0:01

Most of the content created on the internet is created by

0:03

average people. And so kind of the content on average, you

0:05

know, as a whole on average, is average.

0:09

The test for whether your idea is good is how much can

0:11

you charge for? Can you charge the value?

0:14

Or are you just charging the amount of

0:16

work it's going to take the customer to

0:18

put their own wrapper on top of OpenAI?

0:22

The paradox here would be the cost

0:24

of developing any given piece of software

0:26

falls, but the reaction to that is

0:28

a massive surge of demand for software

0:30

capabilities. I think

0:32

this is one of the things

0:34

that's always been underestimated about humans

0:36

is our ability to come up

0:38

with new things we need. There's

0:41

no large marketplace for data. In

0:43

fact, what there are is there are very small markets for data. In

0:47

this wave of AI, Big Tech

0:49

has a big compute and data advantage.

0:52

But is that advantage big enough to drown out all

0:54

the other startups trying to rise up? Well,

0:57

in this episode, A16Z co-founders Mark and

0:59

Dreeson and Ben Horowitz, who both, by

1:01

the way, had a front row seat

1:03

to several prior tech waves, tackle

1:05

the state of AI. So

1:08

what are the characteristics that will define

1:10

successful AI companies? And

1:12

is proprietary data the new oil? Or

1:14

how much is it really worth? How

1:17

good are these models realistically going to get? And

1:19

what would it take to get 100 times better? Mark

1:23

and Ben discuss all this and more, including

1:25

whether the venture capital model needs a refresh

1:28

to match the rate of change happening all

1:30

around it. And of course,

1:32

if you want to hear more from Ben and Mark,

1:34

make sure to subscribe to the Ben and Mark podcast.

1:37

All right, let's get started. It

1:41

is kind of the darkest side of capitalism

1:44

when a company is so greedy, they're willing

1:46

to destroy the country and maybe the world

1:48

to like just get a little extra profit.

1:50

When they do it, like the really kind

1:52

of nasty thing is they claim, oh, it's

1:54

for safety. You know, we've created an alien

1:57

that we can't control, but we're not going

1:59

to stop working We're going to keep building it

2:01

as fast as we can, and we're going to buy

2:03

every freaking GPU on the planet. But we

2:05

need the government to come in and stop

2:07

it from being open. This is literally the

2:09

current position of Google and

2:12

Microsoft right now. It's crazy. The

2:16

content here is for informational purposes only,

2:18

should not be taken as legal, business,

2:20

tax, or investment advice, or be used

2:22

to evaluate any investment or security, and

2:25

is not directed at any investor or

2:27

potential investors in any A16Z fund. Please

2:30

note that A16Z and its affiliates may

2:33

maintain investments in the companies discussed in

2:35

this podcast. For more

2:37

details, including a link to

2:39

our investments, please see a16z.com/disclosures.

2:42

Hey folks, welcome back. We have an exciting show

2:44

today. We are going to be discussing the very

2:47

hot topic of AI. We are

2:49

going to focus on the state of AI as it

2:51

exists right now in April of 2024, and

2:53

we are focusing specifically on the intersection of

2:56

AI and company building. So

2:58

hopefully this will be relevant to anybody working on a startup

3:00

or anybody at a larger company. We have

3:02

as usual solicited questions on X, formerly known

3:04

as Twitter, and the questions have been fantastic.

3:06

So we have a full lineup of listener

3:08

questions and we will dive right in. So

3:11

first question, so three questions on the

3:13

same topic. So Michael asks, in anticipation

3:15

of upcoming AI capabilities, what

3:18

should founders be focusing on building right now? Gwen

3:21

asks, how can small AI startups compete

3:23

with established players with massive compute and

3:25

data scale advantages? And

3:27

Alistair Maclay asks, for startups building on

3:29

top of open AI, et cetera, what

3:32

are the key characteristics of those companies that will

3:34

benefit from future exponential improvements in the base models

3:36

versus those that will get killed by them? So

3:39

let me start with one point, Ben, and then we'll jump

3:41

right to you. So Sam Altman recently gave an interview, I

3:43

think maybe Alex Friedman or one of the podcasts, and he

3:46

actually said something I thought was actually quite helpful. Let's see,

3:48

Ben, if you agree with it. He said something along the

3:50

lines of, you want to assume that

3:52

the big foundation models coming out of the big AI

3:54

companies are going to get a lot better. So you

3:56

want to assume they're going to get like a hundred

3:58

times better. As a startup founder,

4:00

you want to then think, okay, if

4:02

the current foundation models get 100 times better, is my

4:05

reaction, oh, that's great for me and for my startup

4:07

because I'm much better off as a result, or is

4:09

your reaction the opposite? Is it, oh shit, I'm in

4:11

real trouble. So let me just stop right there, Ben,

4:14

and see what you think of that as general advice.

4:17

Well, I think generally that's right,

4:19

but there's some nuances to it,

4:22

right? So I think that

4:24

from Sam's perspective, he was

4:26

probably discouraging people from building

4:28

foundation models, which I don't know

4:30

that I would entirely agree with

4:32

that, and that a lot

4:35

of the startups building foundation models are doing

4:37

very well. And there's many reasons for that.

4:39

One is there are architectural differences, which lead

4:41

to how smart is the model, there's

4:43

how fast is the model, there's how good is the

4:46

model in the domain. When

4:48

that goes for not just text

4:50

models, but image models as well,

4:52

there are different domains, different kinds

4:55

of images that responds to prompts

4:57

differently. If you ask Mid-Journey and

5:00

ideogram the same question, they

5:02

react very differently, depending on the

5:04

use cases that they're tuned for. And

5:07

then there's this whole field of

5:09

distillation where Sam can

5:11

go build the biggest, smartest model in the

5:13

world, and then you can walk up as

5:16

a startup and kind of do a distilled

5:18

version of it and get a model very,

5:20

very smart at a lot less cost. So

5:23

there are things that, yes,

5:25

the big company models are going to get way

5:28

better, kind of way better at what they are.

5:31

So you need to deal

5:33

with that. So if you're trying to go head

5:35

to head, full frontal assault, you probably have a

5:37

real problem just because they have so

5:40

much money. But

5:42

if you're doing something that's

5:45

different enough or a

5:47

different domain and so forth,

5:49

for example, at Databricks, they've

5:52

got a foundation model, but they're using

5:54

it in a very specific way in

5:56

conjunction with their leading.

6:00

data platforms. So, okay, now if

6:02

you're an enterprise and you need a

6:04

model that knows all the nuances of

6:06

how your enterprise data

6:10

model works and what things mean and needs

6:12

access control and what needs to use gear

6:15

specific data and domain knowledge and so forth,

6:17

then it doesn't really

6:19

hurt them if Sam's model gets way

6:21

better. Similarly, 11 Labs with

6:23

their voice model has kind of

6:26

embedded into everybody, everybody uses

6:29

it as part of kind of the AI

6:31

stack. And so it's got kind of a

6:33

developer hook into it. And then, you know,

6:36

they're going very, very fast to what they

6:38

do and really being very focused in their

6:40

area. So there are things that I

6:43

would say like extremely promising that

6:45

are kind of ostensibly, but

6:47

not really competing with open AI

6:49

or Google or Microsoft. So I

6:53

think it sounds a little more coarse grained than

6:55

I would interpret it if I was building a

6:57

startup. Right. Let's dig into this

6:59

a little bit more. So let's start with the question of do we

7:01

think the big models, the god models are going to get 100 times

7:03

better? I kind of

7:05

think so. And then I'm not sure. So if

7:07

you think about the language models, let's do those

7:10

because those are probably ones that people are most

7:12

familiar with. I think if

7:14

you look at the very top

7:16

models, you know, Claude and open

7:18

AI and Mistral and Llama, the

7:20

only people who I

7:23

feel like really can tell the difference as

7:25

users amongst those models are

7:27

the people who study them. You

7:29

know, like they're getting pretty close. So, you

7:32

know, you expect if we're talking 100x better,

7:34

that one of them might be separating from

7:36

each other a lot more. But

7:38

the improvement so 100% better

7:40

in what way? Like

7:43

for the normal priests and using it in

7:45

a normal way, like asking it questions and

7:47

finding out stuff. Let's say

7:49

some combination of just like breadth of

7:51

knowledge and capability. Yeah, like I

7:53

think in some of them they are. Yeah. Right.

7:55

But then also just combined with like sophistication of

7:57

the answers, you know, sophistication of the output.

8:00

the quality of the output, sophistication of the

8:02

output, lack of hallucination, factual grounding. Well,

8:05

that I think is for sure I'm gonna get a

8:07

hundred times better. Like that, yeah, I mean, they're

8:09

on a path for that. The things that are,

8:13

so against that, right? The

8:15

alignment problem where, okay, yeah,

8:18

they're getting smarter, but they're not allowed to

8:20

say what they know. And then that alignment

8:22

also kind of makes them dumber in other

8:24

ways. And so you do have that

8:26

thing. The other kind of question that's come up lately,

8:29

which is kind of, do we need

8:31

a breakthrough to go from what we

8:35

have now, which I would categorize as

8:38

artificial human intelligence as opposed

8:40

to artificial general intelligence, meaning

8:43

it's kind of the artificial version of us. We've

8:46

structured the world in a certain way using

8:49

our language and our ideas and our stuff.

8:52

And it's learned that very well,

8:54

amazing. And it can do kind

8:56

of a lot of the stuff that we can do,

9:00

but are we then the asymptote or

9:04

you need a breakthrough to get to

9:06

some kind of higher intelligence, more general

9:08

intelligence. And I think if we're

9:10

the asymptote, then in some

9:12

ways it won't get a hundred times

9:14

better because it's already like pretty good

9:16

relative to us. But

9:18

yeah, like it'll know more things. It'll

9:20

hallucinate less on all those dimensions. It'll

9:23

be a hundred times better. There's

9:25

this graph floating around. I forget exactly what the

9:27

axes are, but it's basically shows the improvement across

9:29

the different models. To your point, it shows an

9:31

asymptote against the current tests that people are using

9:33

that sort of like add or slightly above human

9:35

levels, which is what you would think if you're

9:37

being trained on entirely human data. Now

9:39

the counter argument on that is, are the tests just too simple, right?

9:41

It's a little bit like the question people have around the SAT, which

9:43

is if you have a lot of people getting 800s on

9:46

both math and verbal on the SAT, is

9:48

the scale too constrained? Do you need a test that

9:50

can actually test for Einstein? Right, right.

