0:00

positioning is one of the most powerful ways

0:02

a startup can have an insight. Most people

0:04

think an insight is just about product. The

0:06

what is the product, the how is how

0:08

you deliver the product. And the how can

0:10

have an insight as well. And when you're

0:12

in a lawsuit, let's say you're Apple and

0:14

you're in a lawsuit with Samsung, there's a

0:16

ton of documents that have to get discovered

0:18

for the core case. They hire these outsourcer

0:20

firms of people to go pour through these

0:22

documents and they charge them on a cost

0:24

plus basis. So what Text IQ

0:26

said is, well, we've got AI Why

0:28

don't you just send us all your

0:30

documents and we'll send you back the

0:32

ones that are discoverable and we'll have

0:35

more accuracy. Now you're not competing for

0:37

software license or per seat revenue or

0:39

even a subscription price. You're saying I'm

0:41

a substitute for that labor spend. You

0:43

used to spend 50 million dollars a

0:45

year on this so I can do

0:47

it for a tenth of the price

0:49

and much better. If I'm a SaaS

0:51

vendor and I charge subscription by the

0:53

seat and that's all I've ever done.

0:55

Think about how embedded that must be

0:57

in the culture, right? Every product manager

0:59

thinks that way. The CFO thinks that

1:01

way. There's nobody in the company who

1:03

knows how to react to your strategy.

1:05

If you change your business model, everyone's

1:07

going to lose their mind. OpenAI is

1:09

moving from being this API developer tool

1:12

to like a product company. They're releasing

1:14

all of these consumer -facing product. A

1:16

lot of founders are thinking about what

1:18

if OpenAI includes this as part of

1:20

ChatGP here, includes this in some new

1:22

product that they release. And I'm curious

1:24

how you would think about... positioning that.

1:26

I'm involved with a company called Applied

1:28

Intuition and they create simulation software for

1:30

autonomous vehicles. If you're GM or if

1:32

you're Porsche or you're these big companies,

1:34

that's pretty valuable. But you can't just

1:36

get that when Sam Altman releases his

1:38

next demo. at a demo day event.

1:40

To succeed as a company like that

1:42

and to really ask for giant contracts

1:44

from these companies, you have to have

1:46

not only AI expertise and products, but

1:48

you have to have multi -discipline expertise.

