, for now,
1:18:02we're only inviting the first 10,000
1:18:04people that join before it closes. So,
1:18:06if you want to join our private close
1:18:08community, head to the link in the
1:18:09description below or go to DOAC
1:18:11circle.com.
1:18:14I will speak to you there.
1:18:16I'm continually shocked by the types of
1:18:17individuals that listen to this
1:18:18conversation um because they come up to
1:18:20me sometimes. So, I hear from
1:18:21politicians, I hear from some royal
1:18:23people, I hear from entrepreneurs all
1:18:24over the world, whether they are the
1:18:26entrepreneurs building some of the
1:18:27biggest companies in the world or
1:18:28they're, you know, early-stage startups.
1:18:31For those people that are listening to
1:18:33this conversation now, that are in
1:18:35positions of power and influence
1:18:37world leaders, let's say.
1:18:39What's your message to them?
1:18:42I'd say what you need is highly
1:18:43regulated capitalism. That's what seems
1:18:45to work best. And what would you say to
1:18:47the average person?
1:18:49Not doesn't work in the industry
1:18:51somewhat concerned about the future
1:18:54doesn't know if they're hopeless or not.
1:18:56What should they be doing in their own
1:18:57lives?
1:18:59My feeling is there's not much they can
1:19:01do. This isn't isn't going to be decided
1:19:04by Just as climate change isn't going to
1:19:06be decided by people separating out the
1:19:09plastic bags from the
1:19:11um compostables. That's not going to
1:19:13have much effect. It's going to be
1:19:14decided by whether the lobbyists for the
1:19:17big energy companies can be kept under
1:19:19control. I don't think there's much
1:19:21people can do to
1:19:24except for
1:19:26try and pressure their governments
1:19:29to
1:19:30force the big companies to work on AI
1:19:32safety.
1:19:33That they can do.
1:19:36You've lived a fascinating fascinating
1:19:39winding life. I think one of the things
1:19:40most people don't know about you is that
1:19:42your family has a
1:19:45big history of being involved in
1:19:47tremendous things. You have a family
1:19:49tree which is one of the most impressive
1:19:51that I've ever seen or read about.
1:19:53Your great
1:19:55great grandfather, George Boole, founded
1:19:57the Boolean algebra logic which is one
1:20:00of the foundational principles of modern
1:20:01computer science.
1:20:03You have uh your great great
1:20:04grandmother, Mary Everest Boole, who was
1:20:07a mathematician and educator who made
1:20:09huge
1:20:10leaps forward in mathematics from what I
1:20:12was able to ascertain. Um I mean I can
1:20:14get the list goes on and on and on. I
1:20:16mean your great great uncle, George
1:20:17Everest,
1:20:19is what Mount Everest is named after.
1:20:22Is that is that correct?
1:20:24I think he's my great great great uncle.
1:20:26His
1:20:28his niece
1:20:30married George Boole.
1:20:33So Mary Mary Boole was Mary Everest
1:20:35Boole.
1:20:36Um she was a niece of Everest.
1:20:39And your first cousin once removed, Joan
1:20:41Hinton,
1:20:42was involved in the new a nuclear
1:20:43physicist who worked on the Manhattan
1:20:45Project which is the World War II
1:20:47development of the first nuclear bomb.
1:20:49Yeah, she was one of the two female
1:20:51physicists at Los Alamos.
1:20:53And then
1:20:55after they dropped the bomb, she moved
1:20:57to China.
1:20:58Why?
1:20:59She was very cross with them dropping
1:21:00the bomb.
1:21:01And her family had a lot of links with
1:21:04China.
1:21:05Her mother was friends with Chairman
1:21:08Mao.
1:21:09Hm.
1:21:10Quite weird.
1:21:13When you look back at your life,
1:21:14Geoffrey,
1:21:16with the hindsight you have now and the
1:21:18retro- retrospective clarity,
1:21:22what might you have done differently if
1:21:23you were advising me?
1:21:26I guess I have
1:21:27two pieces of advice.
1:21:30One is
1:21:31if you have an intuition
1:21:33that people are doing things wrong and
1:21:35there's a better way to do things,
1:21:37don't give up on that intuition just cuz
1:21:39people say it's silly.
1:21:41Don't give up on the intuition until you
1:21:43figured out why it's wrong. Figured out
1:21:45for yourself why that intuition isn't
1:21:47correct.
1:21:48And usually
1:21:50it's wrong
1:21:51if it disagrees with everybody else and
1:21:53you'll eventually figure out why it's
1:21:54wrong.
1:21:56But just occasionally you'll have an
1:21:58intuition that's actually right and
1:22:00everybody else is wrong. Hm.