9:52

It's memorized the test that we have

9:55

and it's great. Right. But

9:57

you can imagine SAT that like really can detect

9:59

gradation. of people who have ultra high IQs,

10:01

who are ultra good at math or something, you

10:03

could imagine tests for AI. You could imagine tests,

10:05

the tests for reasoning above human levels, one assumes.

10:08

Yeah, well, maybe the AI needs to write the test. Yeah,

10:11

it needs to write the test. Yeah, and then there's

10:13

a related question that comes up a lot. It's an

10:15

argument we've been having internally, which is also I'll start

10:18

to take some more provocative and probably more bullish, or

10:20

as you would put it, science fictiony predictions on some

10:22

of this stuff. So there's this question that comes up,

10:24

which is, OK, you take an LM, you train it

10:26

on the internet. What is the internet data? What is

10:29

the internet data corpus? It's an average of everything, right?

10:31

It's a representation of human activity. Representation of human activity

10:33

is going to, because of the distribution of intelligence in

10:35

the population, most of it's somewhere in

10:37

the middle. And so the data set, on average, represents

10:39

the average human. You're teaching it to be very average,

10:41

yeah. Yeah, you're teaching it to be very average. It's just

10:44

because most of the content created on the internet is created

10:46

by average people. And so the content, on average, as

10:48

a whole, on average, is average. And

10:51

so therefore, the answer is our average, right? You're going to

10:53

get back an answer that represents the kind of thing that

10:55

on average 100 IQ, by definition, the

10:57

average human is 100 IQ. IQ is indexed

10:59

to 100. That's the center of the bell

11:01

curve. And so by definition, you're kind of getting back the

11:03

average. I actually argue that may be the case for the

11:06

default prompt today. If you just ask the thing, does the

11:08

Earth revolve around the sun or something, you get the average

11:10

answer to that. And maybe that's fine. This gets

11:12

to the point as, well, OK, the average data might be

11:14

of an average person. But the data

11:16

set also contains all of the things written and thought by

11:18

all the really smart people. All that stuff is in there.

11:20

And all the current people who are like that, their stuff

11:22

is in there. And so then it's sort of like a

11:25

prompting question, which is, how do you prompt it in order

11:27

to get basically, in order to basically navigate to a different

11:29

part of what they call the latent space, to navigate to

11:31

a different part of the data set that basically is like

11:33

the super genius part. And the way these things work is

11:35

if you craft the prompt in a different way, it actually

11:37

leads it down a different path inside the data set. It

11:40

gives you a different kind of answer. And here's another example

11:42

of this. If you ask it write code to do x,

11:44

write code to sort of list, or whatever, render an

11:46

image, it will give you average code to do that.

11:48

If you say, write me secure code to do that,

11:50

it will actually write better code with

11:53

fewer security holes, which is very interesting. Because it's

11:55

accessing a different purpose of training data, which is

11:57

secure code. And if you ask, write this image

11:59

in a row, generation thing the way John Carmack would write it,

12:01

you get a much better result because it's tapping into the

12:03

part of the latent space represented by John Carmack's code, who's

12:05

the best graphics programmer in the world. And

12:08

so you can imagine prompting crafts in many

12:10

different domains such that you're kind of unlocking

12:12

the latent supergenius, even if that's

12:14

not the default answer. Yeah, no, so

12:16

I think that's correct. I think

12:18

there's still a potential

12:21

limit to its smartness. So

12:24

we had this conversation in the firm the other day

12:26

where you have, there's the world, which

12:28

is very complex. And intelligence

12:30

kind of is, you

12:32

know, how well can you understand, describe,

12:35

represent the world? But our

12:37

current iteration of artificial

12:39

intelligence consists of

12:42

humans structuring the world and

12:45

then feeding that structure that we've come

12:47

up with into the AI. And

12:50

so the AI kind of is

12:53

good at predicting how humans have structured the

12:55

world as opposed to how the world actually

12:57

is, which is, you know, something

12:59

more probably complicated, maybe

13:02

the irreducible or what have you.

13:05

So do we just get to a limit

13:07

where like it can be really smart, but

13:09

its limit is going to be the smartest

13:11

humans as opposed to smarter than the smartest

13:13

humans and then kind of

13:15

related, is it going

13:18

to be able to figure out brand new things,

13:20

you know, new laws of physics and so forth.

13:22

Now, of course, there are like

13:24

one in 3 billion

13:26

humans that can do that or whatever. That's

13:28

a very rare kind of intelligence. So it

13:31

still makes the AI is extremely

13:33

useful, but they play

13:35

a different role if they're kind

13:38

of artificial humans than if they're

13:40

like artificial, you know, super

13:42

duper mega humans. So

13:46

let me make the sort of extreme bull case for the

13:48

hundred because okay. So the cynic would

13:50

say the Sam Altman would be saying they're going to get

13:52

a hundred times better precisely if they're not going to. Yeah.

13:56

Yeah. Yeah. Yeah. Right. Because he'd be saying that basically in

13:58

order to scare people in a not competing. Well,

14:01

I think that whether or not they are

14:03

going to get 100 times better, Sam would

14:05

be very likely to say that. Like Sam,

14:07

for those of you who don't know him,

14:10

he's a very smart guy, but for sure

14:12

he's a competitive genius. There's no question about

14:14

that. So you have to take that pick

14:16

up. Right. So if

14:18

they weren't going to get a lot better, he would say that.

14:20

But of course, if they were going to get a lot better,

14:22

to your point, he would also say that. He would also, why

14:24

not? And so let me make the

14:26

bull case that they are going to get 100 times better or

14:28

maybe even on an upper curve for a long time. And

14:31

there's enormous controversy, I think, on every one of the things I'm

14:33

about to say. But you can find very smart people in the

14:36

space who believe basically everything I'm about

14:38

to say. So one is there is

14:40

generalized learning happening inside the neural networks. And we know

14:42

that because we now have introspection techniques where you can

14:44

actually go inside and look inside the neural networks to

14:46

look at the neural circuitry that is being evolved as

14:49

part of the training process. And

14:51

these things are evolving, general computation functions. There was

14:53

a case recently where somebody trained one of these

14:55

on a chess database. And just by training a

14:57

lot of chess games, it actually imputed a world

14:59

model of a chessboard inside

15:02

the neural network that was able to do original moves.

15:04

And so the neural network training process does seem

15:07

to work. And then specifically, not only that, but

15:09

META and others recently have been talking about how

15:11

so-called overtraining actually works, which is

15:13

basically continuing to train the same model against

15:15

the same data for longer, putting more and

15:17

more compute cycles against it. I've

15:20

talked to some very smart people in the fields,

15:22

including there, who basically think that actually that works

15:24

quite well. The diminishing returns people were worried about

15:26

about more training. And they proved it in a

15:28

new LAMAR release, right? That's the primary

15:30

technique they use. Yeah, exactly. Like one guy in the

15:32

space basically told me, basically, he's like, yeah, we don't

15:34

necessarily need more data at this point to make these

15:36

things better. We maybe just need more compute cycles. We

15:38

just trained it 100 times more and it may just

15:40

get actually a lot better. So on

15:42

data labeling, it turns out that supervised

15:44

learning ends up being a huge boost

15:46

to these things. Yeah. So

15:49

we've got that. We've got all of the kind of,

15:51

you know, let's say rumors and reports of various kinds

15:54

of self-improvement loops, you know, that kind of

15:56

underway. And most of the sort of super advanced practitioners in

15:58

the field think that there's now some form self-improvement loop

16:00

that works, which basically is you basically

16:02

get an AI to do what's called chain of thoughts. You get

16:04

it to basically go step by step to solve a problem. You

16:06

get it to the point where it knows how to do that.

16:09

And then you basically retrain the AI on the answers. And so

16:11

you're kind of basically doing a sort of a forklift upgrade across

16:13

cycles of the reasoning capability. And so a lot of

16:15

the experts think that sort of thing's starting to work

16:17

now. And then there's still a raging

16:19

debate about synthetic data, but there's quite a few people who

16:22

are actually quite bullish on that. Yeah.

16:24

And then there's even this trade-off, there's this kind of

16:26

dynamic where like LLMs might be okay at writing code,

16:28

but they might be really good at validating code. You

16:31

know, they might actually be better at validating code than they are at writing

16:33

it. That would be a big help. Yeah. Well,

16:35

but that also means like AI is maybe able to

16:37

self-improvement. They can validate their own code. They can

16:39

validate their own code. And we have this anthropomorphic bias that's

16:41

very deceptive with these things, because you think of the model

16:43

as an it. And so it's like, how

16:45

could you have an it that's better at validating code than

16:47

writing code? But it's not an it, what it is, is

16:50

it's this giant latent space. It's this giant neural network. And

16:52

the theory would be there are totally different parts of the

16:54

neural network for writing code and validating code. And

16:56

there's no consistency requirement whatsoever that the network be equally

16:58

good at both of those things. And so if it's

17:00

better at one of those things, right? So

17:02

then the thing that it's good at might be able to make the

17:04

thing that it's bad at better and better. Right,

17:07

right, right, right, right. Sure, sure. Right, sort

17:09

of a self-improvement thing. And so then on top

17:11

of that, there's all the other things coming, right?

17:13

Which is it's everything is all these practical things,

17:16

which is there's an enormous chip constraint right now.

17:18

So every AI that anybody uses today is its

17:20

capabilities are basically being gated by the availability of

17:22

chips, but like that will resolve over time. You

17:25

know, there's also your point I like data labeling. There is a lot

17:27

of data in these things now, but there is a lot more data

17:29

out in the world. And there's, you know, at least in theory, some

17:31

of the leading AI companies are actually paying to generate new data. And

17:34

by the way, even like the open source data sets are getting much

17:36

better. And so there's a lot of like data

17:38

improvements that are coming and then, you know, there's just the amount of

17:40

money pouring into the space to be able to underwrite all this. And

17:42

then by the way, there's also just the systems engineering work that's happening,

17:45

right? Which is a lot of the current systems, you

17:47

know, where basically we're built by scientists and now they're really

17:49

world-class engineers are showing up and tuning them up and getting

17:51

them to work better. And, you know, maybe that's not a... Which

17:55

makes training, by the way, way more

17:57

efficient as well. Not just inference, but...

17:59

but also training. Yeah, exactly.