1:51

Everybody says, kind of, this is on

1:53

these companies that are just an AI

1:55

wrapper, right? And I'm like,

1:57

well, if the thing that you're wrapping on

1:59

top of involves a process that you really

2:01

know about that most people don't, that may

2:03

be a path to a great company. Mike,

2:19

welcome to the show. Thanks for having me. I've

2:21

been looking forward to this. Yeah. For

2:24

people who don't know you, you are a

2:26

legendary investor at Floodgate, which is one of

2:28

the first seed firms. You're

2:30

an early investor in Twitch, Lyft, Okta, and a bunch

2:32

more. You're also the author of

2:34

the book, Pattern Breakers, which is an excellent book that

2:37

I've read. We've also reviewed it on every which

2:40

I guess I would summarize by it's sort

2:42

of like a guidebook about how there's no

2:44

guidebook to building companies. So

2:46

it's very, it's a little bit Taoist. A little

2:49

bit Zen. Yeah. Yeah, a little

2:51

Zen, which I love. I think that's

2:53

so, I think that's so good and

2:55

so important. You have a lot of

2:57

emphasis on like founders winning by being

2:59

extraordinarily different and breaking the established patterns

3:01

of like how you're supposed to run

3:03

a company. I loved it and I'm

3:05

excited to chat with you about that

3:07

and everything going on in AI on

3:09

the show. Yeah. Well,

3:11

cool. Let's get after it. Let's

3:13

do it. One of

3:15

the things that I'm

3:17

personally curious about is

3:19

you started investing when

3:21

seed wasn't really a

3:24

thing and helped to invent

3:26

this new way of capitalizing

3:28

companies for an

3:30

earlier era of startups, pre -AI startups, let's

3:32

just say that, right? And I think that

3:35

that is an example of the kind of

3:37

thing that you talk about in your book,

3:39

Pattern Breakers, which is like taking a look

3:41

at the landscape of maybe what companies need

3:43

and how companies are funded and being like,

3:45

well, there's this thing that seems to make

3:47

a lot of sense to me that there

3:50

should be a seed stage funding mechanism

3:52

and just going and doing it. And

3:55

I'm kind of curious My

3:59

feeling right now is that AI

4:01

is radically changing the economics of starting

4:03

a business. Software

4:06

orders a magnitude cheaper to make today than

4:08

it was 10 years ago. I'm

4:11

curious, using that same sense

4:14

of, okay, I'm looking at the environment

4:16

and looking at how things change, and

4:18

I'm maybe pushing away the established structures

4:20

for a second. How do you think

4:22

that that might change investing in how

4:24

companies raise money and all that kind

4:26

of stuff? Yeah, I've been

4:28

wondering about this a lot lately. So

4:30

as you know, one of the things

4:33

that I emphasize in startups is the

4:35

power of harnessing inflections, right? So I

4:37

like to say that, you know, business

4:39

never a fair fight. And

4:41

the startup has to have some unfair

4:43

advantage way to win. And the way

4:46

they do that is they harness inflections.

4:48

Inflections allow the startup to wage asymmetric

4:50

warfare on the present and show up

4:52

with something radically different. without

4:54

inflections, they have to play in the incumbent

4:57

sandbox. And so they're limited in their upside.

5:00

So every now and then, though, you get something

5:02

that I like to call a sea change. And

5:04

when I was a kid, the sea change

5:07

was a mass computation in the personal computer.

5:10

And computers used to be really expensive. And

5:13

then they became asymptotically free and ubiquitous.

5:15

And you had one on every desk

5:17

in every home. And a whole

5:19

new set of companies emerged. Software

5:21

became a real business for the

5:23

first time. Software used to be

5:25

what you gave away because mainframes

5:27

were expensive. You had to keep

5:29

them running all the time And

5:31

so so the assumptions got inverted

5:34

and you had a bunch of

5:36

companies using the software licensing model,

5:38

you know Oracle Microsoft SAP companies

5:40

like that Then you had in

5:42

the 90s the era of mass

5:44

connectivity, which I think was extended

5:46

with the iPhone and in mass

5:48

connectivity rather than Processing power becoming

5:50

free, communications bandwidth starts to become

5:52

free, and you start to not

5:54

just have computers everywhere, but

5:56

you have everybody in the world and every device

5:58

in the world connected in these networks. And

6:01

new business models came out of that,

6:03

subscription and SaaS and advertising. It's

6:07

interesting, there aren't any software licensing

6:09

model companies started after 1990 that

6:11

really mattered. All those

6:13

companies got subsumed in Microsoft because they could

6:16

put it in the OS or outcompetum.

6:19

So like why do I think the AIC

6:21

change matters? What I

6:24

see happen with these C

6:26

changes is that some business

6:28

models become relatively more attractive

6:30

and some business models become

6:32

relatively less attractive. And

6:34

there's only nine business models that I

6:36

know of in human history. And

6:39

so the most recent business model I

6:41

know of is 250 years old. It's

6:43

the subscription model. And so You know,

6:45

what I what I like to do

6:47

is I like to say, OK, if

6:49

there's nine business models so far in

6:51

humanity and every time there's a technology

6:53

sea change, there's a migration of attractive

6:55

business models from one set to the

6:57

other. How might that migration

6:59

occur this time? Because what you want

7:01

when you're a startup is to be

7:03

counter positioned to the incumbents. You know,

7:06

this whole the incumbents have the advantage.

7:08

Discussion is wrong headed. Of

7:10

course, the incumbent has the advantage if you play

7:12

by the rules of the incumbents. But

7:14

what you want to do is you

7:16

want to say, how does AI make

7:18

some business models relatively more attractive and

7:21

less attractive? And how can

7:23

I as a startup exploit those

7:25

new opportunities? Not just insight in

7:27

my product, but some type of

7:29

an insight in my business model

7:31

go market strategy that disorient incumbents

7:34

and where they have a disincentive

7:36

to retaliate or to copy your

7:38

strategy. So that's mostly what I'm

7:40

looking at these days from an

7:42

AI point of view. Yeah,

7:45

so I think like one of the things

7:47

that I see a lot from the business

7:49

model perspective and right now we're talking about

7:51

business models for startups. I would also like

7:54

to talk about business models for venture like

7:56

funding startups. Sure. But business

7:58

models for startups just to start there

8:00

for a second. One

8:02

of the things I'm seeing a lot of

8:04

is paying per outcome as opposed to paying

8:06

per month. Yes. Which I think is a

8:09

really interesting one. Is that something you have

8:11

your eye on? Oh, absolutely.

