Latest / Uncomfortable Conversations with Josh Szeps / "Artificial Intelligence in a Human World" with Prof. Toby Walsh

Transcript

"Artificial Intelligence in a Human World" with Prof. Toby Walsh · Uncomfortable Conversations with Josh Szeps

0:00
G'day, humans! Welcome to the safe space for dangerous ideas. Here's a dangerous idea
0:08
for you. Artificial general intelligence, the holy grail of AI, the moment at which computers
0:16
become smarter than us, which many people estimate might be just a year or two away. Could
0:21
actually be underway right now? And not only could we be in the middle of it, but we may
0:27
not even be able to identify even in principle when we are past it until it's too late and we
0:32
can only identify it by looking back and saying, aha, that was the moment that we were all unaware
0:38
when computers became the most intelligent species on the planet. Today's guest Toby Walsh is
0:44
one of the world's leading artificial intelligence academics. He is a professor of artificial intelligence
0:50
at the University of New South Wales in Sydney. He's the chief scientist at the UNSW Artificial
0:56
Institute. He got his degree in theoretical physics and mathematics at Cambridge University
1:02
in the United Kingdom. He got a degree in artificial intelligence from the University of Edinburgh
1:07
and he was literally the editor-in-chief of the Journal of Artificial Intelligence. So he knows
1:12
what he's talking about. And I love Toby because he's not a doomsayer, he's not an arm waiver,
1:17
he just understands all of the facts, he understands the science, he understands the maths, he
1:20
understands what's going on under the hood and he understands and can predict I think with quite
1:25
a level-headed clarity, the likely challenges that we're going to face much sooner than we think.
1:32
We talk about lots here, we talk about the experience of using chat GPT, we talk about the history
1:37
of artificial intelligence and we talk about the disruptions that it's likely to have on the
1:41
labor force. I mean, none of this is incredibly new stuff but I felt it was time for an update.
1:45
I should say a big happy Fourth of July to our American listeners, I was thinking about doing
1:49
something celebrating America for this episode and I thought, ah, you're going to have a lot of
1:54
that anyway, but I wanted to say congratulations and good going on 249 years of something. Love you
2:02
guys. Also, if this episode is less than an hour long then you are on the free feed of this podcast
2:09
and so that's great, thank you, love you, fabulous, but you're only hearing less than half of the entire
2:15
conversation. If this episode is, you know, well over an hour, about two hours long, then thank you
2:22
so much for being a subscriber, for being a paid subscriber. I just want to clarify, the amount of
2:27
revenue that we get from advertising, from the free feed, doesn't even cover 20% of our costs.
2:32
It's nothing. We could just, I mean, sometimes I'm tempted to just forget about it all together
2:36
and just go to a completely free slash paid model. So just be aware that if you are paying,
2:42
then you are funding this entire project. You are funding uncomfortable conversations. You are
2:46
funding the style of conversation that I try to have where one week I'll be talking to somebody
2:51
who thinks that, you know, Israel is a genocidal state and the next I'll be talking to an
2:55
Israeli parliamentarian who is going into bat for Zionism and where I'm trying to bring you a
2:59
diverse range of perspectives on all of the hot button issues that is that are forcing us into
3:04
these echo chambers. This is fresh in my mind at the moment because I've just come off the back
3:07
of this conversation that you're about to hear with Toby, where we go into that, the impact of
3:11
artificial intelligence, which we didn't even realize was deranging the world at the time when
3:16
we were all going crazy with Twitter in the 2010s. And it wasn't front of mind for us that the
3:20
reason why that was happening was because of the deranged preferences of algorithms that were
3:25
showing us things that were likely to incite us, likely to reinforce what we already believe,
3:29
likely to demonize things that we disagree with, likely to pander to us so that we would like and
3:33
click and comment and share. So yes, that was the tip of the iceberg and we are on the precipice
3:40
of another revolution as big as the industrial revolution as big as the discovery of fire.
3:46
And Toby is the man to help guide us through it. So again, if you want to subscribe on Apple
3:50
podcasts, you can click subscribe on this very episode itself and it'll give you a free trial.
3:55
The preferable way to do it is on substack because then in addition to getting just the bonus
3:59
episodes that we produce and access to all of the episodes, you also get me jumping into your
4:04
notifications if you download the substack app with random delightful zeps, zeps live episodes
4:09
where I'll just jump on the live stream with someone who's fascinating about something that's
4:13
going on in the world. This is a podcast which has a, you know, a fairly long lead time. We're not
4:19
always being reactive to the news. That's one of the pleasures of this that we don't have to
4:23
jump up and down when news breaks, but when news does break and there's someone who in the world
4:27
who I really want to talk to about it, I'll jump on a zeps live on substack live. We'll go live
4:32
streaming and if you're a subscriber, then you'll get a notification from your substack app. You can
4:37
just one click and then you can watch me talking to Michael Moynihan on Matt Welch or, you know,
4:41
whoever it might be, Yasha Monk around the world about things that are going on. That is one of the
4:45
great benefits of subscribing on substack rather than just on Apple podcasts. All right, that pitch
4:52
out of the way. Thank you once again for subscribing. I should mention Toby's, oh, also, to the reason
4:58
that why you should be mindful about the conversation that we're about to have about creativity and
5:03
human creativity versus artificial creativity. If you're not aware, because we sort of gloss over
5:08
in this conversation, is this court case that just ended last week, which was one of the first
5:13
big attempts by, you know, to hold artificial intelligence companies accountable for the content
5:19
that they've used to train their data, the data that they've used to train their models. So,
5:25
and I read from the news clipping last week, the technology companies and Thropic and Meta,
5:31
each one landmark victory is in two separate court cases that examined whether or not the
5:36
firms had violated copyright when they trained their large language models on copyrighted books
5:42
without permission. The rulings are the first we've seen to come out of copyright cases of this
5:48
kind. So, this is kind of a big deal, sets the very, very beginning of the very, very beginning of
5:53
the stage for how human creativity and artificial creativity are going to coexist for the rest of the
5:58
duration of human civilization. Enough waffling from me, enjoy as much as I did, please, the one
6:03
the only Professor Toby Walsh. Where are we at the moment? I feel everyone, I feel like
6:14
something's happened since we last spoke where suddenly lay people have noticed.