18:01

And then even, you know, another improvement area

18:04

is basically Microsoft released their five small language

18:06

model yesterday and apparently it's competitive. It's

18:08

a very small model, competitive with much larger models. And

18:10

the big thing they say that they did was they

18:12

basically optimized the training set. So they basically

18:14

deduplicated the training set. They took out all the copies and

18:16

they really optimized on a small amount of training data, on

18:19

a small amount of high quality training data as opposed to

18:21

the larger amounts of low quality data that most people train

18:23

on. You add all these up and you've got eight or

18:25

10 different combination of sort of

18:27

practical and theoretical improvement vectors that

18:29

are all in play. And it's hard for me to

18:32

imagine that some combination of those doesn't lead to like

18:34

really dramatic improvement from here. I definitely

18:36

agree. I think that's for sure going to happen.

18:38

Right? Like if you were, so back to Sam's

18:40

proposition, I think if you were a startup and

18:43

you were like, okay, in two years I

18:45

can get as good as GPT-4, you

18:47

shouldn't do that. Right. With

18:49

a hit, that would be a bad mistake. Right,

18:52

right. Well, this also goes to, you know, a lot of

18:54

entrepreneurs are afraid of, I'll give you an example.

18:56

So a lot of entrepreneurs, here's this thing they're trying to figure

18:58

out, which is okay. I really think, I know how to build

19:00

a SaaS app that harnesses an LLM to do really good marketing

19:02

collateral. Let's just make it very similar, a very, very simple thing.

19:05

And so I build a whole system for that.

19:07

Will it just turn out to be that the

19:09

big models in six months will be even better

19:11

in making marketing collateral just from a simple prompt

19:14

such that my apparently sophisticated system is just

19:16

irrelevant because the big model just does it. Yeah.

19:19

Let's talk about that. Like apps, you know, there

19:21

are, another way you can think about it is

19:23

that the criticism of a lot of current AI

19:25

app companies is their quote unquote, you know, GPT

19:27

wrappers. There's sort of thin layers of

19:29

wrapper around the core model, which means the core model

19:31

could commoditize them or displace them. But the counter argument

19:33

of course is, it's a little bit like calling all,

19:36

you know, old software apps, you know, database wrappers, you

19:38

know, wrappers around a database. It turns

19:40

out like actually wrappers around a database is like most modern software.

19:42

And a lot of that actually turned out to be really valuable.

19:44

And there, it turns out there's a lot of things to build

19:46

around the core engine. So yeah. So Ben, how do we think

19:48

about that when we run into companies thinking about building apps? Yeah.

19:52

You know, it's a very tricky question

19:54

because there's also this correctness gap, right?

19:56

So, you know, why do we

19:58

have co-pilots? Where are

20:01

the pilots? Where are the,

20:03

there's no AI pilots, there are only

20:05

AI co-pilots. There's a human in the

20:07

loop on absolutely everything. And

20:09

that really kind of comes down to this,

20:12

you know, you can't

20:14

trust the AI to be correct

20:16

in drawing a picture

20:18

or writing a program or, you

20:20

know, even like writing

20:23

a court brief without making up citations.

20:27

You know, all these things kind of require a

20:29

human and kind of turns

20:31

out to be like fairly dangerous to not.

20:33

And then I think that, so what's happening

20:35

a lot with the application layer is people

20:37

saying, well, to make it really

20:40

useful, I need to turn this co-pilot into a

20:42

pilot. And can I do that?

20:44

And so that's an

20:46

interesting and hard problem. And then there's

20:48

a question of, is that better done

20:50

at the model level or at some

20:53

layer on top that, you know, kind of

20:55

teases the correct answer out of the model,

20:57

you know, by doing things like using code

20:59

validation or what have you, or

21:02

is that just something that the models will be able to

21:04

do? I think that's one open question. And

21:06

then, you know, as you get into

21:08

kind of domains and, you know, potentially

21:11

wrappers on things, I think there's a

21:13

different dimension than what the models are

21:15

good at, which is what is the

21:18

process flow, which

21:20

is kind of in database for all to, so

21:22

on the database kind of analogy,

21:25

there is like the

21:27

part of the task in

21:30

a law firm that's writing the brief, but there's 50

21:32

other tasks and things that have to

21:34

be integrated into the way a company

21:38

works, like the process flow, the

21:40

orchestration of it. And maybe

21:42

there are, you know, a lot

21:44

of these things, like if you're doing video

21:46

production, there's many tools or music even, right?

21:49

Like, okay, who's going to write the

21:51

lyrics? Which AI will write the lyrics and

21:53

which AI will figure out the music? And

21:55

then like, how does that all come together

21:57

and how do we integrate it and so

21:59

forth? Those things tend

22:02

to just require a

22:04

real understanding of the end customer

22:07

and so forth in a way.

22:10

That's typically been how applications have been

22:12

different than platforms in the past. Like

22:15

there's real knowledge about how the

22:17

customer using it wants to function.

22:20

That doesn't have anything to do with the Intelli

22:23

or is just

22:26

different than what the platform is designed

22:28

to do. To get

22:30

that out of the platform for a

22:33

company or a person turns out to be really hard. Those

22:36

things I think are likely to

22:38

work, especially if the

22:41

process is very complex. It's

22:43

something that's funny as a firm, we're

22:45

a little more hardcore technology oriented, and

22:48

we've always struggled with those in terms

22:50

of, this is

22:53

some process application for plumbers

22:56

to figure out this. We're like, well, where's

23:00

the technology? But a

23:02

lot of it is how do you encode

23:05

some level of domain expertise and how

23:07

things work in the actual

23:09

world back into the software? I

23:13

often think, I haven't told founders that you can think about

23:15

this in terms of price. You can work backwards for pricing

23:17

a little bit, which is to say business value and what

23:19

you can charge for. The

23:22

natural thing for any technologist to do is to say, I

23:24

have this new technological capability and I'm going to sell it

23:26

to people. What am I going to charge for it? It's

23:28

going to be somewhere between my cost of providing

23:30

it and then whatever markup, I think I

23:32

can justify. If I have a monopoly

23:34

providing it, maybe the markup is infinite. But it's

23:37

our technology forward, supply

23:40

forward pricing model. There's

23:43

a completely different pricing model for business

23:45

value backwards, so-called value-based

23:47

pricing. To

23:51

your point, that's basically a pricing model that says, okay,

23:53

what's the business value to the customer of the thing?

23:56

If the business value is a million dollars, then

23:59

can I- I charge 10% of that and get

24:01

$100,000, right? Or whatever.

24:03

And then, you know, why is it cost $100,000

24:06

as compared to $5,000 is because, well,

24:08

because to the customer it's worth a million dollars and so they'll

24:10

pay 10% for it. Yeah,

24:13

actually, so a great example of that, like

24:15

we've got a company

24:17

in our portfolio, Crest AI that

24:19

does things like debt

24:23

collection. Okay,

24:25

so if I can collect way

24:27

more debt with way fewer

24:29

people with my, you know,

24:31

it's a co-pilot type solution, then

24:36

what's that worth? Well, it's

24:38

worth a heck of a lot more than just

24:40

buying an open AI license because

24:43

an open AI license is not

24:45

gonna easily collect debts or

24:48

kind of enable your debt collectors to be

24:51

massively more efficient or that kind of thing.

24:53

So it's bridging that gap

24:55

between the value. And I think you

24:57

had a really important point. The test for whether your idea

24:59

is good is how much can you charge for it? Can

25:02

you charge the value? Or

25:05

are you just charging the amount of work

25:07

it's gonna take the customer to put

25:10

their own wrapper on top of open

25:12

AI? Like that's the

25:14

real test to me of like how deep

25:17

and how important is what you've done. Yeah,

25:20

and so to your point, I like the kinds of

25:22

businesses that technology investors have had

25:24

a hard time with, kind of thinking

25:26

about maybe accurately is sort of it's the company

25:28

that is, it's a vendor that has built something

25:30

where it is a specific solution to

25:33

a business problem where it turns out the business problem

25:35

is very valuable to the customer. And so therefore they

25:37

will pay a percentage of

25:39

the value provided back in

25:42

the terms, if for price for the

25:44

software. And that actually turns out you

25:47

can have businesses that are not very technologically

25:49

differentiated that are actually extremely lucrative. And

25:52

then because that business is so lucrative,

25:55

they can actually afford to go think very deeply about how

25:57

technology integrates into the business, what else they can do. This

26:00

is like the story of a salesforce.com, for example. By

26:03

the way, there's a chance, a

26:07

theory that the models are all getting

26:09

really good. There are open source

26:11

models that are awesome. Llama,

26:15

Mistral, these are great models.

26:19

The actual layer where the value

26:21

is going to accrue is going

26:23

to be tools, orchestration, that thing

26:26

because you can just plug in whatever the best

26:28

model is at the time. Whereas the

26:30

models are going to be competing in a

26:32

death battle with each other and be

26:36

commoditized down to the cheapest one

26:38

wins and that kind of thing.

26:41

You could argue that

26:43

the best thing to do is to connect

26:47

the power to the people. Right.

26:51

That actually takes us to the next question, and

26:53

this is a two-in-one question. Michael asks, and

26:55

I'll say these are diametrically opposed, which is why

26:58

I paired them. Michael asks, why

27:00

are VCs making huge investments in generative AI

27:02

startups? When it's clear these startups won't be

27:04

profitable anytime soon, which is a loaded question,

27:06

but we'll take it. Then Kaiser

27:08

asks, if AI deflates the cost of building

27:10

a startup, how will the structure of tech

27:13

investment change? Of course, Ben,

27:15

this goes to exactly what you just said. It's basically

27:17

the questions are diametrically opposed because if you squint out

27:19

of your left eye, what you

27:21

see is basically the amount of money being invested in

27:23

the foundation model companies going up to the right at

27:25

a furious pace. These companies are raising hundreds of millions,

27:28

tens of billions of dollars. It's just like,

27:30

oh my God, look at these capital

27:32

infernos that hopefully will result in value at the end

27:35

of the process. But my God, look at how much

27:37

money is being invested in these things. If

27:40

you squint through your right eye, you think, wow,

27:42

now all of a sudden, it's much easier to

27:44

build software. It's much easier to have a software

27:46

company. It's much easier to have a small number

27:48

of programmers writing complex software because they've got all

27:50

these AI co-pilots and all these automated software

27:52

development capabilities that are coming online. So

27:55

on the other side, the cost of

27:57

building an AI like application startup might

27:59

crash. it might just be that like

28:01

the AI salesforce.com might

28:04

cost a 10th or 100th or a thousandth

28:06

amount of money that it took to build

28:08

the old database driven salesforce.com. And

28:10

so yeah, so what do we think of that dichotomy, which is

28:12

you can actually look out of either eye and

28:15

see either cost to the moon as

28:17

like for startup funding or cost actually going to

28:20

zero. Yeah, well, like,

28:22

so it is interesting. I mean, we

28:24

actually have companies in both camps, right?

