0:05

Welcome

0:06

to the Making Sense podcast. This

0:09

is Sam Harris. Just

0:11

a note to say that if you're hearing this, you are not currently

0:14

on our subscriber feed and will only be

0:16

hearing the first part of this conversation. In

0:18

order to access full episodes of the Making Sense

0:20

podcast, you'll need to subscribe at SamHarris.org.

0:24

There you'll find our private RSS feed to add

0:26

to your favorite podcatcher, along with other subscriber-only

0:29

content. We don't run ads on the

0:31

podcast, and therefore it's made possible

0:33

entirely through the support of our subscribers. So

0:36

if you enjoy what we're

0:36

doing here, please consider becoming one.

0:46

Welcome to the Making Sense podcast.

0:48

This is Sam Harris. Okay,

0:52

just a reminder that subscribers

0:54

to the podcast can now share full

0:57

episodes by going to the episode

0:59

page on my website

1:01

and getting the link.

1:04

And you can share one-to-one with friends

1:06

and family, or you can

1:08

post to social media, whatever

1:11

you like. Okay,

1:14

today I'm speaking with Mustafa Suleiman.

1:16

Mustafa is the co-founder

1:19

and CEO of Inflection AI

1:21

and a venture partner at Greylock,

1:24

a venture capital firm. Before

1:26

that, he co-founded DeepMind, which

1:28

is one of the world's leading artificial intelligence

1:31

companies, now part of Google.

1:34

And he was vice president of AI product

1:36

management and AI policy at

1:39

Google.

1:40

And he is also the author of a new book,

1:42

The Coming Wave,

1:44

Technology, Power, and the 21st Century's

1:46

Greatest Dilemma, which is the focus

1:49

of today's conversation.

1:51

We talk about the new book. We talk

1:53

about the progress that was

1:55

made in AI by his company,

1:57

DeepMind, various landmarks.

2:00

They achieved Atari DQN,

2:03

AlphaGo, AlphaZero, AlphaFold. We

2:06

discussed the amazing fact that we now

2:08

have technology that can invent

2:11

new knowledge. The risks of our

2:13

making progress in AI,

2:15

super intelligence as a distraction from

2:17

more pressing problems, the inevitable

2:19

spread of general purpose technology,

2:22

the nature of intelligence, productivity

2:24

growth and labor disruption, the

2:27

containment problem,

2:28

importance of scale, open

2:31

source LLMs, changing

2:33

norms of work and leisure,

2:35

the redistribution of value,

2:37

introducing friction into the deployment of

2:40

AI,

2:41

regulatory capture, the

2:43

looming possibility of a misinformation

2:45

apocalypse, digital watermarks,

2:48

asymmetric threats, conflict

2:51

and cooperation with China, supply

2:53

chain monopolies

2:55

and other topics.

2:57

Anyway, it was great to get Mustafa here. He's

2:59

one of the pioneers in this field.

3:01

And as you'll hear, he shares many of my concerns but

3:04

with different points of emphasis.

3:07

And now I bring you Mustafa Suleiman.

3:15

I am here with Mustafa Suleiman. Mustafa,

3:17

thanks for joining me.

3:19

Great to be with you, Sam. Thanks for having me.

3:21

So you have a new book which the

3:24

world needs because this is the

3:26

problem of our time. The title

3:28

is The Coming Wave, Technology,

3:31

Power and the 21st Century's

3:33

Greatest Dilemma. And

3:35

we will get into the book because

3:37

it's really quite a good read. And

3:41

we will talk about what that coming wave

3:44

is. But

3:45

you're especially concerned about AI

3:47

which is your

3:49

wheelhouse, but also you're talking

3:51

about synthetic biology

3:54

and to a lesser degree robotics

3:56

and some other

3:57

technologies that are gonna

3:59

be more and more...

3:59

presence if things don't run totally

4:02

off the rails for us. But before we jump

4:04

into the book, let's

4:06

talk about your background. How would

4:08

you describe the

4:09

bona fides that have brought you to

4:11

this conversation? Yeah, I

4:14

mean, I started life, I

4:16

guess, as an entrepreneur.

4:18

When I was 18, I started my first company,

4:20

which was a point of sale system

4:23

sales company. And we were sort

4:26

of installing these sort

4:28

of very early PDAs back

4:31

in 2002, 2003, and networking equipment. It wasn't successful,

4:35

but that was my first attempt. I

4:38

dropped out of Oxford at the end of my

4:40

second year, where I was reading philosophy

4:43

to start a charity.

4:45

And I helped two or three other people get a

4:47

telephone counseling service off the ground. It was

4:50

a secular service for

4:52

young British Muslims.

4:54

I was a just turned an atheist

4:56

at the time, having been to

4:57

Oxford, discovered

4:59

human rights principles and the

5:01

ideas of universal justice,

5:04

and managed to sort of move

5:06

out of the faith,

5:07

and decided

5:09

that I really wanted to dedicate

5:12

my life to doing good and studying

5:15

philosophy and the theory.

5:16

I was too esoteric

5:19

and too distant from action. I'm a very

5:21

kind of practical action

5:23

focused person.

5:25

So I spent a couple years doing that.

5:28

A little bit after that time, I spent

5:30

a year or so working in local government

5:32

as a human rights policy officer

5:34

for the mayor of London at the time. I think

5:37

I was 21 when I started that job. It

5:39

was very

5:40

big and exciting, but

5:41

ultimately quite unsatisfying and frustrating.

5:45

Who was the mayor? Was that Johnson?

5:47

That was before Johnson, yeah, quite a bit

5:49

before. It was Ken Livingston

5:51

back in 2004. So quite a while back

5:55

in London.

5:56

And then from there, I wanted

5:58

to see how I could... scale up my impact

6:01

in the world and

6:03

you know i helped to start

6:06

a conflict resolution firm i was very

6:08

lucky at the age of twenty two.

6:10

To be able to

6:11

co found this this consultancy with

6:13

a group of some of the most

6:16

practiced. Negotiation

6:18

experts in the world

6:20

some of the people are involved in the piece and reconciliation

6:22

process in south africa post apartheid.

6:25

There's a big group of us coming together with very different

6:27

skills and backgrounds and

6:30

i had incredible three years there

6:32

working all over the world and.

6:34

Inside person for the dutch government

6:37

for you know on the israel-palestine question

6:40

you know many different places

6:42

and it was hugely inspiring and taught

6:44

me a lot about the world. Put a

6:46

sort of fundamentally realized from there that if

6:48

i. If i didn't get back to technology

6:51

i would miss the most important transition

6:54

you know wave if you like happening in my lifetime.