8:14

So, you know, there's a

8:16

business model called tailored services

8:18

with long -term contracts. And

8:21

right now, most people think that's

8:23

unattractive. What is tailored services of

8:25

long -term contracts? That could be

8:28

like the defense subprimes. It could

8:30

be a contract research organization for

8:32

a pharma company. You

8:34

know, it's somebody that you that

8:36

offer services on a contract basis,

8:39

usually is labor intensive, usually is

8:41

cost plus. And the

8:43

conventional wisdom today is those are not

8:45

attractive opportunities for software companies. Like

8:48

a law firm or something? Like

8:50

a law firm, perfect example. So

8:52

like an example, like a law

8:54

firm or legal services, a

8:57

company I was involved with a few years ago was

8:59

called Text IQ. And they would

9:01

go to a big corporation and they would say, you

9:03

know, when you're in a lawsuit, let's say you're Apple

9:05

and you're in a lawsuit with Samsung, there's

9:08

a ton of documents that have to

9:10

get discovered for the court case. And

9:12

so the way that happens in reality

9:14

is they hire these outsourcer firms of

9:16

people to go pour through these documents

9:18

and they charge them on a cost

9:20

plus basis. Yeah. And so what Text

9:22

IQ said is, well, we've got AI.

9:25

Why don't you just send us all your documents and

9:27

we'll send you back the ones that are discoverable and

9:29

we'll have more accuracy. Well, now

9:31

you're not competing for software license

9:34

desktop revenue or proceed revenue or

9:36

even a subscription price. you're

9:38

saying, hey, look, I'm a substitute for that

9:41

labor spend. You used to spend

9:43

$50 million a year on this contract

9:45

outsourcer that sorts through these documents. I

9:47

can do it for a tenth of

9:49

the price and much better. And

9:52

now you're competing over that labor

9:54

cost bucket rather than the software

9:56

spend bucket. And how many seats

9:58

do I get? Well, that's interesting

10:00

because there's cost per task done.

10:02

So it's cost per document processed

10:04

or whatever, which is sort of

10:06

like what OpenAI does when you

10:08

send them a prompt, they send

10:11

a response. But

10:13

even if they send the response and the response

10:15

isn't good, you still pay for it. Right? And

10:18

then there's other companies that are sort

10:20

of capturing the value, part of the

10:22

value that they generate. So it's, it's,

10:24

it's, if they increase your, let's say,

10:26

let's say it's a SDR bot, if

10:29

they increase your sales by some amount,

10:31

your close rate, they take a percent

10:33

of that, only when it's successful. Have

10:36

you looked at those two? Yeah.

10:38

And so I do like

10:40

the outcome based pricing models

10:43

a lot. You know,

10:45

they both have their virtues, right? The thing

10:47

about open AI is like you could use

10:49

Dolly to generate some art that you don't

10:51

think looks pretty enough. But

10:53

open AI probably deserves to be compensated for

10:56

the fact that you did that, right? Yeah.

10:58

It's sometimes hard to know if the job was

11:01

done well or not. It's like, it's not so

11:03

clear. And sometimes it's the customer's fault that the

11:05

job wasn't done well, right? It's tricky. You

11:08

know, back in my ancient days when I was a

11:10

founder, I used to have

11:12

this expression when I would sell enterprise software.

11:15

I called it, what does it take to ring

11:17

the bell? And so like if

11:19

you go into the carnival, you know how there's that thing

11:21

where you have this big mallet and you hit this thing

11:23

and hopefully it goes all the way up and rings the

11:25

bell. But if it doesn't go all the way up, it

11:27

makes no sound. It has to go all the way up

11:29

and ring the bell. It's binary. And

11:31

so what I used to say to

11:33

the folks that I would work with

11:36

is that The customer doesn't care that

11:38

your software ran according to how the

11:40

specification works. That's not what they're buying.

11:42

They have a job to be done.

11:45

They're hiring your product to do a job. And

11:48

so we need to understand what's it going to take

11:50

to ring the bell for doing that job. And

11:52

if we ring the bell, they're going to say,

11:55

this is amazing. I want more of this. If

11:57

it doesn't ring the bell, they're not going to

11:59

care that the mechanism of our system works. They're

12:01

not going to be interested in that. And

12:04

so for me, the outcome -based models

12:06

that we were just talking about a

12:08

minute ago are kind of asking that,

12:10

what is the job to be done

12:12

in a Clay Christiansen sort of lens?

12:15

And then what does it mean to ring the bell? And

12:18

can I get paid if I unambiguously

12:20

succeed at that over and over again?

12:22

And the thing that makes that I think interesting

12:25

over like a SaaS model is that the incumbents

12:27

are all going to be SaaS. And

12:29

if you're guaranteed to get 20 bucks a

12:31

seat or whatever it is, the

12:34

idea of moving to like a pay for

12:36

performance model is like very unappealing. That's

12:38

right. So startups to your counter positioning point,

12:40

like that's a thing that startups can do

12:42

that incumbents like some incumbents already do this.