28:26

Like I think probably

28:29

the companies that have gotten

28:31

to profitability the fastest, maybe

28:33

in the history of the firm have been AI

28:35

companies or been AI companies in the portfolio where

28:38

the revenue grows so fast that

28:40

it actually kind of runs out

28:42

ahead of the cost. And

28:44

then there are like, people

28:46

who are in the foundation model race who

28:49

are raising hundreds of millions,

28:51

even billions of dollars to

28:53

kind of keep pace and so forth.

28:56

They also are kind of

28:58

generating revenue at a fast rate. The

29:00

head count in all of them is small. So

29:02

I would say, where AI

29:05

money goes, and

29:08

even like if you look at OpenAI,

29:10

which is the big spender in startup

29:12

world, which we

29:15

are also investors in is, head

29:18

count wise, they're pretty small against their revenue.

29:20

Like it is not a big company head

29:22

count. Like if you look at the revenue

29:24

level and how fast they've gotten there, it's

29:27

pretty small. Now the total

29:29

expenses are ginormous, but

29:31

they're going into the model creation. So it's

29:33

an interesting thing. I mean, I'm not entirely

29:37

sure how to think

29:39

about it, but I think like if you're

29:41

not building a foundation model, it will make

29:44

you more efficient and probably gets profitability quicker.

29:47

Right. So the counter,

29:50

and this is a very bullish counter argument, but

29:52

the counter argument to that would be basically falling

29:54

costs for like building new software companies are a

29:56

mirage. And the reason for that is this thing

29:58

in economics called the Jevons Paradox. which

30:00

I'm going to read from Wikipedia. So

30:02

the Jevons paradox occurs when technological progress

30:05

increases the efficiency with which a resource is

30:07

used, right, reducing the amount of

30:10

that resource necessary for any one use, but

30:12

the falling cost induces increases

30:14

in demand, right, elasticity, enough

30:17

that the resource use overall is increased

30:19

rather than reduced. Yeah,

30:21

it's certainly possible. Right,

30:23

and so this is, you see versions of this,

30:25

for example, you build in your freeway and it

30:27

actually makes traffic jams worse, right, because

30:30

basically what happens is, oh, it's great, now there's more roads,

30:32

now we can have more people live here, we can have

30:34

more people, we can make these companies bigger and now there's

30:36

more traffic than ever, and now the traffic's even worse. Or

30:39

you saw the classic example is during the

30:41

industrial revolution coal consumption, as the

30:43

price of coal drops, people

30:45

use so much more coal that actually

30:47

the overall consumption actually increased. But

30:49

people are getting a lot more power, but the result was the

30:52

use of a lot more coal in the

30:54

paradox. And so the paradox here would be, yes, the

30:57

cost of developing any given piece of software

30:59

falls, but the reaction to that is a

31:01

massive surge of demand for software capabilities. And

31:03

so the result of that actually is, although

31:05

it looks like starting software companies, the price

31:07

is gonna fall, actually it's gonna happen, it's

31:09

gonna rise, for the high quality reason that

31:11

you're gonna be able to do so much

31:13

more, right, with software, the

31:15

products are gonna be so much better and the roadmap is

31:17

gonna be so amazing of the things you can do and

31:19

the customers are gonna be so happy with it that they're

31:21

gonna want more and more and more. So

31:24

the results of it, and by the way, another example

31:26

of Jeff and Spheradox playing out in other related industries

31:28

in Hollywood, CGI in theory

31:31

should have reduced the price of making movies

31:33

in reality has increased it because audience expectations

31:35

went up. And now you go

31:37

to a Hollywood movie and it's wall-to-wall CGI. And

31:40

so movies are more expensive to make than ever.

31:42

And so the result of it, but the result

31:44

of Hollywood is at least much more, let's say

31:46

visually elaborate movies, whether they're better or not is

31:48

another question, but much more visually elaborate, compelling, kind

31:51

of visually stunning movies through CGI. The version here

31:53

would be much better software, but

31:55

it's like radically better software to the end user, which causes end

31:57

users to want a lot more software, which causes a lot more.

32:00

is actually the price of development to rise. If

32:02

you just think about a simple case

32:04

like travel, okay, booking

32:07

a trip through Expedia is complicated. You're likely

32:09

to get it wrong. You're clicking on menus

32:11

and this and that and the other. An

32:16

AI version of that would be like, send

32:18

me to Paris, put me in a hotel I love at

32:20

the best price, send

32:22

me on the best possible

32:24

kind of airline, an airline

32:26

ticket. Then make

32:29

it really special for me. And maybe

32:31

you need a human to go, okay,

32:37

or maybe the AI gets more complicated and says,

32:39

okay, well, we know the person loves chocolate and

32:42

we're gonna like FedEx in the

32:44

best chocolate in the world from Switzerland into this

32:46

hotel in Paris and this and that and the

32:49

other. And so the

32:51

quality could get to levels

32:55

that we can't even imagine today,

32:57

just cause the software tools aren't what

32:59

they're gonna be. That's

33:02

right. Yeah, I kind of buy

33:04

that actually. I think I brought in your argument.

33:08

How about I'm gonna land in whatever Boston at six

33:10

o'clock and I wanna have dinner at seven with a

33:13

table full of like super interesting people. Yeah,

33:15

right, right, right, right. You

33:17

know, right? Like,

33:21

no travel agent would do that for you today nor would you welcome to.

33:24

No, no. Right,

33:28

well, then you think about it, it's

33:30

gotta be integrated into my personal AI

33:32

and like, and this, you know, there's

33:34

just like unlimited kind of ideas

33:37

that you can do. And I think this is one

33:39

of the kind of things that's always been underestimated

33:42

about humans is like

33:45

our ability to come up with new things we

33:47

need. Like that has been

33:49

unlimited. And there's a very kind of famous

33:52

case where John Maynard Keynes, who

33:54

the kind of prominent economist in

33:56

the kind of first half of

33:58

last century. had this

34:01

thing that he predicted, which

34:03

is like, because of automation,

34:05

nobody would ever work a

34:07

40-hour workweek, because once their

34:10

needs were met, needs

34:12

being shelter and food. And

34:14

I don't even know if transportation was in

34:17

there. That was it. It was

34:19

over. And you would never work past the

34:21

need for shelter and food. Why would you?

34:24

There's no reason to. But of

34:26

course, needs expanded. So then everybody needed

34:28

a refrigerator. Everybody needed not just

34:30

one car, but a car for everybody in the

34:33

family. Everybody needed a television set. Everybody

34:35

needed glorious vacations. Everybody,

34:37

you know. So what

34:40

are we going to need next? I'm quite

34:42

sure that I can't imagine it, but like

34:44

somebody's going to imagine it. And it's quickly

34:46

going to become a need. Yeah,

34:48

that's right. By the way, as Kane famously

34:51

said, his essay, I think, was economic prospects

34:53

for our grandchildren, which was basically that. Yeah.

34:55

And what you just articulated. So Karl Marx

34:57

had another version of that. I just pulled

34:59

up the quote. So when

35:02

the Marxist utopia socialism is achieved,

35:05

society regulates the general production. That

35:07

makes it possible for me to do, to

35:09

hunt in the morning, fish

35:13

in the afternoon, rear cattle in the

35:15

evening, criticize after dinner. What

35:18

a glorious life. What a glorious

35:21

life. Like if I could just list four things that I

35:23

do not want to do, it's

35:25

hunt, fish, rear cattle

35:27

and criticize. Right. And

35:29

by the way, it says a lot about Marx. Those were

35:31

his four things. Well, criticizing being his

35:34

favorite thing, I think, basically

35:37

communism in a nutshell. Yeah,

35:39

exactly. I don't want to get too political,

35:41

but yes. Yes. A hundred

35:43

percent. And so yeah, so it's just this, yeah, what they

35:45

have with Kane's and Marx that in common is just this

35:47

incredibly constricted view of what people want to do. And then

35:49

correspondingly, you know, the other thing is just like, you know,

35:51

people, people who want, people who want to have a mission.

35:53

I mean, probably some people just want to fish and hunt.

35:55

Yeah. But you know, a lot of people, a lot of

35:58

people want to have a mission. They want to have a.

36:00

They want to have a purpose. They want to be useful. They want

36:02

to be productive. It's actually a good thing in life, it turns out. It

36:05

turns out in

36:08

the startling turn of events. OK, so yeah,

36:10

I think that I've long felt a little bit of

36:12

the software eats the world thing a

36:14

decade ago. I've always thought that basically demand

36:17

for software is sort of perfectly elastic, possibly

36:19

to infinity. And the theory there basically is if

36:21

you just continuously bring down the cost of software,

36:24

which has been happening over time, then basically demand

36:27

basically perfectly correlates upward. And the reason

36:29

is because, as we've been discussing, but

36:32

there's always something else to do in software.

36:34

There's always something else to automate. There's always

36:36

something else to optimize. There's always something else

36:38

to improve. There's always something to make better.

36:41

And in the moment with the constraints that you have today,

36:43

you may not think of what that is. But the minute

36:45

you don't have those constraints, you'll imagine what it is. I'll

36:48

just give you an example. I'll give you an example of playing

36:50

out with AI right now. So there have been, and we have

36:52

companies that do this, there

36:55

have been companies that have made software systems

36:57

for doing security cameras forever. And for a

36:59

long time, it was a big deal to

37:01

have software that would do, have

37:04

different security camera feeds and store them on a DVR and be

37:06

able to replay them and have an interface that lets you do

37:08

that. Well, it's like AI security

37:10

cameras all of a sudden can have actual semantic

37:12

knowledge of what's happening in the environment. And so

37:15

they can say, hey, that's Ben. And then they

37:17

can say, oh, hey, that's Ben, but he's carrying

37:19

a gun. Yeah. Right? And by

37:21

the way, that's Ben and he's carrying a gun,

37:23

but that's because he hunts on Thursdays and Fridays

37:26

as compared to that's Mary, and she never carries a

37:28

gun and like, you know, like something is wrong. And

37:31

she's really mad, right? She's got a really esteemed expression

37:33

on her face and we should probably be worried about

37:36

it, right? And so there's like an entirely new set

37:38

of capabilities you can do just as one example for

37:40

security systems that were never possible pre AI. And

37:43

a security system that actually has a semantic understanding

37:45

of the world is obviously much more sophisticated than

37:47

the one that doesn't and might actually be more expensive

37:49

to make, right? Right. Well, and

37:51

just imagine healthcare, right? Like you could

37:55

wake up every morning and

37:57

have a complete diagnostic. you

38:00

know, like how am I doing today?

