0:02

I'm Reed Hoffman. And I'm Aria

0:04

Finger. We want to know

0:06

what happens if, in the

0:09

future, everything breaks humanity's way.

0:11

With support from Stripe, we

0:13

typically ask our guests for

0:16

their outlook on the best

0:18

possible future. But now, every other

0:20

week, I get to ask Reed

0:22

for his take. This is

0:24

possible. All

0:29

right, Reed, so lovely to be here

0:32

with you. So recently, Toby Luka, the

0:34

CEO of Shopify, put out a memo

0:36

that was covered by everyone because it

0:39

had to do with internal employees

0:41

at Shopify and what they were expected

0:43

to do with AI. It talked

0:45

to them about how if you're going

0:47

to request more resources for your team,

0:50

you actually better check if AI could

0:52

do the job better or faster,

0:54

and you actually don't need that additional

0:56

headcount. It also said, like, everyone

0:58

here should be expected to be using

1:00

AI every day. And he said, I'm

1:03

the CEO, I'm no different. We want

1:05

to grow by 20, 30, 40% of

1:07

year, every employee needs to grow as

1:09

well. I think some people were shocked

1:11

by this memo. Other people found it

1:13

reasonable. What did you think about the

1:15

contents of the memo and also

1:17

Toby sort of putting out this

1:20

bold statement for the industry? You

1:22

know, I also found Toby's memo

1:24

to be exactly right. you know,

1:26

Toby does. And then also thinking

1:28

about kind of his classic technologists,

1:30

because he's obviously an engineer, is

1:33

how we use tools, you know,

1:35

AI as amplification intelligence, how it

1:37

is that we get, you know,

1:39

super agency through doing this. And

1:42

his memo, I think, is exactly

1:44

the kind of thing that I

1:46

think everybody, not just technology companies,

1:48

should be doing, every single CEO

1:51

of anything from a, you know, five to

1:53

seven person company. to a tens

1:55

of thousands of people company should

1:57

look at that and say what's

1:59

my version of how I should do that and

2:01

how I should integrate it. And Toby

2:03

obviously has given it enough thought to

2:06

kind of say, look, here are some,

2:08

you know, key, you know, kind of checkpoints

2:10

that work within companies, which is

2:12

you ask for resources, make sure

2:14

that you have, if you're going

2:16

to ask more resources, how you're

2:18

asking for more resources, in a

2:21

context of, and here is how

2:23

I'm already using AI, and here

2:25

are the reasons I need more

2:27

resources, given how I'm using AI,

2:29

either AI, lacks or the AI

2:31

opportunities from doing it. So one

2:33

of the things that I've been

2:36

telling, you know, kind of my

2:38

portfolio companies is to actually have

2:41

kind of weekly monthly check-ins where

2:43

everyone has to bring a, and

2:45

here is the new thing I've

2:48

learned about how to use AI

2:50

to help me do my job,

2:52

help us perform better in our

2:55

mission as a company for doing

2:57

that. Because The answer is

2:59

if you actually haven't found

3:01

something that was useful to

3:03

you, it's useful to your

3:05

group, it's useful to your

3:07

company, you haven't tried hard

3:09

enough. I think Toby's memo is

3:12

the kind of thing that in fact,

3:14

you know, CEOs and all group leaders

3:16

should be looking at saying, great, how

3:18

do I build on that? Thank you

3:20

for the open source, you know, kind

3:23

of, you know, kind of management technique.

3:25

What are the things... that

3:27

I should do specifically for our

3:29

group, for our company, for our

3:32

mission, for our culture, what is

3:34

our version of that, and then

3:36

start iterating in the same way. I

3:38

mean I have to admit last week on

3:40

LinkedIn I saw a marketing agency and they

3:42

sent on LinkedIn you know we promise our

3:44

clients that you will never get a image

3:46

that was started in an image generator that

3:48

was using AI we promise you you're never

3:51

going to get a tagline from us where

3:53

we use chat gPT to create it and

3:55

I literally had to look at the posting

3:57

date because I thought it was April Fool's

3:59

joke. And it wasn't. And like, I

4:01

get the nervousness and the scared

4:04

about your job and about the

4:06

future, but I just sort of

4:08

couldn't imagine that this marketing agency

4:10

was essentially doing the exact opposite

4:12

of Shopify and sort of banning

4:14

AI in their workplace. I'm sure

4:16

it befuddles you as much as

4:18

it befuddles me. Well, I mean,

4:20

I think generally speaking, that's the

4:22

similar idiocy in the education space

4:25

saying our students shouldn't use chat

4:27

chat GBTT. because the whole answer

4:29

is you're preparing them for the

4:31

future, you're preparing them for being

4:33

citizens, for being workers, for being,

4:36

you know, you know, people who

4:38

are navigating life. And here is

4:40

this fundamental tool. It's kind of like

4:42

saying, hey, none of our people. can

4:45

use anything that uses electricity. And that's

4:47

how they learn. They have to use

4:49

pencils and papers and no electricity whatsoever

4:52

in anything. You're like, well, I would

4:54

idiotic close similar to the chat DVD.