6:57

And you know

6:59

i said about

7:01

shortly after the climate negotiations

7:03

that we were working on in two thousand nine in koban

7:05

hagan

7:06

everyone left feeling. Frustrated

7:09

and disappointed that we can manage to reach

7:11

agreement and this was the year that.

7:13

Sort of obama was coming over and everyone had a lot

7:15

of hope

7:16

and it didn't happen turns out for another

7:19

ten or twelve years.

7:20

And i i sort of had this aha moment

7:22

i was like if i don't get back to

7:25

technology then i'm gonna miss the most important thing

7:27

happening and so. I said about this

7:29

question i sort of find anyone that

7:31

any anyone who i knew even

7:34

tangentially who is working in.

7:36

Technology

7:37

and my best friend from

7:40

when we were teenagers his

7:42

older brother was demissus obvious.

7:44

And we were playing poker together

7:47

one night in the victoria casino in

7:49

london and we got.

7:51

Chatting about the ways that at the time

7:53

you know we framed as robots we're gonna transform

7:55

the world and

7:57

deliver enormous productivity boost

7:59

and improve. efficiency in every respect.

8:01

And we were sort of debating like,

8:03

how do you do good in the world? How do you

8:06

get things done? What is the real

8:08

set of incentives and efforts that really makes

8:10

a difference?

8:12

And both very

8:14

passionate about science and technology and having a positive

8:16

impact in the world. And

8:18

one thing led to another and eventually

8:20

we

8:21

ended up starting DeepMind. I

8:23

did that for 10 years.

8:26

Yeah, along with Shane Legge, right?

8:28

Shane is our

8:30

other co-founder exactly. Shane was a Gatsby

8:33

computational neuroscience unit in London

8:35

at the time. And

8:38

he had just finished his PhD a few years early. He was

8:40

doing postdoctoral research

8:42

and his PhD was on definitions of intelligence,

8:44

which was super interesting. It was very

8:47

obscure and really, really

8:49

relevant. He was sort of trying to synthesize 60

8:51

or

8:52

so different definitions of intelligence

8:55

and trying, tried to sort of abstract

8:57

that into an algorithmic construct, one that we could

9:00

use to

9:00

measure progress towards some

9:03

defined goal. And

9:05

his frame was that intelligence

9:07

is the ability to perform well across

9:09

a wide range of environments.

9:11

So the core emphasis was that

9:14

intelligence was about generality, right?

9:15

And we, you know, we can get into

9:17

this. There's lots of different definitions of intelligence,

9:19

which

9:20

place emphasis on different aspects of our capabilities,

9:23

but generality has become the

9:26

core concept that sort of dominated the field

9:28

for the last sort of 12, 15 years. And of course, the

9:32

term AGI, I mean, that predated Shane,

9:35

but I

9:36

think it was very much popularized by

9:38

our

9:39

kind of mission, you know, sort

9:41

of, it was really

9:42

the first time in a long time that a company

9:45

had been

9:45

founded to invent general

9:48

intelligence or AGI.

9:50

And that was our mission to try and build safe and ethical

9:53

artificial general intelligence.

9:56

So I'm trying to remember where we met. I know we

9:58

were both at the Puerto Rico

9:59

conference at the beginning of 2015

10:02

that I

10:03

think it

10:06

was the first

10:07

of these meetings, but it was the first that I was aware

10:09

of that really focused the conversation

10:11

on AI safety and risk.

10:13

And I know I met Demis there.

10:16

I think you and I met

10:18

in LA subsequent to that. Is that right?

10:21

Yeah, I think we

10:23

met... I can't remember if we met before or

10:25

after that, but I think we had a common interest

10:28

in our LA conversation. It might

10:30

have been just before that

10:32

talking about extremism and radicalization

10:34

and terrorism and... Oh, within Islam, yeah. That's

10:37

right, yeah.

10:38

Yeah, so I can't... I don't think we met

10:40

in Puerto Rico, but that conference was very

10:43

formative of my... It

10:45

was really my first impression of

10:47

how big a deal this was going to be ultimately.

10:50

And then there was a subsequent conference

10:52

in 2017 at a Sylmar where

10:56

I think we

10:58

met again. I think I met Shane there as

11:00

well.

11:00

So let's... Before we jump into,

11:02

again, the book and

11:04

what you're doing currently, because you've since moved

11:07

on from DeepMind and you have a new company

11:09

that we'll talk about,

11:10

but let's talk about DeepMind

11:13

because it really was...

11:16

You know, before it has been eclipsed

11:19

in the popular consciousness

11:21

by OpenAI of late with the

11:24

advent of chat GPT

11:26

and large language models.

11:29

But prior to that, really DeepMind

11:32

was the preeminent...

11:34

I may in fact still be the

11:37

preeminent AI company, but

11:39

it's now a branch of Google. Give

11:42

us a little bit of the history there and

11:45

tell us what

11:46

was accomplished because at

11:48

DeepMind, you had several breakthroughs that

11:51

were just fundamental and you

11:54

really put AI back

11:56

on the map and

11:58

prior to what you did there... We were

12:00

in an AI, so-called AI winter, where

12:03

it was just common knowledge that

12:05

this artificial intelligence thing wasn't really panning

12:08

out, and then all of a sudden everything changed.

12:10

So I think pre-acquisition, which

12:13

was in 2014, I think

12:16

there were probably two principal contributions

12:18

that we made. I think the first is

12:21

we made a very early bet on deep

12:23

learning. I mean, the company was founded in 2010, in the summer of 2010.

12:26

And it really wasn't for

12:30

a couple of years that deep learning had even

12:32

appeared on the field, even academically,

12:35

with the ImageNet Challenge a few years after we founded.

12:38

So that was a very significant bet that we

12:40

made early and that we got right. And

12:42

the consequence of that was that we were able

12:44

to

12:45

hire

12:46

some of the best PhDs and postdoctoral

12:48

researchers in the world, who

12:51

at the time were working on this

12:53

very obscure,

12:54

very uninteresting, largely

12:57

not very valuable

12:58

subject. In fact, Jeff

13:00

Hinton was one of our consultants, so

13:03

was his

13:04

student at the time, Ilya Satskiva,

13:06

who's now chief scientist and co-founder

13:08

of OpenAI,

13:10

along with many others from OpenAI and

13:12

elsewhere who basically

13:14

either worked with us full-time

13:16

or worked with us as consultants.

13:19

That was largely reflective of

13:21

the fact that we got the bet right

13:23

early on deep learning.

13:25

The second contribution, I would say, was

13:27

the

13:28

combination of deep learning and reinforcement learning.