38:02

Like what are all my levels of everything? And,

38:05

you know, how should I interpret them? You know, better

38:07

than, you know, this is one thing

38:09

where AI is really good is, you

38:11

know, medical diagnosis because it's a

38:13

super high dimensional problem, but if

38:15

you can get access to, you

38:18

know, your continuous glucose reading, you

38:20

know, maybe sequence your blood now and again,

38:22

this and that and the other, then

38:25

you've got an incredible kind of view of

38:27

things. And who doesn't want

38:29

to be healthier? You know,

38:31

like now we have a scale. That's

38:34

basically what we do. You know,

38:36

maybe check your heart rate or something, but

38:38

like pretty primitive stuff compared to where we

38:40

could go. Yeah, that's right. Okay, good.

38:43

All right, so let's go to the next topic.

38:45

So on the topic of data, so a major

38:48

Tom asks, as these AI models allow for us

38:50

to copy existing app functionality at minimal cost, proprietary

38:53

data seems to be the most important moat. How

38:56

do you think that will affect proprietary data

38:58

value? What other moats do you think companies

39:00

can focus on building in this new environment?

39:02

And then Jeff Weisshopp asks, how should companies

39:05

protect sensitive data, trade secrets, proprietary data, individual

39:07

privacy in the brave new

39:09

world of AI? So let me start with a

39:11

provocative statement, Ben, see if you

39:13

agree with it, which is, you know, you sort of

39:15

hear a lot, this sort of statement or cliche is like data is

39:17

the new oil. And so it's

39:20

like, okay, data is the key input to training

39:22

AI, making all this stuff work. And so, you

39:24

know, therefore, you know, data is basically the new

39:26

resource, it's the limiting resource, it's the super valuable

39:28

thing. And so, you know, whoever has the best

39:30

data is going to win. And you see that directly in how you

39:32

train AIs. And then, you know, you also have like a lot of

39:34

companies, of course, that are now trying to figure out what to do

39:37

with AI. And a very common thing

39:39

you'll hear from companies is, well, we have proprietary data,

39:41

right? So I'm a, you know, I'm a hospital chain,

39:43

or I'm a, you know, whatever, any kind

39:45

of business, insurance company or whatever, and I've got all this

39:47

proprietary data that I can apply, you know,

39:49

that I'll be able to, you know, build things with my

39:51

proprietary data with AI that won't just, you know,

39:54

be something that anybody will be able to have. Let

39:57

me argue that basically, let's see, And

40:00

like almost every case like that, it's not true. It's

40:02

basically what the internet kids would call cope. It's

40:05

simply not true. And the reason it's just

40:07

not true is because the amount of data

40:09

available on the internet and just generally in

40:11

the environment is just a million

40:13

times greater. And

40:16

so while it may not, you know, while it

40:18

may not be true that I have your specific

40:20

medical information, I have so much medical information off

40:22

the internet for so many people

40:24

in so many different scenarios that it

40:26

just swamps the value of quote

40:28

your data. You know, just it's just

40:31

it's just like overwhelming. And so your proprietary data as

40:33

you know, company X will be a little bit useful

40:35

on the margin, but it's not actually going to move

40:37

the needle and it's not really going

40:39

to be a barrier to entry in most cases.

40:41

And then let me cite as proof for the

40:43

for my belief that this is mostly cope is

40:46

there has never been nor is there now any

40:48

sort of basically any level of sort of richer

40:50

sophisticated marketplace for data, market for data. There's no

40:52

there's no there's no large marketplace

40:54

for data. And in fact,

40:56

in fact, what there are is there are very small markets

40:58

for data. So there are these businesses called data brokers that

41:00

will sell you, you know, large numbers of like, you know,

41:03

information about users on the internet or something. And

41:05

they're just small businesses like they're just not large.

41:07

It just turns out like information on lots of people

41:09

is just not very valuable. And

41:11

so if the data actually had value, you know,

41:14

it would have a market price and you would

41:16

see it transacting and you actually very specifically don't

41:18

see that, which is sort of a, you

41:20

know, yeah, sort of quantitative proof that the data actually is

41:22

not nearly as valuable as people think it is. Where

41:25

I agree, so I agree

41:28

that the data, like just as

41:30

here's a bunch of data,

41:33

and I can sell it without

41:35

doing anything to the data is

41:38

like massively overrated. Like I definitely

41:40

agree with that. And like maybe

41:43

I can imagine some exceptions like

41:45

some, you

41:47

know, special population genomic databases or something that

41:49

are, that were very hard to acquire that

41:52

are useful in some way that, you know,

41:54

that's not just like living on the internet

41:56

or something like that. I could imagine where

41:59

that's super. highly structured, very

42:01

general purpose, and not widely available.

42:04

But for most data in companies, it's

42:06

not like that. And that it tends

42:08

to not, it's either widely available or

42:10

not general purpose. It's kind of specific.

42:13

Having said that right, like companies have

42:16

made great use

42:18

of data. For example, a company that

42:20

you're familiar with, Meta, uses

42:22

its data to kind of great

42:24

ends itself, feeding it into its

42:27

own AI systems, optimizing its products

42:29

in incredible ways. And I think

42:31

that us, Andres and

42:33

Horowitz, actually, we just raised $7.2

42:36

billion. And

42:38

it's not a huge deal, but

42:41

we took our data and

42:43

we put it into an AI

42:45

system. And our LPs were able,

42:47

there's a million questions investors have

42:49

about everything we've done, our

42:52

track record, every company we've invested and

42:54

so forth. And for any of those

42:56

questions, they could just ask the AI. They could wake

42:58

up at 3 o'clock in the morning and go, do I really want

43:00

to trust these guys? And go in and

43:02

ask the AI a question, and boom, they'd get an

43:04

answer back instantly. They'd have to wait for us and

43:06

so forth. So we really kind of

43:08

improved our investor relations product

43:11

tremendously through use of our data.

43:14

And I think that almost

43:16

every company can improve

43:18

its competitiveness through use

43:21

of its own data. But the

43:23

idea that it's collected some data that

43:25

it can go like cell, or

43:28

that is oil, or what

43:30

have you. Yeah,

43:33

that's probably not true,

43:35

I would say. And it's kind

43:37

of interesting because a lot of

43:40

the data that you would think would

43:42

be the most valuable would be like your own

43:44

code base, right? Your software that

43:46

you've written. So much of that

43:49

lives in GitHub. Nobody is actually, I

43:52

don't know of any company, we work with whatever,

43:55

a thousand software companies. And

43:57

do we know any that's like building their own programming

43:59

model? on their own code? Or,

44:04

and would that be a good idea? Probably

44:06

not just because there's so much code out

44:08

there that the systems have been trained on.

44:10

So like that's not

44:13

so much of an advantage. So I think it's a very

44:15

specific kind of data that would have value. Well,

44:18

let's make it actionable then. If I'm

44:20

running a big company, like if I'm running an insurance company

44:22

or a bank or a hospital chain or something like that,

44:24

like how to, or, you know, I ended up in a

44:27

consumer packaged goods company, Pepsi or something.

44:30

How should I validate? How should I

44:32

validate that I actually have a valuable proprietary data asset

44:34

that I should really be focusing on using versus maybe,

44:36

versus in the alternate, by the way, maybe there's other

44:38

things like, maybe I should be taking all the effort

44:41

I would spend on trying to optimize use of that

44:43

data. And maybe I should use it entirely trying to

44:45

build things using internet data instead. Yeah,

44:49

so I think, I mean, look, if

44:51

you're, right, if you're in the insurance

44:54

business, then like all your actuarial data

44:56

is both interesting. And then I don't

44:58

know that anybody publishes their actuarial data.

45:03

And so like, I'm not sure how you would

45:05

train the model on stuff off of the internet.

45:08

Yeah, similarly. That's a good, let me, can I challenge that

45:10

one? So that would be a good thing. That'd

45:12

be a good test case. So I'm an insurance company. I've got

45:14

records on 10 million people and you know, the actuarial tables and

45:16

when they get sick and when they die. Okay,

45:19

that's great. But like there's lots

45:21

and lots of actuarial, general actuarial data on

45:23

the internet for large scale populations, you know,

45:25

because governments collect the data and they process

45:27

it and they publish reports. And

45:30

there's lots of academic studies. And so

45:32

like, is your large data

45:34

set giving

45:36

you any additional actuarial information that

45:38

the much larger data set on

45:40

the internet isn't already providing you?

45:43

Like are your insurance clients actuarially

45:45

any different than just everybody? I

45:48

think so, because on intake, on

45:50

the, you know, when you get

45:52

insurance, they give you like a blood test, they

45:54

got all these things. I know if you're a

45:56

smoker and so forth. And in the,

45:58

I think in the general data. is that like, yeah, you

46:00

know who dies, but you don't know what the fuck

46:03

they did coming in. And

46:05

so what you really are looking for is like,

46:07

okay, for this profile of person with this kind,

46:09

with these kinds of lab results, how

46:11

long do they live? And that's, that's

46:14

where the value is. And I think

46:16

that, you know, interesting, like, you know,

46:18

I was thinking about like a company

46:21

like Coinbase where, right, they have incredibly

46:23

valuable assets in the terms of money.

46:26

They have to stop people from breaking

46:28

in. They've done a massive

46:31

amount of work on that. They've seen all kinds of break

46:33

in types. I'm sure they have tons of data on that.

46:36

It's probably like, we're at least specific

46:38

to people trying to break into crypto

46:40

exchanges. And so, you

46:43

know, like I think it could be very useful for them. I

46:45

don't think they could sell it to anybody, but

46:49

you know, I think every company's got

46:51

data that if,

46:54

you know, fed into an intelligent system

46:56

would help their business. And I think

46:58

almost nobody has data that they could

47:01

just go sell. And then

47:03

there's this kind of in-between question, which is what

47:06

data would you want to let

47:08

Microsoft or Google or OpenAI or

47:10

anybody get their grubby little fingers

47:12

on? And that I'm

47:16

not sure. That,

47:18

that's a, that, that I think

47:21

is a question that enterprises are wrestling with more

47:23

than it's not so much. Should

47:25

we go like sell our data, but it should

47:27

we train our own model just so we

47:29

can maximize the value or

47:32

should we feed it into the big model?