6:19
Yeah, I think definitely the public is definitely much more aware and also much more concerned.
6:25
There's much more concern about the impacts that have on people's jobs. I think there's
6:31
reason for that. There's worrying statistics starting to appear.
6:36
various companies, even big tech companies laying off staff.
6:39
And what is the process of the lay off is less about entire industries being laid off in
6:49
heartbeat and more about, I mean, the way that it seems to be unfolding for me is that there are
6:53
teams of 17 designers who now only need four designers and one person who knows how to
6:59
help them to use the AI. There is some truth to the maximum that AI is not going to take a job,
7:05
but someone who uses AI. It's going to be that much more productive that takes your job.
7:09
Yeah. And then the other thing is that a lot of CEOs are saying that the restructuring due to
7:15
AI, it's not clear whether the AI is the excuse or the reality. And then there are other CEOs
7:20
who don't want to talk about AI because they don't want to freak everybody out. And
7:24
you know, they don't want to say the quiet bit out loud and admit that their goal is to massively
7:28
shrink their workforce. But perhaps behind the scenes they are, what's your sense of what industries
7:35
are most vulnerable? I suspect that there's another reason why the public's or why there's
7:41
greater concern because it wasn't, it wasn't the jobs. I think people first thought it was going
7:45
to be, it wasn't, you know, people in factories being replaced by robots. It's white collar people
7:51
being replaced. And in particular, I think the area to be most concerned about is entry-level jobs.
7:58
So graduates entering professions, you already see big, big accounting firms, the big management
8:03
consultants, they're not recruiting as many people this year. I mean, the factory robots did take factory
8:10
workers jobs already, right? I mean, but that happened a generation ago. No one paints a car
8:17
and a factory anymore, no one worlds the car and a factory anymore. And no one is ever going to
8:22
do those jobs again because robots do it so much better, so much more accurately 24-7. So we're
8:30
never going to do those jobs with humans ever again, but we don't make cars here in Australia.
8:35
Right, but the things that we do do, and not just Australia, about the whole rich world,
8:44
indeed, but the latest warehouses that get opened in Australia, the latest Amazon warehouse,
8:49
it's designed for robots. And there's, there's a tenth of the people employed for the size of
8:55
the warehouse than used to be. But this makes me somewhat reassured about AI because it makes the
9:01
analogy comforts me in the sense that we've had, you know, we all used to work in agriculture and
9:06
then we don't work in agriculture anymore. Then we all worked in factories and we don't have to
9:10
work in factories anymore. Now we all do mundane bullshit jobs on computers and we won't have to
9:14
do those anymore. And maybe the fallout will be just the same as the fallout when whaling was
9:20
banned and everyone said, well, how are we going to light our lamps if we don't have whale blubber?
9:23
We'll just use other stuff to light your lamps and the people who hunt whales will go and do
9:26
something else. But I trust that you're not so saying one.
9:29
No, I think in the long term, you're right. We've always created more jobs. New industries have
9:35
appeared. We've used the wealth and the productivity to do other things. I think the challenge is that
9:43
it took 50 years for the Industrial Revolution to unfold for us to leave the fields and find jobs
9:50
and then make sure that we all profited builds, you know, introduced unions and labor laws and
9:56
universal healthcare and education and welfare to support people. I don't think we have 50 years
10:03
for this transition. How long do we have? Most five. Most five. So it's unfolding so much quicker
10:12
and so those transitions are going to be that much harder. So I sometimes hear AI experts talk
10:18
about how we might be only 18 months away from artificial general intelligence and I don't really
10:25
understand what the difference is between what ChatGPT was 12 months ago and the version now
10:32
that's clearly better and what sort of a tipping point they seem to be pointing to. But when you say
10:37
five years, what does it look like? Could put some flesh on those bones? Well, AGI is when machines
10:46
match all of our capabilities and they stop doing the stupid things that ChatGPT sometimes does
10:53
today. Wait a second. Humans do stupid things all the time. So do they have to stop doing stupid
10:57
things at a rate that's worse than humans? Or do they have to stop doing stupid things all together?
11:03
Well, I think the expectation is that once they match humans, they'll exceed humans.
11:09
But you're right. Humans do stupid things, but language models do even more stupid things.
11:15
Right. And they do things that a six-year-old would easily see as foolish.
11:21
But they're funny because they do, what, A, they hallucinate still and be they...
11:26
And we don't want to get rid of hallucination. Why is that? Because we want them to be creative.
11:30
We want them to be inventive. You don't want them to repeat the business letter they
11:36
last wrote. It's a new business letter that actually says something that's never been said before.