13:30

I mean, if deep learning was obscure,

13:33

reinforcement learning was even more theoretical.

13:36

And we were actually quite careful

13:38

to frame our mission among

13:41

academics,

13:42

less around AGI

13:44

and more around applied machine learning. Certainly

13:46

in the very early days, we

13:49

were a bit hush-hush about it. But as

13:52

we got more traction in 2011, 2012, it

13:54

became

13:55

very attractive to people who were otherwise

13:58

quite theoretical in their outlook.

14:00

to come work on problems like reinforcement learning

14:02

in a sort of more engineering focused

14:05

setting albeit still a research lab.

14:08

And it was the combination of deep learning and reinforcement

14:10

learning that

14:11

led to our

14:13

first i think major

14:15

contribution which was the

14:17

Atari dq and i dq

14:20

and so. You can

14:22

was a pretty incredible

14:24

system i mean it is essentially learn

14:26

to play.

14:27

50 or so of the old school

14:29

sort of a tease Atari games Atari

14:32

games

14:33

to human level performance simply

14:35

from the pixels learning to correlate.

14:37

A set of rewarding moments in

14:39

the game via score with

14:42

a set of frames.

14:44

That led to that score in the run

14:46

up to that any actions that were taken there

14:48

and that was a really significant achievement

14:50

it was actually that. Which kor

14:53

larry page's attention and

14:55

led him to email us.

14:57

You know and you know sort of invite us to

14:59

come and be part of google.

15:02

And then google acquired you and

15:06

what was the logic there use good

15:08

to have google's resources to

15:10

scale or. I mean larry made a very

15:13

simple

15:14

claim which was you

15:16

know i've spent

15:18

the last. You know

15:20

ten years or so

15:22

building a platform with

15:24

all the resources necessary. To

15:26

make a really big bet on hg

15:29

i know why should you guys go through all

15:31

of that again

15:32

you know we'll give you the freedom you need to

15:34

carry on operating as a.

15:36

Essentially a independent subsidiary

15:38

even though we were part of google why

15:40

wouldn't you just come and work with us and

15:43

have all the resources you need to.

15:45

To scale

15:46

you know significantly which is what we did and

15:48

it's it's it was a very compelling.

15:51

Proposition because of the time you know monetizing

15:54

deep learning back in 2014 was

15:56

gonna be really tough so. But

15:59

google had its own. AI

16:01

division as well that was just kind of working

16:03

in parallel

16:05

with DeepMind. At some point, you guys

16:08

merged. I don't know if that happened

16:11

after you left or before, but was

16:14

there a firewall between the two

16:16

divisions for a time and then that came

16:18

down or how did that work?

16:20

Yeah, so the division you're referring to is

16:22

Google Brain, which is run by Jeff Dean. And I think

16:24

that started in 2015 with

16:28

Andrew Ung actually as well.

16:30

And in some ways,

16:32

that's the kind of beauty of Google scale, right? That

16:34

it was able to

16:35

run multiple huge billion

16:37

dollar efforts in parallel. And

16:40

the merger,

16:41

which I think has been long coming,

16:43

actually only happened this year. So Google

16:46

plus DeepMind is now Google DeepMind.

16:49

And most of the kind of

16:51

open-ended research on AI is now

16:53

consolidated around Google DeepMind and

16:56

all of the sort of more focused applied

16:59

research that helps

17:01

Google products more directly in the short term

17:03

is focused on a separate division

17:05

Google research.

17:06

Right. So you had the

17:09

Atari game breakthrough,

17:11

which caught everyone's attention because you

17:14

have these memory serves,

17:17

you managed to

17:19

build a system that

17:21

had achieved

17:23

human level

17:25

competence and beyond and also achieved

17:27

novel strategies that

17:30

many humans wouldn't come up with. But

17:32

then the real breakthroughs that

17:35

got everyone's attention

17:36

were with AlphaGo and AlphaZero

17:39

and

17:39

AlphaFold. Perhaps you can run through those

17:42

because that's when, at least

17:44

to my eye, things just became unignorable

17:47

in the AI field.

17:49

Yeah, that's exactly right. I mean, it's

17:52

pretty interesting because sort of after

17:54

we got acquired, it was actually Sergei

17:57

that

17:58

was sort of insisting that we tackle. Go.

18:00

I mean, his point was that

18:03

Go is a massively complex space,

18:06

and all the traditional

18:08

methods that

18:09

had previously been used for games before

18:12

DQN, which essentially involved

18:14

handcrafting rule-based features,

18:17

which is really what drove the work behind

18:19

Deep Blue, IBM's model

18:22

a long time ago, in 97, I think it was.

18:25

Go has something like 10 to

18:27

the power of 170 possible

18:30

configurations of the board. So it's a 19 by 19

18:33

board with black and white stones,

18:35

and the rules are very simple. It's a

18:37

turn-based game where each player simply

18:39

places

18:42

one stone on the board,

18:44

and when you surround your opponent's

18:46

stones, you remove them from the board, and the goal

18:48

is to surround your opponent.

18:51

And so it is a very

18:53

simple rule set, but it's a massively

18:55

complicated

18:57

possible set of different configurations

18:59

that can emerge.

19:00

And so you can't search all

19:02

possible branches of that

19:05

space because it's so enormous. 10

19:07

to the 170 is more atoms

19:09

than there are in the known universe, approximately. Yeah.

19:12

I think that's something like 10 to

19:14

the 80 that gets you

19:16

all the protons in the universe. So yeah,

19:18

it gets bigger still when

19:20

you're talking about Go. Right.

19:23

So this needed

19:25

a new suite of methods, and

19:28

I think it was an incredible experience

19:30

seeing AlphaGo progressively

19:32

get better and better. I mean, we already had

19:34

an inkling for this when we saw

19:36

it play the Atari games, but this was just

19:39

seismically more

19:40

complicated and vast, and yet it was

19:42

using the same basic principle, actually

19:45

the same principle that has subsequently been applied

19:47

in protein folding too.

19:50

So I think that's

19:52

what's really interesting about this is that it's the generality

19:55

of the ideas that simply scale with

19:57

more compute.

19:58

Because a couple of...

19:59

years later,

20:01

AlphaGo became AlphaZero,

20:03

which essentially achieved

20:06

superhuman performance without any

20:08

learning from prior games.

20:12

Part of the trick with AlphaGo is that it

20:14

looked at hundreds of thousands of prior

20:17

games. It's almost like the expert knowledge of

20:19

existing players that has been handed

20:21

down for centuries of playing the game.