1:22:02And I lucked out that way.
1:22:04Early on I thought neural nets are
1:22:05definitely the way to go to make AI.
1:22:09And almost everybody said that was
1:22:11crazy.
1:22:12And I stuck with it because I couldn't
1:22:14it just seemed to me it was obviously
1:22:15right.
1:22:17Now
1:22:18the idea that you should stick with your
1:22:19intuitions
1:22:21isn't going to work if you have bad
1:22:22intuitions. But if you have bad
1:22:24intuitions, you're never going to do
1:22:26anything anyway, so you might as well
1:22:27stick with them.
1:22:30And in your own career journey, is there
1:22:32anything you look back on and say with
1:22:33the hindsight I have now, I should have
1:22:35taken a different approach at that
1:22:36juncture?
1:22:39I wish I spent more time with my wife.
1:22:42Um
1:22:47and with my children when they were
1:22:48little.
1:22:50I was kind of obsessed with work.
1:22:55Your wife passed away. Yeah.
1:22:57From ovarian cancer?
1:22:59No, or that was another wife. Okay. Um I
1:23:02had two wives die of cancer. Oh, really?
1:23:05Sorry.
1:23:05The first one died of ovarian cancer and
1:23:06the second one died of pancreatic
1:23:08cancer. And you wish you'd spent more
1:23:09time with her. With the second wife,
1:23:11yeah.
1:23:12Who was a wonderful person.
1:23:14Why do you say that in your 70s? What is
1:23:17it that you've you've figured out that I
1:23:19might not know yet?
1:23:21Oh, just cuz she's gone and I can't
1:23:22spend more time with her now. Hm.
1:23:26But you didn't know that at the time.
1:23:29At the time you think
1:23:33I mean it was likely I would die before
1:23:35her just cuz she was a woman and I was a
1:23:37man.
1:23:38Um I didn't
1:23:40I just didn't spend enough time when I
1:23:42could.
1:23:43I I think I I inquire there because I
1:23:46think there's many of us that are so
1:23:47consumed with what we're doing
1:23:48professionally that we kind of assume or
1:23:50more immortality with our partners
1:23:52because they've always been there, so we
1:23:53Yeah.
1:23:54I mean
1:23:54She was very supportive of me spending a
1:23:56lot of time working.
1:23:58But
1:23:59And why do you say your children as
1:24:00well? What's the what's the issue?
1:24:02spend enough time with them when they
1:24:03were little.
1:24:05And you regret that now? Yeah.
1:24:10Hm.
1:24:12If you um if you had a closing message
1:24:14for for my for my listeners about AI and
1:24:16AI safety,
1:24:17what would that be, Geoffrey?
1:24:20There's still a chance that we can
1:24:22figure out how to develop AI that won't
1:24:25want to take over from us.
1:24:27And because there's a chance, we should
1:24:29put enormous resources into trying to
1:24:31figure that out cuz if we don't, it's
1:24:32going to take over.
1:24:34And are you hopeful?
1:24:36I just don't know. I'm agnostic.
1:24:40You must get get better get in bed at
1:24:42night and when you're thinking to
1:24:43yourself about probabilities of
1:24:45outcomes, there must be a bias in one
1:24:48direction cuz there certainly is for me.
1:24:50I mean imagine everyone listening now
1:24:51has a
1:24:53internal prediction
1:24:55that they might not say out loud, but of
1:24:57how they think it's going to play out.
1:24:59I really don't know. I genuinely don't
1:25:01know.
1:25:02I think it's incredibly uncertain.
1:25:04When I'm feeling slightly depressed, I
1:25:06think
1:25:07people are toast. AI is going to take
1:25:09over. When I'm feeling cheerful, I think
1:25:12we'll figure out a way.
1:25:13Maybe one of the facets of being a human
1:25:15um is because we've always been here
1:25:18like we were saying about our loved ones
1:25:19and our relationships, we assume
1:25:22casually that we will always be here and
1:25:24we'll always figure everything out. But
1:25:26there's a beginning and an end to
1:25:27everything as we saw from the dinosaurs.
1:25:28I mean
1:25:29Yeah.
1:25:31And
1:25:32we have to face the possibility
1:25:35that unless we do something
1:25:38soon,
1:25:39we're near the end.
1:25:42We have a closing tradition on this
1:25:43podcast where the last guest leaves a
1:25:44question in their diary.
1:25:46And the question that they've left for
1:25:47you
1:25:49is
1:25:54with everything that you see ahead of
1:25:56us,
1:25:57what is the biggest threat you see to
1:25:59human happiness?
1:26:04I think the joblessness is a fairly
1:26:07urgent short-term threat to human
1:26:09happiness.