11:41
Right. But we don't... We want them to hallucinate when we want them to be creative,
11:44
but we don't want them to hallucinate when they're quite certainly with a tone of authority,
11:49
giving us information about something. Like, I mean, ChatGPT just got something totally flat
11:54
wrong with me yesterday, where I was asking it a question and it gave a very confident answer.
11:58
And I said, are you sure about that? Because I don't see... It was about the fair classes that
12:02
finair users to... For there, it's award buckets in premium economy or something. And
12:07
I said, are you sure that it's F class? I don't see any F class. And it said,
12:10
oh, my mistake, I apologize. It's actually, you know, T class or something. You don't want that.
12:15
You don't believe it's apologies anymore. You don't believe it's apologies.
12:18
And it keeps on saying that. I won't do the same mistake. Yeah, the challenge there is it's not
12:23
grounded in reality. It lives in this virtual world that nothing, it ever says or does,
12:30
is actually grounded in the reality of the world. It doesn't experience what is true and what is
12:36
false like we do. I think there's a body of thought that will never build true
12:43
AGI, unless we embodied the intelligence in something like a robot that is out in the world,
12:50
experiencing the world and paying the price for being wrong.
12:54
But how much of our model of the world comes from experiencing the world? I mean, I suppose we need
13:00
to grow up as a child. It's very important. But we do, I mean, we understand gravity,
13:04
we understand how the world works. I don't understand gravity because of my experience in the world.
13:09
I understand. You do because it was painful when you fell over, with things hit you from above.
13:14
Right. You got a very visceral experience of what gravity is. Even if you don't understand
13:21
Newtonian gravitation or even better still, you know, I'd, I'd, I'd sign theory of, you know,
13:26
general relativity, which explains, strangely, I think I understand those things better than I
13:31
understand gravity in a practical sense in the sense that I can't throw, I can't hit a target
13:35
very well if I'm throwing a ball and there's wind disrupting it, right? Like my mind is not
13:41
terribly good at the physics of it, even if my brain can understand it. I mean, my suspicion is
13:47
that the set of things that it's that you need to know through embodiment is a very small subset
13:53
of the total store of knowledge that I have, right? I mean, if I was, if we raised a brain, a human
13:58
brain in a vat matrix style, couldn't it become intelligent? Well, it's an interesting question.
14:05
We don't know any subset language models are that experiment. Yeah. How intelligent can we make
14:11
them when they're, they're not in part of the world? They're just part of this virtual world
14:16
they live in. But the fair buckets of fin air are a purely virtual construct. You know what I mean?
14:22
Like as, as are most of the things that we care about in the 2020s. Yeah.
14:29
But equally, you know, the imaginary fair buckets are also part of the
14:36
so all right, just staying on general intelligence for a moment.
14:42
I'm still confused about whether or not it means that the machine that we're dealing with
14:47
is seems superhuman, whether it simply seems better than a person that most things because I
14:55
feel like chat GPT is already better than a research assistant at most things. It makes mistakes,
15:02
people also make mistakes, but it's breadth of knowledge and its ability to search and
15:08
integrate information and even come up with reasonably creative ways of assisting me,
15:14
far exceeds human capacities. Yeah. So I think you're together to bring AGI and ASI,
15:20
artificial general intelligence and artificial superintelligence. So you know,
15:24
the expectation is that once you get to AGI, why would we stop there? Why would we have the limitations
15:29
of human intelligence? We should surely be able to build machines that would be smarter than us.
15:35
I mean, they have a bunch of features that characteristics that will, you know, give them an
15:40
actual advantage. They are faster than us. I think, you know, in gigahertz, millions of instructions
15:46
a second, we think in the tens of hertz, they're not limited in memory. We forget stuff. They
15:53
don't have to ever forget anything. We run on 20 watts of power. That's the maximum. It's a third
16:00
of our energy. Your brain uses more energy than your heart, interesting fact. But computers don't
16:08
have to be limited to 20 watts of power. We can have almost unlimited. We build a whole nuclear
16:12
power station. Yeah, we are. To power these beasts. So there's various physical limitations
16:21
that they won't have to have. And in the narrow domains that we've got machines to play,
16:28
as well as humans, they've almost always gone way past humans that computers now play go or
16:37
chess far, far better than humans. It's far, far better than the best human on the planet.
16:42
And it almost seems quaint when we say that now. Oh, it can play chess or it can play go a game
16:48
that no one in the West had heard of until a computer beat an Asian person at it. You know, like
16:53
these things were surprising at the time, but they seem like they feel like ancient history now
16:59
to those of us who look around having a daily relationship to what to seemingly sentient
17:05
artificial creatures who help us with everything. I mean, on my commute when I drive
17:09
it, it is ancient history. It's 30 years since computer nearly, nearly 30 years since computers
17:13
got better than humans at chess. But go was less than 10 years ago. To 2016. So that's nearly 10
17:20
years. Wow. So yes. But there's lots of other things now that they're rounding off as well.
17:29
Yeah. Is it just a quantitative difference or is it qualitative between beating someone at go
17:36
and me being able to have a conversation with it in the car with my kids when one of them asks me
17:42
whether or not, you know, reptiles blink and I just turn it on and say, hey, you know,
17:48
we go have a whole conversation about reptilian evolution, about who the first people were who
17:53
used glass, how humans discovered glass, how glass is made, what civilizations, you know,
17:58
we're just going back and forth, man, chat, GPT, the kids are in the back. We do this for,
18:01
you know, 15, 20 minutes a day. We just talk about things. They love it. I love it. That to me is
18:07
if you'd shown that to me when a computer beats someone at go, I would have said, well,
18:14
this is a completely different kettle of fish than teaching it again. And the difference is that
18:18
they've got to come out of language. And perhaps it shouldn't surprise us that language was
18:24
going to be at the center of intelligence because language is at the center of our intelligence.