20:24

Whereas AlphaZero

20:25

was able to learn entirely through self-play,

20:28

almost like I think the intuition is spawning

20:31

instances of itself in order to play

20:33

against itself

20:34

in simulated environments, many,

20:36

many hundreds of millions of billions of times. It

20:40

turns out to be way more valuable than bootstrapping

20:43

itself from the first principles

20:45

of human knowledge, which if you think about the

20:47

size of the state space represents a minor

20:50

subset of all possible configurations

20:52

of that board. That was a kind

20:54

of remarkable insight. Indeed, it did

20:56

the same thing for other games, including

20:58

chess and shogi and so on.

21:00

Yeah, that's a

21:03

really fascinating development where it's

21:05

now uncoupled from the

21:07

repository of human knowledge. It plays

21:10

itself, and over the course of, I think

21:12

it was just a day of self-play,

21:14

it was better than

21:17

AlphaGo and any other

21:19

system, right?

21:21

Right. That's exactly right. Obviously,

21:23

that's partly a function of compute,

21:25

but

21:25

the basic principle gives an important

21:28

intuition which is that

21:29

because these methods are so general, they

21:31

can be paralyzed and scaled up.

21:34

That means that

21:36

we can take advantage of all of the

21:39

traditional assets of computing

21:42

infrastructure rather than relying on

21:44

old-school methods,

21:46

perfect memory, paralyzable compute,

21:49

Moore's

21:49

law, daisy-chaining

21:52

compute together just like we do with GPUs

21:55

these days.

21:56

In some ways, that's

21:58

the key intuition because...

22:00

It means the sort of barrier

22:02

to application of the quality

22:04

of the algorithm is lower because

22:06

it's turbocharged by all these other

22:09

underlying drivers which are also

22:12

improving the power and performance of these models.

22:16

And also AlphaZero in

22:18

when it was playing the world champion

22:20

came up with a move that all

22:23

Go experts thought they immediately

22:25

recognized as a mistake, but

22:27

then when the game played out it turned

22:30

out to be this

22:31

brilliant novel move that no human

22:33

would have made and it just

22:36

a piece of discovered Go knowledge.

22:39

Yeah, I mean, I remember sitting

22:41

in the commentary

22:43

room live watching that unfold

22:45

and listening to

22:47

the commentator who was himself a nine

22:50

down expert

22:52

say that it was a mistake. He was like, Oh, no,

22:54

we've lost. And it took 15 minutes

22:57

for

22:57

him to correct that and sort of come

22:59

back and reflect on it was a really

23:02

remarkable moment.

23:03

And actually,

23:05

for me, it was a great inspiration,

23:07

because this is why we

23:10

started the company. I mean, the quest was to

23:13

try to invent new knowledge. I

23:15

mean, our goal here is to try

23:17

to design algorithms

23:19

that can teach us something that we don't know,

23:21

not just reproduce existing

23:23

knowledge and synthesize information in new ways,

23:26

but genuinely discover new strategies

23:29

or

23:29

new molecules or new

23:32

compounds, new ideas, and

23:34

contribute to the

23:35

kind of well of human

23:37

knowledge and capability. And this

23:39

was a kind of first, well, actually, it

23:42

was the second indication because the first instinct

23:44

I got for that was watching the Atari games player

23:46

learn new strategies from scratch. This

23:49

was kind of the second, I think.

23:59

Well, protein folding

24:02

is a long-standing challenge,

24:06

and we actually started working on this as

24:08

a hackathon, which started in my group back

24:12

in 2016.

24:13

And it was really just an experiment

24:15

to see if, you know, some of the AlphaGo

24:17

models could actually make progress

24:19

here. And the basic idea

24:21

is that if you can

24:23

sort of generate

24:24

an example of the way

24:27

a protein folds, this folding

24:29

structure represents,

24:30

might tell you something about

24:33

the value

24:35

of that molecule in practice, what it can do,

24:38

what

24:38

its strengths and weaknesses are, and so on.

24:40

And so, the nice

24:42

thing about it is because it operated in a simulated

24:45

environment, it was quite similar to some of

24:47

the games that we had been

24:48

playing, you know, teaching

24:50

our models to play. And,

24:53

you know, previously the experiments had

24:55

done something like 190,000 proteins, which is about 0.1% of all the proteins

24:57

in

24:58

existence.

25:02

But

25:03

in AlphaFold 2,

25:05

the team actually open sourced something

25:07

like 200 million protein structures all

25:09

in one go, which is sort of all known

25:12

proteins. This is a massive breakthrough

25:14

that took, you know, four or five

25:16

years of work in development. And

25:19

I think just gives an indication

25:22

of the kinds of things that become possible with

25:24

these sorts of methods.

25:25

Yeah, I forget. Someone

25:27

gave a, well, what reported to be a kind

25:29

of a straightforward

25:30

comparison between what

25:33

AlphaFold did there and

25:35

the academic years of PhD

25:38

theses. And it was something like, you

25:41

know, 200 million PhD theses got

25:43

accomplished in a few years there

25:45

in terms of solving those protein folding

25:47

problems.

25:48

Yeah, I mean, those kinds

25:50

of insights, those kinds of sort

25:53

of compressions are similar to,

25:55

you know, across the board with many technologies.

25:57

Like

25:58

another one that's sort of similar to that is... that

26:00

the amount of labor

26:02

that once produced 50 minutes

26:04

of light in the 18th

26:07

century now produces 50 years

26:09

worth of light. And that just

26:11

gives us a sense for

26:14

how technology has this massive

26:16

compressive effect that is

26:19

hugely leveraging in terms of what we can do.

26:21

Yeah, there's another crazy analogy

26:23

in your book talking about the

26:25

size of these

26:28

parameters of these new large

26:30

language models, which we'll get to. But the

26:32

comparison was something like executing

26:34

all of these floating point operations. If

26:38

every operation were a drop

26:39

of water, the

26:41

largest large language models

26:43

execute as many calculations

26:46

as would fit into the entire Pacific

26:48

Ocean. So it's just the scale

26:51

is astounding. Right.

26:53

So your book was a bit of a surprise

26:56

for me because you

26:58

are more worried than

27:00

I realized about how all of this can go

27:03

wrong. And I got

27:05

the sense in, you and I haven't spoken

27:07

very much, but

27:08

in talking to you and Demis and Shane,

27:10

I

27:11

got the sense that,

27:13

and this is these conversations are now

27:15

several years old, that you were

27:17

more sanguine about our

27:20

solving all of the relevant

27:23

problems, alignment being the

27:25

chief among them, but

27:27

other concerns of bad incentives and arms

27:29

race conditions and

27:31

et cetera.