47:35

And if we feed it into the big model, do

47:37

all of our competitors now have the thing that

47:39

we just did? And, you know, or

47:41

could we trust the big company to not

47:43

do that to us? Which

47:47

I kind of think the answer I'm trusting the big company

47:49

not to F with your data is

47:51

probably I wouldn't do that. If

47:55

your competitiveness depends on that, you probably shouldn't

47:57

do that. Well, there are at least

47:59

reports that sort of. certain big companies are using all kinds

48:01

of data that they should be using to train their

48:03

models already. Yep,

48:06

I think those reports are very

48:08

likely true. Right. Or

48:11

they'd have open data, right? Like this is, you know,

48:13

we've talked about this before, but you

48:16

have the same companies that are saying

48:18

they're not stealing all the data from

48:20

people or taking it in

48:22

an unauthorized way, refuse to

48:24

say open their data. Like

48:26

why not tell us where your data came from? And

48:29

in fact, they're trying to shut down all openness, no

48:31

open source, no open weights, no open data, none open,

48:33

nothing, and go to the government and try and get

48:35

to do that. You know, if you're

48:37

not a thief, then why are you doing that? Right,

48:40

right, right, what are you hiding? By the

48:42

way, there's other twists and turns here. So for

48:44

example, the insurance example, I kind of deliberately loaded

48:47

it because you may know, it's actually illegal to

48:49

use genetic data for insurance purposes, right?

48:51

So there's this thing called the Geno

48:54

Law, Genetic Information Non-Discrimination Act of

48:56

2008. And basically it

48:58

basically bans health insurers in the US

49:01

from actually using genetic data for the

49:03

purpose of doing health assessment, actuarial assessment

49:05

of, which by the way, because now

49:07

the genomics are getting really good, like that data probably actually

49:09

is among the most accurate data you could have if you

49:11

were actually trying to predict like when people are gonna get

49:14

sick and die. And they're literally not

49:16

allowed to use it. Yeah,

49:18

it is, I think

49:21

that this is an interesting, like

49:23

weird misapplication of

49:26

good intentions in

49:28

a policy way that's probably going

49:30

to kill more

49:32

people than ever get

49:35

saved by every kind

49:37

of health, FDA, et cetera, policy

49:39

that we have, which is, you

49:43

know, in a world of AI, having

49:46

access to data on all humans, why they

49:48

get sick, what their genetics were, et cetera,

49:50

et cetera, et cetera, that

49:53

is, you know, you don't talk about data being the

49:55

new oil, like that is the new oil, that's the

49:57

healthcare oil is, you know, if you could match

49:59

that data, those up, then we'd

50:01

never not know why we're sick. You

50:03

know, you could make everybody much healthier,

50:06

all these kinds of things. But, you

50:09

know, to kind of stop the insurance

50:11

company from kind of

50:13

overcharging people who are more likely

50:15

to die, we've kind of locked

50:18

up all this data. A

50:21

kind of better idea would be to

50:23

just go, okay, for the people who

50:25

are likely to, like we subsidize healthcare

50:28

like massively for individuals anyway,

50:30

just like differential, differentially

50:33

sub, you know, subsidize.

50:37

And, you know, and then like you

50:39

solve the problem and you don't lock up all the data.

50:41

But, yeah, it's typical of politics

50:44

and policy. I mean, most of them are like

50:46

that, I think. Well,

50:48

there's this interesting question, it's like an insurance, like

50:50

basically, one of the questions people have asked about

50:52

insurance is like, if you had perfectly predictive information

50:54

on like individual outcomes, does the

50:56

whole concept of insurance actually still work? Right?

50:58

Because the whole theory of insurance is risk

51:00

pooling, right? It's precisely the

51:02

fact that you don't know

51:04

what's going to happen in the specific case. That means

51:06

you build these statistical models and then you risk pool,

51:08

and then you have variable payouts depending on exactly what

51:10

happens. But if you literally knew what was

51:13

going to happen in every case, because for example,

51:15

you have all this predictive genomic data, then

51:17

all of a sudden it wouldn't make sense to risk pool,

51:19

because you just say, well, no, this person is going to

51:22

cost X, that person is going to cost Y, there's no...

51:24

Health insurance already doesn't make sense in

51:26

that way, right? Like, in current, the

51:29

idea of insurance is kind of like

51:31

the... It started with crop insurance where

51:33

like, okay, you know, my

51:35

crop fails. And so we all

51:37

put money in a pool in case like

51:40

my crop fails so that, you know, we can cover

51:42

it. It's kind of designed

51:44

for a risk pool for

51:46

a catastrophic, unlikely incident. Like,

51:49

everybody's got to go to the doctor all

51:52

the fucking time. And some people get sicker

51:54

than others and that kind of thing.

51:56

But like, the way our health insurance

51:58

works is like, all medical... gets

52:01

paid for through this insurance

52:03

systems, which is this layer

52:05

of loss and bureaucracy

52:07

and giant companies and all

52:09

this stuff. When like, if we're

52:12

gonna pay for people's healthcare, just pay for people's

52:14

healthcare. Like what are we doing? Right?

52:17

Like, and if you want to disincent people from

52:19

like going for nonsense reasons and

52:21

just up the copay, like it's,

52:24

like what are we doing? Just,

52:27

yeah. From a, yeah, from a justice standpoint, from a

52:29

fairness standpoint, like what it makes sense for me, you

52:31

know, what it makes sense for me to pay more

52:33

for your healthcare. If I

52:35

knew that you were gonna be more expensive than

52:37

me, like, you know, I'm direct, you know, if

52:39

you, if everybody knows what future healthcare costs is

52:41

per person, there has a very good

52:43

predictive model for it. You know, societal willingness to all pool

52:46

in the way that we do today might really diminish. Yeah,

52:48

yeah. Well, and then like, like,

52:50

you could also, if you knew, like

52:53

there's things that you do genetically and maybe we give

52:55

everybody a pass on that. It's like, you can't control

52:57

your genetics, but then like there's

52:59

things you do behaviorally that like dramatically increases

53:01

your chance of getting sick. And

53:04

so maybe, you know, we incentivize

53:06

people to stay healthy instead of

53:08

just like paying for them not

53:10

to die. The, there's

53:12

a lot of systemic fixes

53:15

we could do to the healthcare system. It

53:17

couldn't be designed in a more ridiculous way, I

53:19

think. Well, it couldn't be designed in

53:21

a more ridiculous way. It's actually more ridiculous in some

53:24

other countries, but it's pretty crazy here. Nathan,

53:27

Nathan Odie asks, what are the strongest common

53:29

themes between the current state of AI and

53:31

web 1.0? And so let

53:34

me start there. Let me give you a theory, Ben, and

53:36

see what you think. So I guess it's questionable, you know,

53:38

because of my role in, you know, Ben, you with me

53:40

at Netscape, you know, we get this question a lot because

53:42

of our role early on with the internet. And so there's

53:44

an, you know, the internet boom was like a major, major

53:46

event in technology, and it's still within a lot of, you

53:48

know, people's memories. And so, you know,

53:50

the sort of, you know, people like to reason from

53:52

analogy. So it's like, okay, the AI boom must be

53:54

like the internet boom starting an AI company, and it

53:56

must be like starting an internet company. And

53:59

so, you know, what is this? And we actually got a bunch

54:01

of questions like that, you know, that are kind of an analogy questions

54:03

like that I actually think you know, and then Ben, you know, you

54:05

and I were there for the internet boom So we you know, we

54:07

live through that and the bus and the boom and the bust So

54:10

I actually think that the analogy

54:12

doesn't really work for the most but it works in certain

54:15

ways But it doesn't really work for the most part and

54:17

the reason is because the the internet The

54:20

internet was a network Whereas

54:23

AI is a computer. Yep Okay,

54:26

yeah, so it's so some people understand what we're saying

54:29

the PC boom Or

54:32

even I would say the microprocessor like my

54:35

best Yeah,

54:38

or even to the like the original computers like back to the

54:40

mainframe era And the reason is

54:42

because yeah look what the internet did was the internet, you

54:44

know, obviously was a network But the network connected together many

54:46

existing computers And then of course people built many other new

54:48

kinds of computers to connect to the internet but fundamentally the

54:50

internet was a network and then and and

54:53

that's important because most of most

54:55

of the sort of industry dynamics

54:57

competitive Dynamics startup dynamics around

54:59

the internet had to do with basically building either

55:01

building networks or building applications that run on top

55:04

of networks and this you know then the internet

55:06

generation of startups was very consumed by network effects

55:08

and You know all these positive

55:10

so positive feedback loops that you get when you connect a

55:12

lot of people together You know

55:14

things like met you know, so-called Metcalfe's law Which is sort

55:16

of the value of a network, you know expands, you know

55:18

Kind of the way it expands is you have more people

55:20

to it and then you know There were

55:22

all these fights, you know these fights, you know All the social networks

55:24

or whatever fighting to try to get network effects and try to steal

55:26

each other's users Because the network

55:29

effects and so it's kind of it's dominated by

55:31

network effects Which is what you expect from from

55:33

it from a network business AI Like

55:35

there are some networks effects in AI that we

55:37

can talk about but it's it's more like a

55:39

microprocessor. It's more like a chip It's more like

55:41

a computer in that. It's

55:44

a system that basically right it Data

55:46

comes in data gets processed data comes out things

55:49

happen. That's a computer. It's an information

55:51

processing system. It's a computer It's a new kind of

55:53

computer. It's a you know, we like to

55:55

say the the sort of computers up until now have

55:57

been What are called by Norman machines, which is to

55:59

say that deterministic computers, which

56:01

is they're like, you know, hyper literal, and they do exactly

56:04

the same thing every time. And if they make a mistake,

56:06

it's the programmer's fault. But they're

56:08

very limited in their ability to interact with people and understand

56:10

the world. You know, we think

56:12

of AI and large language models as a

56:14

new kind of computer, a probabilistic computer, a

56:16

neural network based computer that, you know,

56:18

by the way, is not very accurate and is, you know,

56:20

doesn't give you the same result every time. And in fact,

56:22

might actually argue with you and tell you that it doesn't

56:25

want to answer your question. Yeah, which makes

56:27

it very different in nature than

56:30

the old computers. And it makes

56:32

it kind of comprehensibility, you know,

56:34

the ability to build things,

56:36

big things out of little things

56:39

more complex. Right.

56:42

But the capabilities are new and different and

56:44

valuable and important because it can understand language

56:46

and images and, you know, do all these

56:48

things that you see when you use. All

56:50

of our domains we can never solve with

56:52

deterministic computers we can now go after, right?