18:28
You know, what's going on back here seems to me to be language.
18:34
Thoughts emerge in your mind, formed in English or Spanish or Mandarin or whatever you think in.
18:42
And, you know, I've learned another language and I've woken up and realized I've been dreaming
18:47
in that language. My thinking is in that language, why don't you get yourself immersed?
18:52
What's the language that you learn? French in the time. Interesting.
18:57
And I remember thinking, oh my god, the French are taking over my brain.
19:03
You've realized, you know, there was a switch when you know one night I woke up and I realized
19:06
I'd be dreaming in French. Yeah. That is the language of thought. And that is the thing that
19:13
makes us unique. Or that's the thing that we thought made us unique was that we were the only
19:19
animal that had, you know, a sophisticated language. And there's only very basic rudimentary
19:27
languages you can find in other animals. No one has the richness of language that we have.
19:32
Right. But the animal has allowed us to be so smart because we've been able to
19:38
communicate ideas, write ideas, downshare ideas, construct the world based upon that language.
19:46
Right. I mean, there are so there are two possible mechanisms that you're pointing to.
19:50
There one is the collaborative, which is a kind of a, you know, you've all know,
19:54
our R.E. makes a big deal out of this. Which is big. I mean, I stand, you stand on the
20:00
shoulders of giants. Yes. Yes. I mean, yes. I didn't have to go back and invent all of
20:05
artisans' indulges. I could stand on all the artisans' research. They've been done by the people
20:11
who came before me and before them, you know, the people who invented computers and all the people
20:16
who invented physics and everything else that we get to stand on top of. Yes. Yes. I'm not
20:20
pooping. We don't have to reinvent everything. No, absolutely. And we can also, you know,
20:25
everything else in the animal world has to start from scratch when it's born and you find everything
20:32
out for itself. Yeah. Or just see its parents prodding an antil with a stick and, you know,
20:37
or whatever, whatever rudimentary tool use they might have if you're a chimpanzee or something.
20:42
But science is, in that sense, you know, that's, you know, the greatest exponential process
20:48
in play is science. You know, we get to build on the cumulative knowledge of everyone who came
20:54
before us. There is many scientists alive today has ever lived. Which if you understand the
21:02
mathematics of exponential process tells you even the number of scientists is an exponential.
21:07
Hmm. So the one mechanism by which language is a game changer is it allows us to collaborate,
21:13
learn from past generations, and complex sophisticated. Right. That's the second one that I just
21:20
want to point to just because I think they're slightly different things because when you said,
21:22
when you talked about dreaming in French, for example, there, there's a closer, a tighter and
21:28
potentially more controversial link between language and thought, which is that you actually
21:33
sort of need to articulate the idea in order for it to be a fully realized idea that without
21:40
language, you wouldn't even be thinking the same way. Indeed, and the most beautiful language
21:46
is mathematics. And that's the heart of science is that we've invented a special language of very
21:51
precise language, which allows us to express really complex, powerful ideas that we can then,
21:59
build machines around awful. Because it is that language that allows us to do that.
22:04
And now we're using the language of mathematics to build machines that can
22:08
ape the language of language and talk at us through large language models.
22:12
Yeah, they have a decent command of language. That's, that is a step change. That is
22:19
an important step on the way to AGI, for sure. I want to get back to the whole language thing and
22:26
artificial intelligence being dismissed by some people as just a very sophisticated order
22:30
complete. The other thing about language is that it's, I think it's surprised everyone working
22:37
in the field how easy it was with enough statistics. How easy what was language was for the AI?
22:45
Yes, we just, if we have enough data, collect enough statistics, these are very statistical
22:51
models. They draw on upon statistics, meaning the likelihood that one word is going to follow
22:56
another, basically, out of all the corpus of human writing. And we'd spend a long time
23:02
trying to formalize the grammar of language and so it turns out no. Simple statistics with enough
23:09
data gives us, you know, a good command of language and they seem to be working quite well.
23:15
Is that not sort of also how humans learn language? I mean, most of us don't understand grammar
23:20
because we were taught grammar, right? We sort of just, we glean it and probabilistically there's
23:25
probably an apostrophe here, a comma probably goes here. Aren't we sort of, we are the same thing.
23:31
But, you know, this is where I have a huge respect for humans, which is that we're much better,
23:37
we're much more efficient. If you look at, you don't have to read half of the internet to be able
23:42
to speak English, whereas chat GPT read half of the internet to be able to read English. Half,
23:48
I thought it was like 90% or something. No, it was less than half, but you know, it was a
23:54
significant fraction. But just that that's more than a human could read in the lifetime of reading,
23:58
right? Right. Okay. You know, that's hundreds of thousands of books. Yeah. I mean, including all my
24:04
books and all your books. Yeah, which we're not getting paid for. But, you know, a child doesn't read
24:12
that many books before they've already got a cook monitor. Right. Yes. So we're better at learning.