27:33

You all were putting a fairly

27:35

brave face on

27:37

a problem that was making many

27:40

of us increasingly

27:41

shrill and not to say hysterical.

27:44

And so

27:47

I guess the most hysterical voice of

27:49

the moment is someone like Eliezer Yudkowsky.

27:52

And there was obviously Nick Bostrom and others who were

27:56

issuing fairly grave warnings

27:58

about how it was more likely

28:00

than not that we were going to screw this up and build

28:03

something that we really can't

28:05

control ultimately and that could well

28:07

destroy us. On the way to

28:09

the worst possible outcome,

28:11

there are many bad, very

28:14

likely outcomes like a

28:16

misinformation apocalypse and

28:18

other risks. But

28:21

in your book, you

28:22

don't give the risks

28:25

short shrift. I mean, you

28:27

do seem to suggest that,

28:30

and certainly when you add in the attendant

28:32

risks of synthetic biology

28:34

here, which we'll talk about, you

28:37

are quite worried and yet there's no,

28:40

as you agree with a

28:42

point I made

28:43

early on here, which is that

28:46

as worried as we are, there really is no

28:48

break to pull. I mean, the incentives are such

28:51

that we're going to build this. And so

28:53

we have to sort of figure out

28:55

how to repair the rocket as it's

28:58

taking off and align it properly as it's

29:00

taking off because there's just no getting

29:02

off this ride at the moment, despite the fact

29:04

that people are calling for a moratorium

29:07

or some people are.

29:08

So I guess before we jump into the book,

29:12

when did you get worried? Were you always

29:14

worried or

29:17

are you among the newly worried people like

29:19

Jeff Hinton, who I mean, like Jeffrey

29:22

Hinton, who you mentioned is

29:24

really the godfather of this technology. And

29:27

he just recently resigned

29:30

from Google so that he could express his worries

29:32

in public.

29:33

And he seems to have just become worried

29:36

in the presence of these large

29:38

language models.

29:39

And it's quite inscrutable to me

29:42

that he

29:44

suddenly had this change of heart because in

29:47

my view, the basis for this concern

29:49

was always self-evident. So

29:52

give me the memoir of your

29:54

concerns here.

29:55

Yeah, so this is not a new consideration

29:58

for me. I've been worried about it.

29:59

about this from the very

30:02

first days when we founded the company. And in

30:04

fact,

30:05

our strap line on our business plan

30:08

that

30:08

we took to Silicon Valley in

30:11

2010 was building artificial general intelligence

30:14

safely and ethically for the benefit of everyone. And

30:17

that was something that was critical to me all the

30:19

way through. And when we sold the company, we

30:22

made it a condition of the acquisition that we have

30:24

an ethics and safety board with some

30:26

independent members overseeing technology

30:29

in the public interest that our technologies

30:31

wouldn't be used for military purposes

30:34

like lethal autonomous weapons or surveillance

30:36

by the state. And since

30:39

then at Google, I went

30:41

through lots and lots of different efforts to experiment

30:43

with different kinds of oversight boards

30:45

and charters and external

30:48

scrutiny and independent audits and all

30:50

kinds of things. And so I'd say I've definitely been

30:53

top of mind for me all the way through. I think

30:55

where I diverge from the

30:57

sort of Bostrom camp a bit

30:59

is that I think that

31:01

the language around super intelligence

31:04

has actually been a

31:05

bit of a distraction.

31:07

And I think it was quite obviously a distraction

31:09

from fairly early

31:11

on. I think that the focus

31:13

on

31:14

this sort of intelligence

31:17

explosion, this AI that recursively

31:19

self improves and suddenly takes over everybody

31:21

and turns the world to paperclips.

31:23

I think has consumed way more

31:25

time than the

31:27

idea justifies.

31:29

And actually, I think there's a bunch

31:31

of more near term, very practical

31:34

things that we should be concerned about.

31:36

They shouldn't create trill alarmism

31:39

or panic, but they are

31:41

real consequences that if we don't take

31:43

them seriously, then they have the potential to

31:45

cause serious harm. And

31:48

if we continue down this path

31:51

of complete openness

31:52

without any sort of checks

31:54

and balances on how this technology arrives

31:57

in the world, then essentially

32:00

it has the potential to cause a great deal of chaos.

32:02

And I'm not talking about

32:04

AIs running out of control and

32:06

robots and so on. I'm really talking

32:08

about

32:09

massively amplifying the spread

32:12

of misinformation

32:13

and more generally reducing the

32:16

barrier to entry to be able to

32:18

exercise power. That

32:20

is fundamentally what this technology is.

32:23

In my book, I have a framing, which

32:25

I think is more helpful around a

32:27

modern Turing test, one that evaluates

32:30

capabilities, like what can an

32:33

AI do. And I think that we should be much

32:35

more focused on

32:36

what it can do

32:38

rather than what it can say. What it

32:40

can say is important and it has huge influence,

32:42

but increasingly it's going to have capabilities.

32:45

And so an artificial capable

32:48

intelligence, an ACI,

32:50

is something that has the potential

32:52

not just to influence and persuade,

32:55

but also to learn to

32:57

use APIs and initiate actions,

33:00

queries, calls in third party

33:02

environments. It'll be able to use

33:05

browsers and parse the pixels

33:07

on the browser to be able to click buttons and

33:09

take actions in those environments.

33:11

It'll be able to call,

33:13

phone up and speak to,

33:15

communicate with other AIs

33:17

and other humans. So

33:20

these technologies are getting smaller

33:22

and smaller and more and more capable,

33:25

are getting cheaper to build. And so

33:27

if you look out over a 10 to 20 year period,

33:30

I think the story is one of a

33:32

proliferation of power in the conventional

33:34

sense,

33:35

not so much an intelligence explosion, which

33:38

by the way, just for the record,

33:39

I think is an important thing for us to think about.

33:42

And I care very deeply about

33:44

existential risk and AGI safety,

33:47

but I think that the more practical

33:49

risks are not getting enough consideration.

33:52

And that's actually a big part of the book. In

33:54

no way does that make me a pessimist. I mean,

33:56

I'm

33:57

absolutely an optimist.

33:59

I'm hoping.

33:59

hopeful and positive about technology. I want to

34:02

build things to make people's lives

34:04

better and to help us create more value in

34:06

the world and reduce suffering. And I think

34:08

that's

34:09

the true upside of these technologies, and we will

34:11

be able to deliver them

34:12

on that upside.

34:13

But no technology comes without risk,

34:16

and we have to

34:17

consciously and proactively attend

34:20

to the downsides.