12:44

Like in the customer service world, this has

12:46

been a thing for forever. But in general,

12:49

Um, this is not a thing. And so

12:51

incumbents are not going to be able to

12:53

do this very well. Yeah. I think that

12:55

this counter positioning thing is a really important

12:57

thing to maybe double click on. And so

12:59

like, um, a great example

13:01

is in the nineties, if you were

13:03

a startup, the words that you dreaded

13:06

to hear was Microsoft has decided to

13:08

compete in your market because you're just

13:10

like, okay, I guess I'm out of

13:12

business because even if they start losing,

13:14

they're just going to bundle this thing

13:16

in windows. And I'm just host, right?

13:18

And so. that was happening to a

13:21

lot of companies you know netscape just

13:23

disappeared basically because microsoft decides to bundle

13:25

the browser you know in the operating

13:27

system and go go full ham right

13:29

against netscape. Well then

13:31

the internet happens and then some

13:33

people start to discover that you

13:35

can monetize not by selling by

13:37

the cedar by the desktop by

13:39

selling ads. And that was

13:42

google and microsoft had no answer to

13:44

that. You can't bundle something in your

13:46

operating system and deal with the fact

13:48

that Google is pricing by ads, right?

13:51

It doesn't solve the problem. It doesn't

13:53

impact their business at all. And

13:56

so Google was counter positioned to Microsoft

13:58

from business model perspective. And

14:00

like counter positioning is one of the

14:03

most powerful ways a startup can have

14:05

an insight. Most people think an insight

14:07

is just about product, but it can

14:09

also be about the what is the

14:11

product, the how is how you deliver

14:13

the product. And the how can have

14:16

an insight as well. And quite often,

14:18

the very best, most valuable companies have

14:20

an insight around business model that's facilitated.

14:22

Google's business model couldn't work before the

14:24

internet. The technology wouldn't

14:26

have provided the empowerment necessary for Google

14:29

to monetize with ads, but now all

14:31

of a sudden it did. And

14:34

so that's what we look for with

14:36

this counter positioning. And to

14:38

your point, right? Like now, it's sell the

14:40

work, not the software. If I'm

14:42

a company, if I'm a SaaS vendor and

14:44

I charge subscription by the seat, and that's

14:46

all I've ever done, think about

14:48

how embedded that must be in the culture,

14:50

right? Every product manager thinks that way. The

14:53

CFO thinks that way. You

14:56

know, there's nobody in the company who knows

14:58

how to react to your strategy because the

15:00

investors think that way. Everybody does. You know,

15:02

if you change your business model, everyone's going

15:04

to lose their mind. Yeah. So how, how

15:06

would you even think about changing it? Midstream

15:09

you just it's just that even if

15:11

you knew to have the insight that

15:13

perhaps you should consider it you just.

15:15

You just wouldn't have the wherewithal to

15:17

do it because it just it's so

15:19

embedded in your culture your entire value

15:21

delivery system is predicated on different model.

15:24

Yeah well let's keep talking about about

15:26

counter positioning and I want to bring

15:28

up I think like if I have

15:30

to pick who the Microsoft is of

15:32

the AI world. Like

15:34

huge huge huge tech companies like Microsoft

15:36

and Google aside. I think

15:39

the one right now to think about

15:41

kind of positioning or at least a

15:43

lot of startups are afraid of is

15:46

OpenAI. OpenAI is moving from being this

15:48

API developer tool to like a product

15:50

company. They're releasing all these consumer -facing

15:53

products. ChatGBT is sort of like taking

15:55

over. And so I think a lot

15:57

of founders are thinking about, well, what

16:01

you know, Chatchapiti includes this in their,

16:03

OpenAI includes this as part of Chatchapiti

16:05

or includes this in some new product

16:07

that they release. And I'm curious how

16:09

you would think about counter positioning that.

16:13

Yeah. So there, there are a couple of

16:15

ways. There are a couple of things I

16:17

find really interesting about OpenAI from a counter

16:19

positioning. So maybe, maybe we start with startups

16:21

and then just, there's some general stuff too,

16:23

like with deep seek and things like that.

16:25

But, but, um, so. Like,

16:28

let's just take an example. I'm

16:30

involved with a company called Applied Intuition, and

16:33

they create simulation software. I

16:35

love that name, by the

16:37

way. Yeah, it's pretty good.