56:54

Right. Right. Yeah, exactly. And so I

56:57

think, I think, Ben, I think the analogy and I think

56:59

the lessons learned are much more likely to be drawn from

57:01

the early days of the computer industry or from the early

57:03

days of the microprocessor than the early days of the internet.

57:05

Does that, does that sound right? I

57:07

think so. Yeah, I definitely think so. And that

57:09

doesn't mean there's no like boom and bust and

57:11

all that because that's just the nature of technology.

57:13

You know, people get too excited and then they

57:15

get too depressed. So

57:18

there will be some of that, I'm sure. There will

57:20

be over build outs, you know, potentially eventually

57:22

chips and power and that kind of thing.

57:24

You know, we start with the shortage.

57:27

But I agree. Like I think networks are fundamentally

57:30

different in the nature of how they

57:32

evolved in computers and the kind of

57:34

just the adoption curve and all those

57:36

kinds of things will be different. Yeah.

57:40

So then this kind of goes to where, how I think

57:42

the industry is going to unfold. And so this is kind

57:44

of my best theory for kind of what happens from here,

57:46

this kind of this, you know, this giant question of like,

57:48

you know, is the industry going to be a few God

57:50

models or, you know, a very large number of models of

57:52

different sizes and so forth. So the

57:54

computer, like famously, you

57:56

know, the original computers, like the original

57:58

IBM mainframes, you know, the big computers,

58:00

they were very, very large and expensive and there

58:03

were only a few of them. And

58:05

the prevailing view actually for a long time was that's

58:07

all there would ever be. And there

58:09

was this famous statement by Thomas Watson, Sr.,

58:11

who was the creator of IBM, which was

58:13

the dominant company for the first 50 years

58:15

of the computer industry. And

58:18

he said, I believe this is actually true, but he said, I don't know

58:20

that the world will ever

58:22

need more than five computers. And

58:24

I think the reason for that, it was literally, it was

58:26

like the government's going to have two and then there's like

58:28

three big insurance companies and then that's it. Who

58:32

else would need to do all that math? Exactly.

58:35

Who else would need to, who else needs to

58:37

keep track of huge amounts of numbers? Who else

58:39

needs that level of calculation capability? It's

58:41

just not a relevant, you know, it's just not a

58:43

relevant concept. And by the way, they were like big

58:45

and expensive. And so who else can afford them, right?

58:48

And who else can afford all the headcount required to manage them and

58:50

maintain them? I mean, this is in the days, I mean, these things

58:52

were big, these things were so big that you would have an entire

58:54

building that got built around a computer, right?

58:56

And they'd have like, they famously have all these guys

58:59

in white lab coats, literally like taking care of the

59:01

computer because everything had to be kept super clean or

59:03

the computer would stop working. And so,

59:05

you know, it was this thing where, you know, today we have

59:07

the idea of an AI God model, which is like a big

59:09

foundation model that, you know, then we have the idea of like

59:12

a God mainframe. Like there would just be a few of these

59:14

things. And by the way, if you watch old science fiction, it

59:16

almost always has this sort of conceit. It's like, okay,

59:19

there's a big supercomputer and it either is like doing

59:21

the right thing or doing the wrong thing. And

59:24

if it's doing the wrong thing, you know, that's often the plot of

59:26

the science fiction movies is you have to go in and try to

59:28

figure out how to fix it or defeat it. And

59:30

so it's sort of this idea of like a single

59:32

top down thing, of course. And

59:34

that held for a long time. Like that held for, you

59:36

know, the first few decades. And then, you know, even when

59:38

computers, computers started to get smaller, so then you had so-called

59:41

mini computers was the next phase. And so

59:43

that was a computer that, you know, didn't cost $50 million instead

59:45

it costs, you know, $500,000, but even still $500,000 is a lot

59:47

of money. People

59:50

aren't putting mini computers in their homes. And so

59:52

it's like mid-sized companies can buy mini computers, but

59:54

certainly people can't. And then of course with

59:57

the PC, they shrunk down to like $2,500. And

59:59

then with the smartphone. and they shrunk down to $500. And

1:00:02

then sitting here today, obviously, you have computers

1:00:04

of every shape, size, description, all the way

1:00:06

down to computers that cost a penny. You've

1:00:09

got a computer in your thermostat that basically controls

1:00:11

the temperature in the room. And it probably cost

1:00:13

a penny. And it's probably some embedded arm chip

1:00:15

with firmware on it. And there's many

1:00:17

billions of those all around the world. You buy

1:00:19

a new car today. It has something. New cars

1:00:21

today have something on the order of 200 computers

1:00:23

in them, maybe more at this point. And

1:00:25

so you just basically assume with the chip sitting

1:00:27

here today, you just assume that everything has a

1:00:29

chip in it. You assume that everything, by the

1:00:31

way, draws electricity or has a battery because

1:00:34

it needs to power the chip. And then increasingly, you

1:00:36

assume that everything's on the internet because basically all computers

1:00:38

are assumed to be on the internet or they will

1:00:40

be. And so

1:00:42

as a consequence, what you have is the computer industry

1:00:44

today is this massive pyramid. And you

1:00:46

still have a small number of these supercomputer

1:00:48

clusters or these giant mainframes that

1:00:50

are like the god model, the god mainframes.

1:00:53

And then you've got a larger number of many computers.

1:00:55

You've got a larger number of PCs. You've got a

1:00:58

much larger number of smartphones. And then you've got a

1:01:00

giant number of embedded systems. And it turns out the

1:01:02

computer industry is all of those things. And what

1:01:05

size of computer do you want is

1:01:08

based on what exactly are you trying to do and who are you and

1:01:10

what do you need? And so

1:01:12

if that analogy holds, it basically

1:01:14

means actually we are going to

1:01:16

have AI models of every conceivable

1:01:18

shape, size, description, capability based

1:01:21

on trained on lots of different kinds of data running

1:01:23

at very different kinds of scale, very different privacy, different

1:01:26

policies, different security policies. You're

1:01:28

just going to have enormous variability

1:01:31

and variety. And it's going to be an entire ecosystem

1:01:33

and not just a couple of companies. Yeah,

1:01:36

let me see what you think of that. Well,

1:01:38

I think that's right. And I also think

1:01:40

that the other thing that's interesting about this

1:01:42

era of computing, if you look at priors

1:01:44

of computing from the mainframe to

1:01:46

the smartphone, a huge source

1:01:49

of lock-in was basically the

1:01:52

difficulty of using them. So

1:01:54

nobody ever got fired for buying IBM because

1:01:57

you had people trained on it.

1:02:00

you know, people knew how to use the

1:02:02

operating system. Like it was, you

1:02:05

know, it was just kind of like a

1:02:07

safe choice due to the massive complexity of

1:02:09

like dealing with a computer. And

1:02:12

then even with a smartphone, like the

1:02:14

re you know, why is

1:02:17

the Apple computer smartphone so dominant?

1:02:20

You know, what makes it so

1:02:22

powerful as well because like switching off of it

1:02:24

is so expensive and complicated and so forth. It's

1:02:27

an interesting question with AI because AI is the

1:02:30

easiest computer to use by far it speaks

1:02:32

English. It's like talking to a person. And

1:02:35

so like, what is the lock in there?

1:02:39

And so are you completely free

1:02:41

to use the size, price, choice,

1:02:43

speed that you need for your

1:02:46

particular task? Or are you locked

1:02:48

into the God model? And you know,

1:02:51

I think it's still a bit of

1:02:53

an open question, but it's pretty interesting

1:02:55

in that that that thing could be

1:02:58

very different than prior generations. Yeah,

1:03:02

yeah, that makes sense. And then just to complete the

1:03:04

question, what would we say? So Ben, what would you

1:03:06

say are lessons learned from the internet era that we

1:03:08

live through that would apply that people should think about?

1:03:12

I think a big one is probably

1:03:14

just the boom

1:03:17

bust nature of it that like,

1:03:20

you know, the demand, the

1:03:22

interest in the internet, the recognition

1:03:24

of what it could be was so

1:03:26

high that money just kind

1:03:28

of poured in and buckets. And

1:03:32

you know, and then the underlying thing,

1:03:34

which in internet age was the telecom

1:03:36

infrastructure and fiber and so forth, got

1:03:38

just unlimited funding and unlimited fiber was

1:03:40

built out. And then eventually we had

1:03:42

a fiber glut and all

1:03:44

the telecom companies went bankrupt and, and

1:03:47

that was great fun. But you know, like

1:03:49

we ended in a good place. And I

1:03:51

think that that's something like that's probably pretty

1:03:53

likely to happen in AI where

1:03:55

like, you know, every company is going to

1:03:58

get funded. We don't need that AI

1:04:00

companies, so a lot of them are going to bust.

1:04:02

There's going to be a huge, you know, huge

1:04:05

investor losses. There will be an overbuilt

1:04:07

out of chips for sure at

1:04:10

some point. And then, you know,

1:04:12

we're going to have too many chips and yeah, some

1:04:14

chip companies will go bankrupt for sure. And

1:04:17

then, you know, and I think probably

1:04:19

the same thing with data centers and so forth,

1:04:21

like, well be behind, behind, behind, and then we'll

1:04:24

overbuild at some point. So

1:04:27

that, that, that'll all be very

1:04:29

interesting. I think that, and that's kind of

1:04:31

the, that's

1:04:33

every new technology. So Carlotta Perez has a

1:04:35

great kind of, has done, you know, amazing

1:04:38

work on this where like, that is just

1:04:40

the nature of a new technology is that

1:04:42

you overbuild, you underbuild, then you overbuild and,

1:04:44

you know, and there's a hype cycle that

1:04:47

funds the build out and a lot of

1:04:49

money is lost, but we get the infrastructure and

1:04:51

that's awesome because that's when it really gets adopted

1:04:54

and changes the world. I want

1:04:56

to say, you know, with the internet,

1:04:58

the other, the other kind of big kind of thing

1:05:01

is the internet went through a couple

1:05:03

of phases, right? Like it went through a

1:05:05

very open phase, which was unbelievably great. It

1:05:08

was probably one of the greatest booms to

1:05:10

the economy. It, you

1:05:12

know, it, it certainly created tremendous

1:05:14

growth and power in America, both,

1:05:16

you know, kind of economic power

1:05:18

and soft cultural power and these

1:05:20

kinds of things. And then, you

1:05:22

know, it became closed with the

1:05:24

next generation architecture, with, you know, kind

1:05:26

of discovery on the internet being

1:05:28

owned entirely by Google and, you know, kind

1:05:30

of other things, you know, being owned by

1:05:33

other companies. And,

1:05:35

you know, AI, I think could go either

1:05:37

way. So it could be very open or

1:05:39

like, you know, with kind of misguided regulation,

1:05:42

you know, we could actually force our way

1:05:45

from something that, you know, is open

1:05:47

source, open weights, anybody can build it.