24:17
Yes. But I mean, that's sort of a cheat because if it's easy for an artificial intelligent and
24:24
officially intelligent, I mean, we have to stop and think how wonderful evolution was to be able to
24:29
come up with something as efficient as that. Right. What I mean is giving us credit over the AI is
24:35
a bit of a cheat because it smuggles in the assumption that doing lots of reading is a bad thing in
24:42
order to build your world model. Like if that's easy for the AI to do it. Yeah, it's easy.
24:46
Right. We were able to do it. We had the internet. We can't do it. We can't read. No,
24:50
well, you know, we can read trillions of pages. Computers can read at speed and they had all the
24:55
internet available to them. So yeah, they could do that. So yeah, it was a engineering solution to
24:59
the problem. But you know, I mean, there's an interesting question arising, which is, you know,
25:04
why were the only why were the only animal that had language? Why is it that nothing else
25:10
has acquired language? And something that's going to happen in the next couple of years is that
25:16
we are going to be able to decode the language, whatever it is, of the whales and the
25:23
corvids and all the other things that have somewhat seems to us from outside some rudimentary
25:30
language. We have to speak whale and speak all the other, speak ape and speak all the other
25:36
things that all the other animals that do some limited form of communication and understand
25:41
their language and be able to speak to them. Amazing. And that's courtesy of AI, which will
25:47
have worked out what those languages are. That's super cool. And it would do that by comparing
25:51
whale song, all of the whale song that we've got and looking at the context in which the
25:56
song was sung and making inferences that humans can't. Looking for the correlations and yeah,
26:01
we've got enough data just like they manage to acquire English by throwing enough
26:07
data and spotting the correlations and the statistics. They will hopefully learn and we're
26:12
starting to see that they'll learn the language of elephants and the language of whales and
26:18
yeah, we've already got glimpses into that. So elephants, for example, have names for each other.
26:26
Well, so do you think I'll be able to walk into the house and when my cat meows,
26:31
there'll be a little speaker on the wall that can tell the difference between I'm hungry or
26:36
good to see you or I wish you hadn't gone away for so long or cold. There's an AI that already does
26:44
that. But does it work or is it just a gimmick? I think it works reasonably well. It depends,
26:54
I suspect how sophisticated you think cats are. Yeah, but there's clearly something
26:58
different going on. There's probably a difference between I'm terrified and I'm uncomfortable.
27:06
They've worked on the face as well. So the cat will display different faces as well,
27:11
depending on whether it's angry or hungry or happy and so on. This criticism that I alluded to,
27:18
which you hear naysayers of AI talk about a lot, it's a complicate. It's a sophisticated
27:24
auto-complete. It's basically just like you're typing something into Google, Google can predict that
27:28
most people who've written those four words end up writing the fifth word, it's like when you're
27:33
writing an email and it says you write yours at the bottom and it fills in sincerely for you.
27:38
It's not doing any thinking, it's not doing anything creative, it's not really contributing to
27:41
anything. AI is just an extremely sophisticated version of that. It's looked at all of the things
27:46
that have ever been written and when you ask it something, it just probabilistically spews out
27:52
mindlessly whatever it thinks the most probabilistically correct response is. That's a good criticism
27:59
not of artificial intelligence, but of human intelligence. Well, that's what we have as well.
28:04
Rest amazing. What human intelligence is and a lot of human intelligence is also complete.
28:10
And if you think about it, a lot of the time we have a conversation, you're not actually thinking,
28:15
stop and think deeply before you start speaking. Not auto-complete.
28:19
Of course, you come into the studio before this recording. Hi, how you doing? Thanks for coming,
28:25
impressed that you rode your bike. It's all basically, I wasn't really really
28:29
care about the fact that you rode your bike, but you know, it's something to say. It's so much
28:33
of what we're talking about. I mean, that's why I think people like, well, at the risk of blowing
28:38
my own trumpet shows like this, where we're not on auto-complete. I think it's clear, like,
28:42
there are very rare glimpses when you're having a great conversation around a dinner party with
28:46
friends or some deep flow state with another person where it's not auto-complete. It's a tap dance
28:54
or there's repartee. There's actual ingenuity and creativity and contemplation going on,
28:59
but 95% of our interactions aren't that. They aren't. If you ring some up in the core center,
29:05
they're reading from a script. Quite literally in that case.
29:09
Quite literally. Quite re-from a script.
29:11
Even in a more metaphorical sense, you walk into a store and I have an exchange with the cash
29:15
here. The weather hasn't the storm been nasty. And so it turns out, we've taught those
29:20
scripts to machines. I don't think a criticism of the machines. That's a criticism of us.
29:26
Is it true of the machines in the narrowest sense that people try to make it? Because I think
29:32
the experience of people who are frequent users of AI is just that if it's true,
29:40
that dismissive explanation, it must be trivially true and missing something fundamentally important.
29:48
Because I have a relationship to my GPTs that I wouldn't have where it's simply auto-complete.
30:00
We're quick to project our humans.
30:02
No, but I understand that. I'm not inferring that it cares about me.
30:07
I'm just getting useful creative content out of it that I wouldn't be able to get
30:13
from a machine that wasn't doing something other than regurgitating.
30:22
There's a certain amount of regurgitation and I think a lot of the time when people
30:27
the latest advice, it's past the bar exam. It's past the medical first-year entrance exam.