16:39

It creates simulation software for

16:41

autonomous vehicles and also technology

16:43

stacks for electric vehicles. And

16:46

these car companies, other than Tesla, don't really

16:48

know how to do EVs, don't know how

16:51

to do AVs. They don't really even know

16:53

how to do software, right? Their entire business

16:55

model is predicated on a supply chain that's

16:57

100 years old where they get parts from

17:00

Bosch and chips from all these people and,

17:02

you know, parts from different tool and dye

17:04

shops and everything else. So,

17:06

so applied intuition says, OK, we've got

17:08

a bunch of people from Google and

17:10

Waymo and now some people from Tesla

17:13

and all the best autonomous vehicle, all

17:15

the best EV companies in the world.

17:18

We can build the entire thing that

17:20

you need. to sort of

17:22

update your strategy and roadmap to have

17:24

the software to find car, which is

17:26

where the future is going. Now,

17:30

if you're GM or if you're

17:32

Porsche or you're these big companies,

17:34

that's pretty valuable, but you can't just

17:37

get that when Sam Altman releases his

17:39

next demo at a demo day event,

17:41

right? Like, you know, if you're gonna

17:43

have a software to find car, there's

17:46

a whole lot of things that you

17:48

have to know intimately. the

17:50

processes of how cars are made

17:52

and manufactured and tested the whole

17:54

supply chain and you know how

17:56

the delivery system works. And

17:59

so you to succeed as a company like

18:01

that and to really ask for giant contracts

18:03

from these companies. You have

18:05

to have not only AI expertise and products

18:08

but you have to have multi discipline expertise.

18:11

you know, so like Casper and Peter, they

18:13

grew up in Detroit. But, you know, before

18:15

they got in Google and Waymo, they were,

18:17

you know, in the car industry at GM.

18:20

And so, yeah, so I like companies

18:22

like that where, you know,

18:25

one way I like to think about is

18:27

everybody says, is kind of disses on these

18:29

companies that are just an AI wrapper, right?

18:32

And I'm like, well, if the

18:34

thing that you're wrapping on top

18:36

of involves a process that you

18:38

really know about, that most people

18:40

don't, that may be a path

18:43

to a great company. And

18:45

so I think that that's what I'm interested

18:47

in, is some of those. The AI wrapper

18:50

thing was so silly. I see

18:52

less of that now, which is nice. But

18:54

it was a very silly thing when it first

18:56

started. So one other thing

18:58

about this counter positioning and open AI

19:00

that I think is interesting, and I'd

19:03

love to get your read on, is

19:05

one way I have internalized the deep

19:07

seek stuff, is that in the early

19:10

days of the internet, all

19:12

of the researchers from Bell

19:14

Labs and folks from AT

19:16

&T, Time Warner, the

19:19

government said, this internet thing's a

19:21

toy. It's never to be good enough.

19:25

We've tried this before. It doesn't work. These

19:27

protocols are not going to be robust enough.

19:30

And in the short term, you would have been right. None

19:33

of these things looked all that interesting or

19:35

impressive. But, you know, was

19:37

talking to Steve Sinovsky about this the other

19:39

day. You know, was at Microsoft at

19:41

the time when the internet took off. And he was

19:43

at Cornell and he saw, see, you see me and

19:46

he goes to Gates. You know, this is going to

19:48

be a tidal wave. This is going to be a

19:50

giant new phenomenon that we got to really pay attention

19:52

to. Deepseek reminds

19:55

me of that. So like the culture

19:57

in AI, the hyperscalers right

19:59

now is you can solve all problems

20:01

by throwing money at it. And

20:04

the deepseek guys. who said, well,

20:06

if we're limited with some fundamental

20:08

constraints, what would we do? I

20:11

think that there's going to be

20:13

a cultural shift in AI where

20:15

many people adopt that mindset. And

20:17

you know, that's important because the

20:19

early days of mass computation, the

20:22

IBM PC had a 640K memory limit.

20:25

And so like the Microsoft programmers had

20:27

an advantage because they could write small

20:29

fishing code. It wasn't how many thousand

20:31

lines of code anymore. It was how

20:33

efficient is your code. And

20:35

I think that we might

20:37

see the same phenomenon here

20:39

where people come from the

20:42

bottoms up with very frugal,

20:44

sort of low cost by

20:46

design solutions. And it'll

20:48

be hard for the open AIs and

20:50

the anthropics and those guys. I mean,

20:52

I have huge respect for what they're

20:54

doing, but it'll be hard for them

20:57

to respond to that because they're culturally

20:59

embedded in their operating model is to

21:01

solve everything by throwing money at it.