34:22

Otherwise,

34:24

we haven't really achieved our full

34:26

objective. And that's the purpose of speaking up about

34:28

it. Well,

34:29

before we get into details

34:31

about the downsides,

34:33

let's talk about how this

34:34

might go well.

34:37

I guess before we talk

34:39

about the upside, let's just define the

34:41

terms in the title of your book. The

34:43

title is The Coming Wave. What

34:45

is the coming wave?

34:47

So when you look back over

34:49

the millennia, there have been

34:52

waves of general purpose technologies,

34:55

from fire to the invention of the wheel to

34:57

electricity.

34:59

And each of these waves,

35:01

to the extent that they have been lasting and valuable,

35:03

are general purpose technologies,

35:06

which enable other technologies. And that's

35:08

what makes them a wave. They're enablers of other

35:10

activity. They're general purpose in

35:12

nature.

35:14

And as they get more useful,

35:16

naturally, people experiment

35:18

with them. They iterate, they invent, they

35:20

adapt them, and they

35:23

get cheaper and easier to use. And

35:25

that's how they proliferate.

35:27

So in the history of technologies,

35:29

all technologies that have been useful, that are real

35:32

general purpose technologies, have spread far

35:34

and wide and got cheaper.

35:36

And

35:37

almost universally, that is an incredibly

35:39

good thing. It has transformed

35:42

our world.

35:43

And I think that that's

35:46

an important but very simple

35:48

concept to grasp. Because

35:51

if that is a law of technology,

35:53

if it is a fundamental property

35:55

of the evolution of technology, which

35:58

I'm arguing it is, then

35:59

that happens. real consequences for the

36:01

next wave, because the next wave

36:04

is a wave of intelligence

36:07

and of life itself. Intelligence

36:11

is the ability to take actions. It

36:14

is the ability to synthesize

36:17

information,

36:18

make predictions,

36:20

and affect the world around you.

36:23

It's almost the definition of power.

36:25

Everything that is in our visual

36:28

sphere, everything in our world, if you look around you

36:30

at this very minute today, has

36:31

been affected

36:33

in a very material way by intelligence.

36:36

It is the thing that has produced all of the value

36:39

and all of the products and all of

36:41

the affected the landscape that you can

36:43

see around you in a huge

36:45

way.

36:46

And so the prospect of being able to

36:48

distill what makes us unique

36:50

as a species into

36:52

an algorithmic

36:54

construct that can benefit from being

36:56

scaled up and paralyzed, that can

36:58

benefit from

36:59

perfect memory and compute and consuming

37:02

vast amounts of data, trillions

37:04

of words of data,

37:06

is enormous. That in itself

37:08

is almost like gold.

37:12

It's like being able...it's like alchemy. It's like

37:14

being able to capture the essence of what has

37:16

made us capable

37:18

and add more

37:19

knowledge and

37:21

essentially science and technology

37:23

into the

37:24

human ecosystem. So

37:26

imagine that

37:27

everybody will now in the future,

37:29

in 10 years, 15 years, have access to

37:32

the very best doctor

37:35

in the world, the very best

37:37

educator,

37:39

the very best personal assistant and chief

37:41

of staff. And any one of these roles,

37:44

I think, is going to be very,

37:46

very widely available to billions of people.

37:48

And people often say to me, well, aren't

37:52

the rich going to benefit first? Or is it going to be unfair

37:54

in terms of access? Yes, for a

37:56

period of time.

37:57

That's true.

37:58

But we're actually living...

37:59

in one of the most meritocratic moments

38:02

in the history of our species.

38:04

Every single one of us, no matter how

38:06

wealthy you are, every one of us

38:08

in the Western world,

38:10

really the top 2 billion people on the planet,

38:12

have access to the same smartphone.

38:16

No matter how much you earn, you cannot buy

38:18

a smartphone or a laptop that

38:20

is better than the very richest.

38:22

That's an unbelievably meritocratic moment

38:25

that is worth really meditating on.

38:28

And that is largely a function

38:30

of these exponentials. The

38:33

cost of chips has exponentially

38:35

declined over the last 70 years,

38:38

and that's driven mass proliferation.

38:40

And if intelligence and life are

38:43

subject to those same exponentials,

38:45

which I think they are, over the next two to

38:47

three decades,

38:49

then the primary trend that we have

38:51

to cope with in terms

38:53

of our culture, and our politics and commerce,

38:56

is this idea that intelligence,

38:58

the ability to get stuff done,

39:00

is about to proliferate.

39:02

And that's going to produce a

39:04

Cambrian explosion of productivity.

39:08

Everybody is going to get access to a tool that enables

39:10

them to pursue their agenda, to make us all

39:12

smarter and more productive and more capable.

39:15

So I think it might be one of the most

39:18

productive periods in the

39:20

history of humanity.

39:22

And I think, of course, the challenge

39:24

there is that it may also be one

39:26

of the most unstable over the next 20 years.

39:29

Yeah. So that

39:31

cornucopia

39:33

image immediately

39:35

begets the downside

39:38

concern of

39:39

massive labor disruption, which

39:42

many people doubt in principle. They

39:44

just think that we've learned over the course

39:46

of the last 200 years of

39:49

technological advancement and economic

39:51

thinking that there is no such

39:53

thing as a true

39:56

canceling of a need for human

39:58

labor.

39:59

And so people people draw the obvious analogies

40:01

from agriculture and other previous

40:03

periods

40:04

of labor disruption

40:06

and

40:08

conclude that

40:09

this time is no different and while

40:12

there might be a few hiccups,

40:15

what's going to happen here is that all

40:17

of these productivity gains and job

40:19

canceling innovations born

40:22

of AI will

40:24

just open new lanes for

40:27

human creativity and there'll be better

40:29

jobs and, you know, we're

40:32

just as we were happy to get rid of jobs

40:34

in agriculture and coal mines

40:37

and open them up in the service sector,

40:40

we're going to do the same with AI. I remain

40:42

quite skeptical

40:45

of that this

40:47

time is the same given

40:49

the nature of the technology. This is the as

40:51

you just said, this is the first moment

40:54

where we are

40:56

envisioning a technology

40:59

which is a true replacement

41:02

for human intelligence.

41:04

If we're talking about general intelligence

41:07

and we're talking about the competence

41:09

that you just described, the ability

41:12

to do things in addition to saying

41:14

things,

41:15

we are talking about the

41:17

cancellation of human

41:20

work,

41:21

at least in principle, and

41:24

strangely, I mean, this

41:26

is not a terrible surprise

41:28

now, but it would have been a surprise probably 20 years

41:30

ago, this

41:31

is coming for

41:33

the higher cognitive, higher status,

41:35

white-collar jobs before it's coming for

41:38

blue-collar jobs. How do you view

41:40

the prospect of labor

41:42

disruption here, and how

41:46

confident are you that everyone

41:48

can be retrained

41:51

with their nearly omniscient

41:53

AI

41:54

assistants and chiefs of staffs

41:56

and find something worth

41:59

doing that other people

41:59

people will pay them to do.