30:34
You look closely at what they've done and you discover that the questions were actually
30:40
on the internet or questions very similar to the questions that were answered were on the
30:44
internet already. A lot of it is sophisticated information retrieval from trillions of bites of
30:53
data. That's impressive because humans can't look at trillions of bites of data, but the machines
30:58
can. A lot of things that we thought require quite a bit of intelligence turn out to be things
31:04
that you can do. You've got a big enough database, you've trained on a big enough corpus of knowledge,
31:10
you can do by information retrieval.
31:13
Right. I guess it trivially true in the sense that if you can write a symphony, a brand-new symphony
31:22
on the basis of the existence of all other symphonies, okay, you can call that auto-complete if
31:26
you want, but what are we even talking about here? Then everything is auto-complete. Unless you
31:32
believe that humans have a soul where there's a sprinkle of pixie dust and God touches us to write
31:37
a symphony that's different from the one that AI would write.
31:39
I mean, that's one way in which AI will be an amazing tour that will, in some sense, surpass us,
31:48
because it can do things that humans can ever look at. Look at data sets that humans don't have
31:54
the attention span to be able to look at. Look at more symphonies than a human could look at,
32:02
or games of go than a human could look at. Alpha go got to be good at go because it played
32:07
millions of games of go. No human could play a million games of go in a lifetime of playing go.
32:15
But what it wasn't doing, which I think I had thought it was doing when the news broke,
32:20
is simply using brute force to extrapolate out every possible scenario from every move and
32:28
then select the winning move, which would be impressive. It was using quite a bit of compute,
32:32
so I mean, they didn't. No, but it was, you know, it had used all the training,
32:40
the hundreds of thousands of hours training beforehand as a way of compiling that knowledge.
32:47
Yes, but correct me if I'm wrong, Toby, my understanding was that if you play a game of hangman
32:53
or checkers or, you know, monopoly with a computer, it's just going to assess if I do this,
33:00
then here are the 150 million different possible outcomes. The best one is this, the best one is
33:06
that. The best one is this. Just go through them all, read them all, look at them all, and I can
33:10
look into the future and I can see all the things that he might do, the Josh might do, and I can
33:14
beat him. But the the set of possible outcomes in a game like go is so vast.
33:21
More than the essence of the universe. Yeah, you'd need the computing power of, yeah,
33:25
the universe isn't big enough. If the universe was a computer, it's not big enough.
33:29
Right. So all that analysis. What was different here is that it's not just using brute force
33:34
to crunch through all of the outcomes and then pick the best one. It is using something akin to
33:39
intuition. It's learned by seeing so many games ago, what are good looking moves and what are
33:46
bad looking good moves and what are good looking ball positions and what are bad looking ball
33:50
positions. And that's why, you know, move 37 of the second game, it made that divine move,
33:57
which just explain that. So on the second game, on the 37th move, AlphaGo played on the fifth row.
34:06
That's early in the game, right? So that's quite early in the game. And the conventional wisdom
34:12
learnt from humans playing go for 2,000 years of playing go, oldest strategy game in the world
34:19
was that you would never play on the fifth row early in the game. That's, you know, a loser move.
34:25
And at least at all, if you watch the tape, you know, you can see the look of, he was stunned,
34:31
you look on this face, he stood up, well, went away to think about it for 15 minutes.
34:36
And a lot of go experts who were watching it,
34:38
thought, oh, the computer's over. Yeah, the computer's time. The kid was going to win.
34:41
Begin a loser move. And it turns out, you know,
34:46
subsequent analysis has shown that that was an absolutely perfect move that has changed the way
34:52
that we see the game of go. That's sort of an entry point to the question of, will AI change the
35:00
way that we see the entry point to everything? Yes, I think you know, one of the most promising
35:10
areas that after intelligence would contribute to is science is going to help us discover new
35:15
science, discover new drugs, accelerate the way that we do science because it could do things like,
35:20
you know, get humans couldn't possibly do. And we're starting to see that already. I mean,
35:25
as an example, there's the UK gene bank. So half a million people in the UK have contributed
35:31
their genotype, which takes a couple hundred dollars to read your genes. And all their physiological
35:36
information, their medical history, their family history. So this is a huge great database,
35:41
terabytes of data. Humans can't do anything with that on their own, when we just, it's just too
35:46
much data. We cannot see the word for the trees. But you could throw AI at it and they're starting
35:52
to do that. And, you know, discovering interesting things. So, you know, one of the first things
35:57
that they discovered was some of the genetic markers for bowel cancer, third most deadly cancer.
36:06
Mostly treatable, but unfortunately by the nature of bowel cancer, by the time you unfortunately
36:11
notice you've got it, it's often too late. But if you caught it early, it would have been treatable.
36:16
Well, now we know the genetic markers. So that you can, well, we don't understand the mechanisms,
36:22
yet. We just understand those genes are implicated. And just because you have those genes,
36:29
does not mean you will get bowel cancer because there are environmental triggers that we don't
36:33
understand, or various other things. But it means, you know, you've got a, you know, 50% greater chance
36:38
of bowel cancer. And you can be monitored from an early age. You can have a blood test and so on
36:44
at an early age. So to see if, you know, if it has been triggered or not. And so then be saved,
36:50
you know, an untimely death. And what about cures for cancer? I hear people talk about that,
36:55
but I don't understand the mechanism by which that would AI would assist. A couple of ways.
37:01
I mean, one is that, you know, we're now using AI to work out the biology, the biochemistry of the
37:08
cell. So, you know, for example, how proteins fold and so on with alpha-fold allows us to understand
37:14
the, you know, the three-dimensional structure of proteins, which is very important to their function.