4:56

And so I think that marketing agency,

4:58

the question is really, when is

5:00

it going to have to shift or it's

5:03

probably going to die or become very esoteric

5:05

boutique. Right. Absolutely. If you

5:07

want to be the most boutique agency,

5:10

perhaps that's the way to go. So

5:12

another concern people have with AI though

5:14

is misinformation, disinformation, all of this synthetic

5:17

media that was created. And actually last

5:19

week, and this wasn't created by AI,

5:21

this was a tweet that sent the

5:23

market, made an $8 trillion worth of

5:25

market volatility because someone tweeted that the

5:28

tariffs were off when they in fact

5:30

were not. And so... If a single

5:32

tweet can move the market by $8

5:35

trillion, like what does this mean for

5:37

the future when disinformation, misinformation is increasing

5:39

and perhaps with algorithmic trading and AI

5:42

able to do this at sort of

5:44

greater quantities and greater speeds, like how

5:46

do we protect against that for

5:48

the future? There's a combination

5:51

of a free market response, which I

5:53

think is partially correct, and a societal

5:55

response, which is also partially correct.

5:58

And so that's the balance of...

6:00

makes us challenging. So the free

6:02

market responses simply say, well, if

6:04

people who are doing trading are

6:06

going to be idiots and not

6:09

track false posts and so forth,

6:11

they're going to lose money and

6:13

eventually they will be disempowered. And

6:16

so what you principally need to

6:18

do is to just make sure

6:20

that there is validated sources of

6:23

information, you know, kind of are

6:25

the anchors and then to increase

6:27

that validation. you know accuracy availability

6:29

and then allow the market to

6:31

sort it out and that's a

6:34

partial answer and my principal you know

6:36

thought there is like we should not be trying

6:38

to restrict technology as much as

6:40

we should be trying to shape

6:42

technology because the question isn't like

6:45

let's not have algorithmic training and

6:47

it's like okay that's kind of

6:49

foolish it's Let's have algorithmic trading

6:51

work in the following way,

6:53

generating the following awards, making

6:56

sure it's involving the following

6:58

kinds of data is only

7:00

deployable by entities that have

7:02

a method by which they

7:04

participate in the market in

7:06

a way that is healthy

7:08

for not creating, you know,

7:10

crazy volatility swings that damaged society.

7:12

It's a little bit like similar

7:14

to saying. Hey, you know, car

7:16

manufacturers don't want to manufacture seat

7:18

belts, drivers don't want to wear

7:20

seat belts, but actually in fact,

7:22

because the cost to the society

7:24

and the health care system and

7:26

everything else is so high, like

7:28

you would say, hey, it's a

7:30

free market, you're going to take

7:33

their risk and they're going to

7:35

dive. It's like, no, no, no,

7:37

actually in fact, there's so many

7:39

injuries and so many costs here,

7:41

and the cost of enforcing you

7:43

to wear a seat belt seat

7:45

belt is very low. for making

7:47

the overall system work is I

7:49

think an ongoing and kind of

7:52

thoughtful thing that banks and regulators

7:54

and intellectuals and economists should think

7:56

about like what are those minimal

7:58

kind of also where shaping

8:00

technology or technology

8:02

ads that keep the cost

8:04

of the cost of not

8:06

having an overly centralized system

8:09

and the benefits of all

8:11

the free market and broad

8:14

network, you know, working while

8:16

navigating the fact that we

8:18

kind of live in a more volatile

8:21

space now. On

8:26

this podcast, we like to focus on

8:28

what's possible with AI, because we know

8:30

it's the key to the next era

8:32

of growth. A truth well understood

8:34

by Stripe. Makers of Stripe billing,

8:36

the go-to monetization solution for

8:39

AI companies. Stripe knows that

8:41

when launching a new product,

8:43

your revenue model can be just

8:45

as important as the product itself.