42:01

Yeah, I mean, I'm with you. I've

42:04

long been skeptical of people who've

42:06

said that this will be just like

42:09

the agricultural revolution, or

42:11

this will be like the horse and cart and

42:14

cars. People will

42:16

have more wealth.

42:18

The productivity will drive wealth

42:20

creation, and then that wealth creation will drive

42:22

demand for new products. And we couldn't possibly imagine

42:25

what people are going to want to consume

42:28

and what people are going to create with this new

42:30

wealth and new time. And that's typically

42:32

how the argument goes. And I've

42:35

never found that compelling. I mean,

42:37

I think

42:38

that if you look at it, it's been quite

42:41

predictable the last decade. I mean,

42:43

these models are deliberately

42:46

trying to

42:47

replace human cognitive abilities.

42:50

In fact, they have been slowly climbing

42:52

the ladder of human cognitive

42:55

abilities for many years. I mean, we started

42:57

with image recognition and

43:00

audio recognition, and then moved

43:02

on to audio generation,

43:05

image generation, and

43:06

then text, understanding,

43:09

text recognition, and then

43:11

now text generation. And

43:13

it was kind of interesting because

43:15

if you think even just two or

43:17

three years ago, people would have said,

43:20

well, AIs will never

43:22

be creative. That's not achievable.

43:25

That creativity will always be

43:27

the preserve of humans, and

43:30

judgment is somehow unique and special to

43:32

what it means to be human. Or like AIs

43:34

will never have empathy. We'll always be able

43:36

to do care work, and emotional

43:38

care is something that's special. You can never

43:40

replace that connection. I mean, both

43:42

of those are now self-evidently not

43:44

true, and I think have been quite predictable.

43:47

So I think that

43:49

the honest way to look at this is that these

43:51

are only temporarily augmenting of human

43:53

intelligence. If you think about the trajectory

43:56

over 30 years, I mean, let's not quibble over whether

43:58

it's five years, 10 years, or 15 years. engineers, just

44:00

think about it long-term. I think we can all agree long-term.

44:03

If these exponential trajectories

44:05

continue, then they're

44:08

clearly only temporarily going

44:10

to turbocharge an existing human.

44:13

And so we have to really think, okay, long-term,

44:15

what does it mean to have systems that

44:18

are this powerful, this cheap,

44:20

this widely proliferated?

44:22

And that's where I think the broad concept

44:25

I have in the book of containment comes in, because you

44:27

can start to get an intuition for

44:29

the massive consequences of

44:31

the spread of this kind of power,

44:33

and then start to think about what are the sorts

44:36

of things we would want to do about it. Because on the face

44:38

of it, like you said earlier, the incentives

44:40

are absolutely overwhelming. Technology

44:43

has always been a machine

44:45

of statecraft. It's been used

44:48

by militaries and used by nation states to

44:50

serve citizens and drive

44:53

us forward. And now

44:54

it is the fundamental driving force of

44:56

nation states, being commercially

44:58

competitive, having the best companies, having

45:01

the best labor market,

45:02

that drives our competitive edge. So

45:05

from a state perspective, a nation state

45:07

perspective, from an individual

45:09

scientific perspective, the huge

45:12

drive to explore and invent and discover,

45:14

and of course from a commercial

45:16

perspective, the profit

45:18

incentive is phenomenal. And all of these

45:20

are good things, provided they can be

45:23

well managed and provided we can mitigate

45:25

the downsides. And I think we have to be focused

45:28

on those downsides and not be afraid to talk

45:30

about them. So I definitely

45:34

experience when I bring up these topics over

45:36

the years, this kind of, what

45:39

I describe in the book as a pessimism aversion.

45:41

There's people who are just sort

45:44

of constitutionally unable to

45:46

have a dark conversation about how things

45:48

may go wrong. And I'll get accused

45:51

of not being an optimist or something

45:53

as though that's like a sin

45:55

or something, or that being a pessimist

45:57

or an optimist is somehow...

45:59

a good way of framing things. To me, both

46:02

are biased. I'm just observing the facts

46:04

as I

46:06

see them. And I think that's an important

46:08

misconception and unhelpful

46:10

framing of pessimism and optimism, because

46:13

we have to start with our best assessment

46:15

of the facts and try to reject those facts

46:17

if they're inaccurate in some way,

46:19

and then try to collectively predict

46:21

what the consequences are going to be like. And

46:24

I think it's another trend over the last decade

46:26

or so. Post-financial crisis,

46:29

I feel like people,

46:30

public intellectuals and elites in general

46:32

and everyone in general, has sort of just

46:34

got a bit allergic to predictions. We've

46:37

got a bit scared of being wrong. And I think

46:39

that that's another thing that we've got to shed.

46:41

So we've got to focus on trying to make some of these

46:44

predictions. They may be wrong. I may have

46:46

got this completely wrong, but it's

46:48

important to lay out a

46:50

case for what might happen and

46:53

start taking steps towards mitigation

46:55

and adaptation. Well,

46:57

you invoke the concept of containment,

47:00

which does a lot of work in the book, and you have

47:02

this phrase, the containment problem

47:05

that you use throughout. What

47:07

is the containment problem?

47:09

In its most basic form, the idea

47:12

of containment is that we

47:14

should be able to demonstrate to ourselves that

47:16

technologies that we invent

47:19

should always be

47:20

accountable to humans

47:23

and within our control. So it's the

47:25

ability to close down or

47:27

constrain or limit a new technology

47:30

at any stage of its development or

47:32

deployment.

47:34

And that's a grand claim.

47:36

But actually put in the most simple terms,

47:39

it basically says we shouldn't

47:41

allow technologies to run out of our control.

47:43

If we can't say what

47:46

destiny we want for how a technology

47:48

impacts our species, then we're at the

47:50

mercy of it. And I think the

47:54

idea is if we don't

47:56

have mechanisms to shape that and restrict

47:58

its capabilities... then it potentially

48:02

leads us into some quite catastrophic

48:04

outcomes over a 30-year period. Do

48:08

you think

48:09

we've lost the moment

48:12

already? I mean, it seems like the digital

48:14

genie is more or less out of the bottle.