21:03

you know, hire the best people, throw

21:05

money at it and just keep going,

21:07

keep going faster. That's

21:10

so interesting. You said so many things

21:12

I want to talk about. So one

21:14

is sort of like this, this toy

21:16

thing where people and governments are like

21:18

big companies, like sort of ignore the

21:20

internet at first because they're like, we

21:22

tried it and it doesn't work. It

21:24

doesn't scale or whatever. You have the

21:26

same history with neural networks where like

21:28

in the beginning of AI and symbolic

21:30

AI, like in the 50s, neural networks

21:32

were around then. But they were mostly

21:34

ignored because the early AI people, particularly

21:36

like Marvin Minsky, proved

21:40

that single layer neural networks

21:42

were not as powerful as

21:44

other types of like Turing

21:47

machines, basically, or couldn't

21:49

do certain types of computations. And

21:52

I think academia sort of... by

21:55

and large felt like neural networks were

21:58

not understandable enough. There was no theory,

22:01

and so it felt like a

22:03

toy and it was basically ignored,

22:05

except for a few kind of

22:07

neural network researchers in the 80s

22:09

and 90s, and then industry adopted

22:11

it, and it blew up because...

22:13

well, it just works. Who cares?

22:16

Who cares what the theory is,

22:18

which I think happens all the

22:20

time. And I'll stop there. I'm

22:22

curious. Curious if you have anything to add

22:24

to that. Yeah. And it's funny because I

22:26

even like in when I was working on

22:28

this book, you know, with pattern breaker stuff,

22:30

one of the examples I used was the

22:32

Wright brothers with the airplane. And

22:34

so all the experts said it's going

22:37

to take a million years for to

22:39

create a flying contraption. that can fly

22:41

humans in it. The New

22:43

York Times ran an ad

22:45

called Flying Machines That Won't

22:48

Fly, and it

22:50

said that it wasted time to

22:52

try, and they had quote from

22:54

the head of engineering of the

22:56

army and all this stuff. 69

22:59

days later, the Wright brothers at Kitty

23:01

Hawk flew their first plane, and

23:04

there were a couple of bicycle mechanics. And

23:06

so what you see is that the time and

23:09

again, the experts are attached

23:11

to their mental model

23:13

how the world works.

23:15

And it's the tinkerers.

23:17

It's the people who

23:19

have permissionless innovation, who

23:21

just tinker with stuff and make something work.

23:24

And before you know it, they

23:26

have to even change the science.

23:28

People's understanding of Bernoulli's equation and

23:31

all that stuff got modified and

23:33

improved because of the success of

23:35

the Wright brothers with their planes.

23:38

You know, people tend to think

23:40

that the abstract science precedes the

23:42

engineering, but quite often the engineering

23:44

and the tinkering causes the science

23:46

to evolve to explain the unexplainable.

23:49

And yeah, that's what I see happen more

23:51

often in practice. 100%.

23:53

I think the next point that

23:55

you made is sort of like

23:57

this big money versus small team

24:00

thing, which I think happens all

24:02

the time too. Constraints

24:05

breed creativity. And I think, in

24:07

general, being

24:10

able to throw money at a problem means you don't have

24:12

to spend time thinking about how to make it more efficient.

24:15

And so I think your question

24:17

about, like, are the open

24:20

AI's and anthropics of the world in trouble? I

24:22

think that's an interesting one. I

24:24

would bet not. Right. I would, too. My

24:26

feeling about that is,

24:30

Uh, I mean, obviously the sort of cliche

24:32

thing is like, okay, it's going to stimulate

24:34

demand or whatever, which is fine. I think

24:36

that's true. I think that they'll be able

24:39

to integrate most like integrate this and having

24:41

more efficient servers and that can, that can

24:43

serve the, uh, the demand that they currently

24:45

have. I think, uh, I

24:47

think it's, I think we'll work. Um,

24:49

the thing that it seems like to me

24:52

that this opens up is I think we

24:54

have like mass AI figured out, which is

24:56

like, how do you scale these models up

24:58

so that like, a billion people can use

25:01

chat GBT. And how

25:03

do you make that efficient and smart enough

25:05

to work and all that kind of stuff?