1:05:49

We'll have a plethora of this

1:05:52

technology will be like, use

1:05:54

all of American innovation to

1:05:57

compete or will, you know,

1:05:59

will. cut it all off, we'll force it

1:06:01

into the hands of the

1:06:03

companies that kind of own

1:06:05

the internet today and we'll

1:06:08

put ourselves at a huge disadvantage,

1:06:10

I think, competitively against China in

1:06:12

particular, but everybody in the world. So

1:06:16

I think that's something that definitely

1:06:18

that we're involved with trying

1:06:20

to make sure it doesn't happen, but it's

1:06:22

a real possibility right now. Yeah.

1:06:25

The certain irony is that networks used to

1:06:27

be all proprietary. Yeah. Yeah,

1:06:30

right. Landman, Apple Talk,

1:06:32

Netbuoy, Netbios. Yeah,

1:06:34

exactly. And so these are all the early proprietary

1:06:36

networks from all individual specific vendors and then the

1:06:38

internet appeared in kind of TCP, IP and everything

1:06:40

opened up. The AAI is trying to

1:06:43

go the other way. The big companies trying to take

1:06:45

AAI the other way. It started out as like open,

1:06:47

just like basically just like the research. Everything was open

1:06:49

source in AAI. Right. Right. And

1:06:52

now they're trying to lock it down. So it's

1:06:54

a fairly nefarious turn of events. Yeah,

1:06:56

very nefarious. You know,

1:06:58

like it's remarkable

1:07:00

to me. I mean, it is kind of

1:07:03

the darkest side of capitalism when a company

1:07:06

is so greedy, they're willing to destroy the

1:07:08

country and maybe the world to like just

1:07:10

get a little extra profit. But, you know,

1:07:14

and they do it like the really kind

1:07:16

of nasty thing is they claim, oh, it's for

1:07:18

safety. You know, we've created an

1:07:20

alien that we can't control, but we're not

1:07:22

going to stop working on it. We're going to keep building

1:07:24

it as fast as we can. And we're going to buy

1:07:26

every freaking GPU on the planet. But we

1:07:28

need the government to come in and stop it from

1:07:30

being open. This

1:07:33

is literally the current position of

1:07:35

Google and Microsoft right now. It's

1:07:37

crazy. And we're not going

1:07:39

to secure it. So we're going to make sure that like

1:07:41

Chinese spies can just like steal our chip plans, take them

1:07:44

out of the country. And we won't even realize for six

1:07:46

months. Yeah, it has nothing to do with security. It only

1:07:48

has to do with Monopoly. Yes.

1:07:50

The other, you know, just been going back on

1:07:53

your point of speculation. So there's this critique that

1:07:55

we hear a lot, right? Which is like, OK,

1:07:57

you idiots. Basically, it's like you idiots, you idiots,

1:07:59

entrepreneurs, investors. You idiots, it's like there's

1:08:01

a speculative bubble with every new technology. Like basically like

1:08:03

when are you people going to learn to not do

1:08:05

that? And there's no

1:08:08

joke. There's no joke that relates to this, which is

1:08:10

the foremost dangerous words in investing are this time is

1:08:12

different. The 12 most dangerous

1:08:14

words in investing are the foremost dangerous words

1:08:16

are investing are this time is different. Right?

1:08:19

Like, so like, does history repeat? Does it not

1:08:21

repeat? My sense of

1:08:23

it, and you referenced Carlotta Perez's book, which I agree is

1:08:25

good, although I don't think it works as well anymore. We

1:08:28

can talk about some time, but, but you know, it's a

1:08:30

good at least background piece on this. It's

1:08:32

just like, it's just incontrovertibly true. Basically every

1:08:34

significant technology advance in history was greeted by

1:08:36

some kind, some kind of financial bubble, basically

1:08:38

since financial markets that existed. And

1:08:41

this, by the way, this includes like everything from,

1:08:43

you know, radio and television, the railroads, you know,

1:08:45

lots and lots of prior, by the way, there

1:08:47

was a, there was actually a, a so-called, there

1:08:49

was an electronics boom bust in the sixties called

1:08:51

the, it was called the Tronics. Every company had

1:08:53

the name Tronics. And so, you know,

1:08:55

there, there was that. So

1:08:57

there, you know, there was like a laser boom, bust cycle there.

1:08:59

There were all these like boom, bust cycles. And so basically it's

1:09:02

like any new tech, any new technology. That's

1:09:04

what economists call a general purpose technology, which is to say

1:09:06

something that can be used in lots of different ways. Like

1:09:08

it inspires sort of a speculative mania and you know, and

1:09:11

look, the critique is like, okay, why do you need to

1:09:13

have this speculative mania? Why do you need to have the

1:09:15

cycle? Because like, you know, people, you know, some people invest

1:09:17

in the things, they lose a lot of money and

1:09:20

then there's this bus cycle that, you know, causes everybody to get

1:09:22

depressed. Maybe the ways to roll out. And

1:09:24

it's like two things. Number one is like, well, you

1:09:26

just don't know. Like if it's a general purpose technology

1:09:28

like AI is and it's potentially useful in many ways,

1:09:30

like nobody actually knows upfront, like what the successful use

1:09:32

cases are going to be or what successful companies are

1:09:34

going to be. Like you actually have to, you have

1:09:37

to learn by doing. You're going to have to miss

1:09:39

this. That's venture capital. Yeah, exactly.

1:09:41

We need to follow. Yeah, exactly.

1:09:43

So yeah, the true venture capital model kind of

1:09:45

wires this in, right? We, we, we basically, core

1:09:47

venture capital, the kind that we do, we sort

1:09:49

of assume that half the companies fail, half the

1:09:51

projects fail. And you know,

1:09:53

if, if, if any of us, if we are, anybody

1:09:55

fail completely, like lose money, yeah, like lose

1:09:57

money. Exactly. Yeah. If

1:10:00

we or any of our competitors could figure out how to do

1:10:02

the 50% that work without doing the 50% that

1:10:04

don't work, we would do that. But

1:10:06

here we sit 60 years into the field

1:10:08

and nobody's figured that out. So there

1:10:11

is that unpredictability to it. And

1:10:13

then the other interesting way to think about this is

1:10:15

like, okay, what would it mean to have a society

1:10:18

in which a new technology did not inspire speculation? And

1:10:21

it would mean having a society that

1:10:23

basically is just inherently super pessimistic about

1:10:25

both the prospects of the new technology,

1:10:27

but also the prospects of entrepreneurship. And

1:10:30

people inventing new things and doing new things. And

1:10:32

of course, there are many societies like that on

1:10:34

planet Earth that just fundamentally

1:10:36

don't have the spirit of invention

1:10:38

and adventure that a place like

1:10:40

Silicon Valley does. And

1:10:42

are they better off or worse off? And generally speaking,

1:10:45

they're worse off. They're

1:10:47

just less future oriented, less focused

1:10:49

on building things, less focused on figuring

1:10:51

out how to get growth. And

1:10:53

so I think there's, at least my sense, there

1:10:55

comes with the territory thing. We

1:10:59

would all prefer to avoid the downside of a speculative boom-bust

1:11:01

cycle, but it seems to come with the territory every single

1:11:03

time. And at least I have not, nobody,

1:11:05

no society I'm aware of has ever figured out how to

1:11:07

capture the good without also having the bad. Yeah.

1:11:10

And like, why would you? I mean, it's

1:11:13

kind of like the whole Western

1:11:15

United States was built off the gold rush. And

1:11:18

every kind of treatment in the popular

1:11:20

culture of the gold rush kind of

1:11:22

focuses on the people who didn't make

1:11:25

any money. But there were people who made a

1:11:27

lot of money. And

1:11:29

found gold. And in the

1:11:32

internet bubble, which was completely ridiculed

1:11:34

by every movie, if you go

1:11:36

back and watch

1:11:39

any movie between like 2001 and 2004, they're

1:11:41

all like how only morons did a dot com

1:11:48

and this and that and the other. And there

1:11:50

were all these funny documentaries and so forth. But

1:11:54

that's when Amazon got started. That's

1:11:57

when eBay got started. That's when Google got

1:11:59

started. And these companies that

1:12:01

were started in the bubble

1:12:04

in the kind of time of this

1:12:06

great speculation, there was gold in those

1:12:09

companies. And if you hit

1:12:11

any one of those, like you funded

1:12:13

probably the next set of companies, which

1:12:16

included things like Facebook and X and

1:12:19

Snap and all these things.

1:12:21

And so, yeah, I mean, like that's

1:12:23

just the nature of it. I mean, like that's what makes it

1:12:25

exciting. And it's

1:12:28

an amazing kind of thing

1:12:31

that, look, the transfer of

1:12:34

money from people who have excess money to

1:12:37

people who are trying to do new things and make

1:12:39

the world a better place is

1:12:41

the greatest thing in the world. And

1:12:44

if we, some of the people with excess money

1:12:47

lose some of that excess money and

1:12:49

trying to make the world a better place, like why

1:12:51

are you mad about that? Like that's the thing that

1:12:53

I could never have seen. Like, why would you be

1:12:55

mad at young, ambitious

1:12:57

people trying

1:13:01

to improve the world, getting funded

1:13:04

and some of that being misguided? Like, why is

1:13:06

that bad? Right, right.

1:13:08

As compared to, yeah, as compared to, especially

1:13:11

as compared to everything else in the world

1:13:13

and all the people who are not trying

1:13:15

to do that. As you'd rather like, we

1:13:17

just buy lots of mansions and boats and

1:13:19

jets and, really, like, what are you talking

1:13:21

about? Right, exactly. Donate

1:13:23

money to ruin us. You mean, ruin

1:13:26

his causes, right? Such

1:13:29

as ones that are on the news

1:13:31

right now. Okay, so, all right, we're at a minute 20. We made

1:13:33

it all the way through four questions. We're

1:13:35

doing good, we're doing great. So let's call it here. Thank

1:13:37

you, everybody, for joining us. And I believe we should do

1:13:39

a part two of this, if not parts three through six,

1:13:41

because we have a lot more questions to go, but thanks

1:13:43

everybody for joining us today. All right, thank you.

1:13:46

Thank you.