48:17

I mean, this is something that, if anything

48:19

surprised me, and I know

48:22

certainly surprised the people who were more

48:24

focused on AI safety, and again,

48:26

people like Yudkowsky, in

48:28

recent developments around these LLMs, was

48:31

that we missed a moment that many

48:34

of us, more or less expected, were more

48:36

or less sure was coming, which was there'd

48:39

be a breakthrough at some company

48:41

like DeepMind where

48:43

the people building the technology

48:46

would recognize that they had finally gotten

48:48

into the end zone or close

48:51

enough to it so that they're now in the presence of something

48:53

that's fundamentally different

48:55

than anything that's come before, and

48:58

there'd be this question, okay, is

49:00

this safe to work

49:02

with? Is this safe to release

49:04

into the wild? Is this safe to create

49:07

an API for? So

49:10

the idea was that you'd have this digital

49:13

oracle in a box that

49:16

would already have

49:18

been air gapped from the internet and incapable

49:21

of doing anything until we let it out, and

49:23

then the question would be, have we done enough

49:25

safety testing to let it out? But now

49:28

it's pretty clear that everything

49:30

is already more or less out, and we're building

49:32

our most powerful models

49:35

already in the wild, right?

49:38

And they're already hooked up to things, and they

49:40

already have millions of people playing with them,

49:42

and they're open source versions of the

49:44

next best model. And so is containment

49:48

even a dream at this point?

49:50

So it's definitely not too late.

49:52

We're a long, long way away. This is

49:54

really just the beginning. We have

49:57

plenty of time to address.

49:59

this. And the more

50:02

that these models and these ideas

50:04

happen in the open,

50:06

the more they can be scrutinized and

50:08

they can be pressure tested and held accountable.

50:10

So I think it's great that they're happening in

50:12

open source at the moment. So,

50:15

you like Sam Altman. This is what

50:17

Sam has always said, that the

50:20

philosophy behind OpenAI is do this stuff

50:23

out in the open, let people play

50:25

with it, and we will learn a lot as

50:28

we get closer and closer to building

50:30

something that we have to worry about.

50:32

I think that we have to

50:34

be humble about the practical

50:36

reality about how these things emerge. So

50:39

the initial framing

50:42

that it was going to be possible to invent this

50:44

Oracle AI that stays in a box, and

50:46

we'll just probe it and poke it and test it

50:48

until we can prove that it's going to

50:50

be safe, and that we'll

50:52

stay in the bunker and keep it hidden from everybody.

50:55

I mean, this is a complete nonsense and it's attached

50:57

to the super intelligence framing. It

50:59

was just a completely wrong metaphor

51:02

that totally ignores the history of all technologies.

51:04

And actually, this is one of the core motivations for me in

51:06

the book is that I had time during the pandemic

51:09

to really sleep and reflect

51:11

and really deeply think, okay, what is

51:13

actually happening here on a multi-century

51:16

scale? And what are the patterns of history

51:19

around how inventions end

51:21

up proliferating? And it's

51:24

really stating the obvious, it's almost like ridiculously

51:26

simplistic, but it needed to be said that actually,

51:30

as soon as something as an idea is

51:32

invented,

51:33

millions of other people

51:35

have approximately the same idea

51:37

within just weeks, months, years,

51:39

especially in our modern digitized

51:42

world. And so we should

51:44

expect, and as we do see, the open

51:46

source movement to be right

51:48

hot on the heels of the

51:50

absolute frontier. And so,

51:52

I mean, just one small example

51:54

of that to give an intuition,

51:56

GPT-3, it was

51:58

launched in the summer of 2020.

51:59

So three years ago, 175 billion

52:02

parameters, and is now regularly

52:05

being trained at 2 billion parameters.

52:08

And so that is a massive

52:10

reduction in serving cost, you

52:12

know, that now means that people

52:15

can have open source versions

52:17

of GPT-3 that have broadly

52:19

the same capabilities, right, but

52:21

are actually extremely cheap to serve and

52:24

indeed to train.

52:25

So if that trajectory

52:28

continues,

52:29

then we should expect that what is

52:31

cutting edge today, frontier models

52:33

like ARSA inflection and like

52:36

GPT-4, GPT-3.5

52:39

even, will be open source in the next

52:41

two to three years. And so what does it

52:43

mean that those capabilities are available

52:46

to everybody, right? And I think that is a great

52:48

thing for where we are today. But if

52:50

the trajectory of exponentially

52:52

increasing compute and size of models

52:55

continues for an other three,

52:57

four, five generations, which we all

52:59

expect it to,

53:00

then that's a different question. We have

53:03

to step back and honestly ask ourselves, what does it mean

53:05

that this kind of power is going to proliferate

53:07

in open source, number one?

53:09

And number two, how do we hold accountable

53:11

those who are developing these mega models, even

53:14

if they are centralized and closed, myself

53:16

included, OpenAI, DeepMind, et cetera?

53:18

And if you just look at the amount of compute, it's predictable

53:22

and breathtaking. And I think people forget

53:24

how predictable this is.

53:26

So going back to Atari DQN,

53:28

we developed that model in 2013 and it used two

53:30

petaflops of

53:35

computation, right? So

53:38

a petaflop

53:39

is a billion million

53:41

operations, right? So imagine

53:43

a billion people,

53:44

each holding one million calculators each

53:47

and doing a complex calculation

53:50

all

53:50

at the same time pressing equals, right?

53:52

So that's, that would be one petaflop and Atari

53:55

used two petaflops over several

53:57

weeks of computation.

53:59

later, the cutting

54:01

edge models that we develop at Inflection for

54:03

Pi, our AI, use

54:07

five billion times the

54:09

compute that was used to play Atari DQN.

54:11

So 10 billion, billion, million.

54:14

It's just like-

54:17

Now you're sounding like Ali G.

54:19

Exactly. That's

54:23

basically 10 orders of magnitude

54:25

more compute in a decade. So

54:27

one order of magnitude every year. So 10X

54:30

every year for 10 years,

54:32

which is way more than

54:34

Moore's Law. Everyone's familiar with Moore's Law, 70

54:37

years of doubling, doubling every 18 months

54:39

or whatever. I mean, that is miniscule

54:42

by

54:42

comparison now. Of course, there's a very good handle, maybe. If

54:45

you'd like to continue listening to this conversation, you'll

54:48

need to subscribe at SamHarris.org. Once

54:51

you do,

54:51

you'll get access to all full-length episodes of

54:53

The Making Sense podcast, along with other

54:55

subscriber-only content, including

54:57

bonus episodes and AMAs, and

55:00

the conversations I've been having on the waking up app.

55:03

The Making Sense podcast is ad free and

55:05

relies entirely on listener support, and

55:07

you can subscribe now at SamHarris.org.