25:08

But I think one thing

25:10

that people don't talk about

25:12

nearly enough is that the

25:14

capabilities of models today are

25:16

in many ways not limited

25:19

by the intelligence or the

25:21

intelligence of the technology. They're

25:23

limited by the risk profiles

25:25

of the companies that are

25:27

serving them. And if

25:29

you're a gigantic company, you're open AI and you're

25:32

literally like, you know, you have to go give

25:34

government briefings before you launch anything. You're

25:37

going to be pretty like careful about

25:39

what you, what you put out. And

25:41

I think the deep seek stuff is

25:43

interesting because it means that, and I

25:45

mean risk in all sorts of different

25:47

ways. There's lots of different ways to

25:49

take risk, but it means that small

25:51

teams can build little models for like

25:54

Problems that look like toys that you know

25:56

an open AI would be like well, we

25:58

wouldn't do this and I think that is

26:00

the like that's the big thing I don't

26:02

think that that takes away chat GBT, but

26:05

it does mean that it we we have

26:07

way more AI in different corners of the

26:09

world than we would have otherwise which I

26:11

think is not good Yeah, you know what

26:13

Dan one of my favorite examples of this

26:16

actually comes from the field of rocketry because

26:18

it's it's so visceral so like you know

26:20

Elon Musk he'll launch a Starship if it

26:22

blows up He's like, okay, well, we instrumented

26:24

it, we got telemetry, we'll make it better

26:27

next time. NASA's not going to do

26:29

that. If NASA launches

26:31

a rocket, they don't sit there and

26:33

say, easy come, easy go, it blew

26:35

up. And so the

26:37

fact that Elon has a different

26:40

risk profile and is not attached

26:42

to whether it's successful or capital

26:44

S, it changes the

26:46

calculus of what he can do. And it

26:48

changes the speed with which he can move.

26:51

And so, like I like to say that

26:53

in many cases, the big company is not

26:55

quote unquote dumb compared to the small company.

26:57

It's to your point, they have a different

26:59

risk profile. And they can't,

27:01

they're just certain things they can't do. Like when

27:03

I was working with the guys at Justin TV,

27:05

which became Twitch, if

27:08

they launched something that's insecure, so

27:10

what? Nobody knew who they were. But

27:12

Google can't do that, right? And

27:14

Microsoft can't do that. And the big

27:17

companies can't do that. Hollywood guys can't

27:19

do that. Netflix can't do that. So

27:21

not having to be burdened by what

27:23

could go wrong is a big factor

27:25

in trying things that could go right.

27:27

That makes total sense. I want to

27:29

go back to something that you were

27:31

talking about earlier, talking about

27:33

this company Applied Intuition, which

27:35

she said sells into large car manufacturers. And

27:37

I assume when a large car manufacturer buys

27:40

them, it goes into a Ford vehicle and

27:42

a customer is maybe using it and maybe

27:44

has no idea what it is, but they're

27:46

using it. Is that sort of how it

27:48

works? I think so. It is less of

27:50

an end user type of thing, although that

27:52

might change. I need to be careful what

27:54

I'd say. But the

27:57

primary customer is the car company that

27:59

says, oh my god, the architecture of

28:01

cars has changed. What do I do?

28:04

Yeah. So the strategy question

28:06

I want to ask you is how you

28:08

think about OEM relationships like that. Because I

28:10

think that's going to be a common thing

28:12

for a lot of AI companies, especially if

28:14

you're working on more foundational model type things,

28:16

is you're going to be integrated into something

28:18

else that has a consumer layer. And

28:21

that's where opening i started and then they

28:24

were like actually we want to own the

28:26

ux layer because that's how everything took off

28:28

as they figured out a form factor that

28:30

worked and then they have a data flywheel

28:32

there's all this stuff right. And my last

28:34

company was it was an oam and. That

28:36

is a difficult position when you're serving to

28:38

customers. There's like an end user and then

28:40

there's customer you need to sell to. It's

28:42

hard to generate a lot of power or strategic

28:45

advantage in that situation. And it's

28:47

hard to make a great product. And

28:49

I'm curious how you think about OEM

28:51

type strategies and when they work versus

28:53

when they don't. Yeah, it's tricky. What

28:57

are some examples of where it's worked? I'd

28:59

say applied is working really well. Intel?

29:01

It's been great. Intel was a

29:04

good one for the PCs. Another

29:07

good example would be Qualcomm

29:09

back in the day, with

29:12

licensing their spread spectrum technologies

29:14

and chips. And

29:16

so it can work. Broadcom would

29:19

be another Twilio, I guess. Twilio

29:22

is an interesting one. I like that. In

29:24

fact, I like thinking of Twilio as a

29:26

design -win business more than a dev tool.