1:26:10I think if you make lots and lots of
1:26:12people unemployed,
1:26:13even if they get universal basic income,
1:26:16um they're not going to be happy.
1:26:19Because they need purpose. Because they
1:26:21need purpose, yes. And struggle.
1:26:23to feel they're contributing something.
1:26:25They're useful.
1:26:27And do you think that outcome that
1:26:29there's going to be huge job
1:26:29displacement is more probable than not?
1:26:32Yes.
1:26:33I do. And what's the
1:26:34That one I think is definitely more
1:26:36probable than not. If I worked in a call
1:26:38center, I'd be terrified.
1:26:41And what's the time frame for that in
1:26:42terms of mass job displacement?
1:26:44it's beginning to happen already.
1:26:46I wrote an article in the Atlantic
1:26:47recently
1:26:48that said it's already getting hard for
1:26:51university graduates to get jobs.
1:26:53And part of that may be that people are
1:26:56already using AI for the jobs they would
1:26:58have got.
1:27:00I spoke to the CEO of a major company
1:27:02that everyone will know of, lots of
1:27:03people use, and he said to me in DMs
1:27:07that they used to have seven just over
1:27:087,000 employees. He said uh by last year
1:27:11they were down to I think 5,000. He said
1:27:13right now they have 3,600 and he said by
1:27:15the end of summer because of AI agents,
1:27:17they'll be down to 3,000.
1:27:19So you've said
1:27:20It's happening already. Yes. He's halved
1:27:22his workforce because AI agents can now
1:27:24handle 80% of the customer service
1:27:26inquiries and other things.
1:27:28So it's it's happening already.
1:27:30Yeah.
1:27:31So urgent action is needed. Yep. I don't
1:27:33know what that urgent action is.
1:27:36That's a tricky one cuz that depends
1:27:37very much on the political system.
1:27:40And political systems are all going in
1:27:42the wrong direction at present.
1:27:44And what do we need to do? Save up
1:27:45money? Like do we save money? Do we move
1:27:47to another part of the world?
1:27:49I don't know.
1:27:50What would you tell your kids to do?
1:27:53They said, "Dad, look, there's going to
1:27:54be loads of just job displacement."
1:27:56Because I worked for Google for 10
1:27:57years, they have enough money. Okay.
1:28:00Okay. [ __ ]
1:28:01So they're not typical. What if they
1:28:03didn't have money?
1:28:04Train to be a plumber.
1:28:05Really? Yeah.
1:28:10Geoffrey, thank you so much. You're the
1:28:12first Nobel Prize winner that I've ever
1:28:15had a conversation with, I think, in my
1:28:17life.
1:28:18So that's a a tremendous honor and you
1:28:20you you received that award for a
1:28:22lifetime of exceptional work in pushing
1:28:23the world forward in so many profound
1:28:25ways that will lead to great
1:28:27and that have led to great advancements
1:28:29in things that matter so much to us. And
1:28:31now you've turned this season in your
1:28:32life to shining a light on some of your
1:28:34own work, but also on the the the
1:28:36broader risks of AI and how um
1:28:40and how it might impact us adversely.
1:28:41And there's very few people
1:28:43that have worked inside the the machine
1:28:45of a Google or a big tech company that
1:28:47have contributed to the field of AI that
1:28:50are now at the very forefront of warning
1:28:52us against the very thing that they
1:28:53worked upon.
1:28:55There are actually a surprising number
1:28:57of us now.
1:28:58They're not as uh
1:29:00as public and they're actually quite
1:29:02hard to get to have these kinds of
1:29:03conversations because many of them are
1:29:04still in that industry.
1:29:06So, you know, someone who tries to
1:29:08contact these people often and ask
1:29:09invites them to have conversations, they
1:29:11often are a little bit hesitant to speak
1:29:13openly, so they speak privately,
1:29:15but they're less willing to openly
1:29:16because maybe maybe they still have
1:29:17something at
1:29:18at some sort of incentives at play. I
1:29:20have an advantage over them which is I'm
1:29:23older so I'm unemployed so I can say
1:29:24what I have. Well there you go.
1:29:26So thank you for doing what you do it's
1:29:27a real honor and please do continue to
1:29:29do it. Thank you. Thank you so much.
1:29:34Many people think I'm joking when I say
1:29:36that but I'm not. What are you coming
1:29:38for? Yeah.
1:29:41And plumbers are pretty well paid.

About this transcript

This page contains the full transcript of Godfather of AI: They Keep Silencing Me But I’m Trying to Warn Them! by The Diary Of A CEO, generated from the public captions YouTube serves with the video. The transcript has 16,545 words across 2,762 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.