37:20
And then, you know, to start designing drugs that exploit that structure. We all know that
37:27
structure is really important. Remember, you know, COVID, it was the spike protein that was, you know,
37:34
central to its mechanism. So once we start, you know, having AI to help us to understand that
37:39
biochemistry, then we can understand the mechanisms and we can also start designing specific drugs
37:46
that will work to those mechanisms. And I mean, the implementation here, I suppose, is the
37:52
sort of golden question about how we use the information that we get. Because when you're talking
37:57
about the bowel cancer predisposition, I do, I have thought for some time that it's weird that when
38:03
I go to the doctor, they say, you know, do you have a family history of something or other?
38:08
Well, because I was like, well, why don't you just, we've known enough, for long enough,
38:14
that why haven't I had my blood, I mean, I have, because I'm a nerd, but why hasn't the normal
38:19
person had their blood taken and a DNA scan, and then, you know, a screening done for the obvious
38:27
markers of genetic predispositions, so that my doctor doesn't have to ask me about whether my
38:32
half-uncle, you know, died of bowel cancer, or like, that seems, it seems very old-fashioned
38:37
in the 2020s to just be asking people about their family history when it's right there in
38:41
myself. We could just look at it. We could, except we don't understand it well enough yet, but we
38:46
will. I understand a lot, we understand a lot of basics, right? We don't understand it as well
38:51
as we will once AI takes it. It throws up how, you know, this is going to create interesting
38:58
challenging societal questions, right? So is everyone going to get the same level of care then
39:05
and so on? And, you know, for as an example, the extreme ethical questions, right? So we can tell
39:13
from your genes, we can tell at birth, whether you're going to be susceptible to bowel cancer,
39:20
we can tell in utero, and so now we have the possibility, we have the scientific capability at
39:27
least to be able to select zygotes, you know, when we're doing IVF, for those people who are not
39:36
going to be not going to have bowel cancer, but it's not going to be just bowel cancer, we know,
39:41
that seems very, very little good to, you know, not screening out bowel cancer from the population,
39:50
but, you know, it will be all conditions, you know, it will be autism, it will be anything you
39:56
can think of, you know, blue eyes. Yeah, intelligence perhaps. But again, we've been grappling with
40:04
this for decades and the session, we have, we can biopsy blastocysts and look at that. We have,
40:09
by color, but we'll be able to be so much more discriminatory. But if we're not doing it in
40:14
a discriminatory way now, why would anything change when we get greater fidelity? What? Because
40:21
we've already seen, you know, the first, you know, the first bioengineered human genotype done
40:27
illegally in China was actually for intelligence. Right. Yeah, and I mean, there are other, you know,
40:34
like Denmark is proud of the fact that they are basically eliminating Down syndrome. Yeah,
40:39
according to them, which as a recent guest, meant pointed out to me, that just means that they're
40:45
killing all the tiny Down syndrome babies. And Tommy's, it doesn't mean they're getting rid of
40:49
Down syndrome, it just means that those children aren't being born, right? Yeah. You know, at the risk
40:53
of touching a provocative subject like abortion. But we grapple with this in the sense that we
40:59
don't allow China babies in the worst. Like I might twins because they were the product of IVF.
41:03
We had them. But we do. We do allow designer babies. We choose our spouses. We choose that
41:09
exactly right. So there's a really inefficient. And we chose the embryos that were the fittest and
41:14
the least likely to have, you know, obvious genetic diseases as best we could tell when they
41:21
were being born. But we weren't selecting for their eye color or their height or their hair color
41:25
or anything. And we, I mean, the technology existed. We could have looked at it. I don't know if
41:29
we would have been allowed to do it. I didn't want to do it. But we've done now AIs being used
41:33
and Australian success story, life whisperer, to select the most likely eggs to implant, the ones
41:42
that look, you know, was these, watch them down the microscope and divide something that humans,
41:48
you know, no good at doing that. We didn't evolve to watch cells dividing down a microscope.
41:53
I don't know. My IVF specialist was pretty good at picking out the, you know, just like your
41:58
IVF specialist has slowly learned that process. Yeah. That's an ideal task to teach an AI to do.
42:04
And we've taught AIs to do that. So, you know, the positive of that, of course, is you can
42:08
increase the success rate of IVF. But also, you know, we don't know yet, you know, whether there's
42:17
some unfortunate or unexpected biases in what it's selecting for. And then, you know, of course,
42:24
we can start asking other questions. We can start selecting for other things.
42:28
So, looking at the way that an embryo is developing in its first five days of life and inferring
42:35
from that, whether or not it's going to grow into a human being or miscarry, is the perfect kind of
42:41
thing that AI will be good at, given enough data, right? It's good after ready. Yeah. And so,
42:49
I understand that things like that and analysis of radar patterns and radiology examinations.
42:57
We weren't evolved for that, you know. No, that wasn't. Yeah, we evolution.
43:02
No, there was not a lot of it in the savanna right now. We're looking at microscopes.
43:06
Not a lot of microscopes out there. Not surprising. We don't have those skills.
43:10
Right. And yeah, spotting pre-cancer was growth on a radiology scan of a person's lung or data entry
43:19
or even convincing and going over reams of legal contracts in the purchase of a property
43:27
or something like that. These are all things that obviously an artificial system is going to have
43:31
a keen eye for them we will. But the creative stuff is the stuff that I'm sort of interested in
43:39
because most of what the most terrified dystopian scenarios reflect on is the possibility that
43:47
we won't even need to really be thinkers anymore because we'll just outsource all of our thinking
43:53
about things to these machines. Do you believe that? But whether the machines are creative
44:01
is a question that's haunted AI from the very beginning and still haunts us today, still.