8:48

In fact, every single one of

8:50

the Forbes Top 50 AI companies

8:52

that has a product on the

8:55

market today uses Stripe to monetize

8:57

it. See what Stripe can do

8:59

for your business at stripe.com. On

9:06

a lighter note, if there's any parents out

9:09

there who are navigating this, I just

9:11

read the book Lemonscello's library with my

9:13

nine and seven year olds and a

9:15

main plot point is a fake Wikipedia

9:17

post that leads to ruining someone's reputation

9:19

and the kids who like don't believe

9:21

it. So anyway, try that out if

9:23

you're looking to teach your kids about

9:25

misinformation on the internet. But actually moving

9:27

on to another thing that people think

9:29

is childlike and play, one of the

9:31

fun things about our conversation with Demis

9:34

Asabas last week was we talked about

9:36

games. And it was so clear that

9:38

Demis grew up playing chess and games

9:40

were so important to him, both in

9:42

terms of his scientific research, but also

9:44

in the progression of AI, whether it

9:46

was Alpha Go or the famous IBM

9:48

Watson chess chess competition. And so when

9:50

you think about the future as AI

9:52

is more in mesh in our daily

9:54

lives, like will that give. Humans the

9:56

opportunity to play more are we going

9:58

to be playing with AI? Are we

10:00

going to be interacting it solo with

10:02

teams as a game? Like, how do

10:04

you see that connection between games and

10:06

sort of our AI future? There's

10:08

a fun book, which, you know, Demis

10:11

also knows about homoludence, which is like,

10:13

we're not just sapiens, we're game players.

10:15

Obviously, you know, I have this version

10:17

of homo technically. because I think part

10:20

of games is as technologies and the

10:22

technologies that enable different kinds of gameplay

10:24

as part of it, but games is

10:27

a way we think. And as you

10:29

know, you know, I tend to approach

10:31

like most of my strategic thinking through

10:33

the lens of games. So it's like,

10:36

it's like with a startup, what's your

10:38

theory of the game, with creating.

10:40

a book super agency, what's your

10:43

theory of the game? And so,

10:45

because game playing brings tactics and

10:47

strategies and transformation, like large language

10:50

transformers together, and also has a

10:52

notion of increasing learning and competence

10:55

because how well you're playing the

10:57

game, what are the conceptual tools

10:59

you're bringing to it, etc. So,

11:02

so games is a way that we operate

11:04

across, you know, kind of,

11:06

called intelligent experience, like

11:08

almost like is species X intelligence,

11:10

how do they play games is

11:12

actually in fact, you know, kind

11:14

of directly correlated to that. It's

11:17

one of the reasons why we

11:19

know that other kinds of mammals

11:21

and other things have intelligence because

11:23

we see dolphins playing games. We

11:25

see chips playing games. We play

11:27

games with our dogs and we

11:29

play games with with our cats

11:31

and kind of that initiating gameplay

11:33

and everything else is part of

11:35

how that tends to operate. We

11:37

don't just play games solo. We

11:39

don't just play solitaire. We don't

11:41

just play games one-on-one. We play

11:44

games as teams. You know, sports

11:46

games and all of us. And

11:48

that's part of how you model

11:50

what companies go. And when it

11:53

gets to this kind of

11:55

superagency future of saying, well,

11:57

how it is that we're deploying? It's

11:59

like, with agents. I should deploy with

12:01

these tools. And by the way, we

12:04

as teams should deploy with these tools.

12:06

We as companies should deploy with these

12:08

tools. We as individual scientists

12:10

as groups of scientists should deploy

12:12

with these tools. And that's the

12:14

kind of the pattern that we're

12:16

on and that's a the model

12:18

of games is a good way for us

12:21

thinking about it. But it's also a

12:23

good way for thinking about like how

12:25

do we construct these devices and

12:27

also how do we interact

12:29

with them? like the very

12:31

first genius moment that you know

12:33

Demis and Shane and Mustafa

12:35

brought to scalable AI was

12:38

realizing here is a way

12:40

you can apply scalable computer

12:42

learning systems to

12:44

creating amazing cognitive

12:47

as opposed to like we program

12:49

the AI the AI learns and it

12:51

learns at scale because you can

12:53

use self-play as a way

12:56

of doing it. And that was actually

12:58

part of like seeing this genius moment

13:00

by them was part of what got

13:03

me back in AI from my you

13:05

know kind of my undergraduate days where

13:07

I concluded that they that the the

13:10

the the mindsets and program and

13:12

programming AI would actually not work

13:14

you know I hadn't gone to

13:16

well what what are the

13:18

scalable compute learnable systems

13:20

because back then by

13:22

the way like a

13:24

single computer was super

13:26

expensive let alone like

13:28

you know how do

13:30

you create a server

13:32

farm of a hundred

13:34

thousand kind of working

13:36

in concert and all

13:39

the rest and and

13:41

by the way the

13:44

computers back then were

13:46

less powerful than

13:49

the smartphone that's

13:52

in your pocket.

13:54

Possible is produced by Katie

13:57

Sanders, Eddie Allard, Sarah Schlead,

13:59

Vanessa Haynes. Andy, Leah

14:01

Yates, Palomo Moreno

14:03

Jimenez, and Malia

14:06

Agudello. Jenny Kaplan

14:08

is our executive

14:10

producer and editor.

14:13

Special thanks to

14:15

Suria Yalomanchili, Saeda

14:18

Sepiava, Thanasi Delos,

14:20

Ian Alice, Greg

14:22

Viato, Park Patel,

14:24

and Ben Relis.