44:07
For many people, I think it's the last refuge in the sense of, okay, yes, it can do,
44:12
it can look at the x-rays, it can do this, but it won't be creative like humans,
44:19
something special. Will it make art? Will it make art? I think the answer to that is probably,
44:28
it will make art, but it won't mean anything to us. So, you know, on a functional level,
44:34
it will look like the paintings and sound like the music and read like the books that humans make.
44:42
But, you know, art speaks to the human condition, it speaks to
44:46
falling in love and losing a loved one and contemplating your mortality, those human experiences.
44:54
And they speak to us because, you know, we're humans and we understand those things that
44:59
I've been told to learn. I haven't learned off anything apart from human experiences.
45:03
Right, and just like learning all the go-moo's and then, you know, the fifth time.
45:07
It will be reciting the right things, it will sound right. But, you know, I'm not going to
45:12
care about, you know, its feelings about its mortality because it's not mortal. But,
45:17
are you, when you read a novel, are you caring about the novelist's feelings about mortality or
45:21
are they speaking to us? They speak to us because they, you know, they've had those tortured
45:26
experiences themselves. Not sure that's true, Toby. They speak to me because they speak to me.
45:34
Yes. And, you know, sometimes the machine will say things that will speak to us.
45:38
Yeah. But, you know, it's one area that the AI is going to
45:44
is inherently limited, which is that it doesn't have a deep internal emotional life inside.
45:51
But do you need reflect on it? So, there are a few ways that we use art, right?
45:55
We're talking when the number one box office hit is a movie called F1 where Brad Pitt
46:01
races around a track very fast in a car and it's by the same director. I'm not sure I'd call that
46:06
art as top gun too. Well, I would know. I would know.
46:09
There's a few years ago as the world over disagree. Well, yeah, we're entertaining, but is it art?
46:14
Right. So, yes, I do think that's it. And then there's, you know, the movie that was recently
46:19
named the best movie of the 21st century so far, which is Parasite, this Korean movie that won
46:24
the Academy Award a few years ago. You know, something that seems genuinely fresh and unusual,
46:30
right? So, no, I'm not saying that a blockbuster couldn't be art. I mean, you know, something like
46:34
Oppenheimer does speak to, you know, there's, you know, the torch and soul, the person who created
46:41
the nuclear bomb. That's, you know, there's some great human themes in there to explore that the movie
46:47
does. Right, but I'm just trying to set up a spectrum of, I haven't seen F1, but I'm not
46:52
expecting. You know, I'm trying to set up a spectrum of innovation, right? So, that there are things
46:56
that speak to us in ways that are either that are probably easier to replicate than others. So,
47:04
F1 and the Marvel movies and such are probably the first kinds of things that AI will be useful for
47:13
because they're quite formulaic. And then at the very far end of the spectrum, there's like
47:18
Proust or like something highly esoteric and that really when it arrives has not been seen before.
47:28
And I'm not sure whether or not there's a difference of kind between those two things that there's
47:35
a spark of sort of divine human creativity in a way that like is required for the totally
47:41
original thing. But I can be moved by the non-original thing. I mean, I will, I'll sometimes read a
47:48
novel. It is a great novel because I want to do the the slog of getting into the writer's mind.
47:54
But after I've read one of those, I'll pick up a Lincoln lawyer book from this serialized version,
48:02
like series of detective and lawyer books that are written by a former Los Angeles Times crime
48:09
reporter who just smashes out 70 books a year, you know, and they're and they're pot boilers and
48:16
they're very predictable. And if you fed them all into an AI and AI could probably come up with a
48:19
pretty good facsimile of a new one of those. I suspect you may be using AI. And it's and I love
48:26
them. And and I don't think I'm getting into the mind of the writer. I think I'm just being touched
48:33
by plot and circumstance and the joy of a good read. So I'm not sure the I'm not sure you need
48:43
the writer in the picture if the AI has learned everything that every human has ever written before
48:47
and can mimic it. Yeah. Well, I mean, we can be entertained. But I don't know whether we'd be touched
48:55
as much. But you know, like many of my colleagues, I think we're biological machines. There's no
49:03
nothing, no spark and nothing divine inside. And if we're biological machines, why couldn't
49:09
silicon machines do the same?
49:39
One more platform to your life if you're not already on sub stack. But here's the thing we're now on
49:44
Apple premium podcast as well. So with one click, Apple will just subscribe you up to the premium feed.
49:49
You'll be able to listen to the rest of this conversation right now, right now, right now. So open
49:53
up your Apple podcasts, click the old pay old Zepsi button. That's probably what it's called pay old
49:59
Zepsi. And you click the button and then you no longer have to be a free writer. You can feel like
50:04
you're contributing to the mission of this show, which is really what we're all here for,
50:08
better conversations, a freer, more freewheeling attitude towards the things that we can talk about,
50:14
a way to steal man our opponents and to take the most gracious approach towards disagreement.
50:20
That is what you're voting for when you vote for uncomfortable conversations by going into your
50:24
Apple podcasts and clicking subscribe. I love you. If you don't do that, that's fine. I'll see you
50:30
the next free episode, I guess. Take care.