29:28

I like that framing a lot, actually. The

29:31

term I like to use to

29:34

describe it is a design -win

29:36

model, where you want to become

29:38

viewed by the customer as integral

29:41

to their product strategy. If

29:43

they have a slide that shows all

29:45

these blocks and triangles and arrows and

29:48

stuff, you need to be a big

29:50

square. in that slide, what you provide.

29:54

Sometimes, like Twilio, you

29:56

solve a problem that they really have, but that

29:58

they just have no interest in solving on their

30:01

own. So if you're Uber, do

30:03

you really want to have an entire

30:05

team building a messaging update texting platform

30:07

that's a substitute for Twilio? Probably none

30:10

of your best developers want to do

30:12

that inside of Uber. And so you're

30:14

like, hey, I'll just pay. I'll

30:17

pay Twilio every time the earth turns a click

30:19

or I send a message. I'll send them

30:21

tiny fractions of a penny. That's okay. So

30:24

that can work. The other way

30:26

I think it can work is

30:28

if you solve something existential for

30:30

the customer. So like in the

30:32

case of the car companies. The

30:34

end customer or the customer the

30:36

customer you're selling to? For the

30:38

OEM actually. So like the problem

30:40

that the car companies have is

30:42

that the Tesla is just a

30:44

fundamentally different architecture than ice vehicles

30:47

right and it's not just it's

30:49

got a battery and they don't

30:51

it's it has to do with

30:53

how many what what their operating

30:55

system is like and how many

30:57

chips they have and how how

30:59

messages flow throughout their messaging bus

31:01

and like Tesla is designed the

31:03

way a car would be designed

31:05

by Silicon Valley type of thinkers

31:07

whereas the you know the ice

31:09

vehicles of today are mostly you

31:11

know an amalgam of a bajillion

31:13

parts suppliers that they've done business

31:15

with for very long time. And

31:18

it's kind of like whatever Bosch has

31:20

this year is going to be the

31:22

new windshield wiper sensor thingy that I

31:24

put in the Mercedes, right? And that's

31:26

how they've operated. So they

31:29

look at it and they're just like, look,

31:31

it's just a completely different paradigm of how

31:33

you'd build a car. And so

31:35

you need somebody that can be your

31:37

thought partner in how to build those

31:39

things. And so that can be

31:42

another kind of design win model that works.

31:45

That's interesting. Yeah. Hey

31:47

there, Dan here. I wanted

31:49

to take a one minute break from the episode to

31:51

tell you about our latest sponsor. All

31:53

right, let's play a game. What powerhouse

31:55

productivity tool is also free

31:58

for individuals? Not

32:02

that one. Try again.

32:07

You may not expect this, but

32:09

it's Microsoft Teams. Yep, the same

32:11

teams that Big Enterprise is swear

32:13

by also has a free plan

32:15

for individuals. Whether you're jamming on

32:17

a side project or bootstrapping a

32:19

startup or building a community, Teams

32:22

has all of the features that other platforms,

32:24

Nickel and Diny for using. You

32:26

can get unlimited chat, 60 -minute video

32:28

meetings, file sharing, and collaborative workspaces all

32:31

for free. And the real

32:33

magic is that everything is integrated

32:35

in one seamless collaborative workspace. That

32:37

means there's no more hopping between

32:39

different applications for messages, meetings, and

32:41

file sharing. Teams puts it all

32:43

at your fingertips to save you

32:45

time and money. So

32:47

ditch the app overload and the

32:50

subscription fatigue and use Teams to

32:52

experience effortless collaboration today. Are you

32:54

ready to streamline your workflow? Head

32:57

to aka .ms slash every to use Teams

32:59

for free. Your productivity will thank you and

33:01

so will your wallet. And now back to

33:03

the episode. First of all, the thing that

33:06

makes me think of is like there's this

33:08

like knife's edge, which is interesting of this

33:10

strategy, which is you have to be Um,

33:13

critical to their business, but somehow they

33:15

don't want to do it themselves, which

33:17

is like, there's very few things that

33:19

are like that. That's right. Um,

33:21

that, and that's really hard. Either you're critical and

33:24

they're like, maybe we'll work with you, but then

33:26

we'll buy you or we'll just replace you or

33:28

you're not critical. And then it's, it's horrible to

33:30

like try to sell that product. No one wants

33:32

to do that. That's, that's, and I love that

33:35

framing of it. I haven't quite. internalize

33:38

it that way, but you're right. They either don't

33:40

want to do it themselves because they just don't

33:42

want to, or they don't want to do it

33:44

themselves because they can't conceive of how they would.

33:48

And they're just like, even if I want

33:51

to, that's kind of academic. I can't. But

33:53

in both cases, it's something that they actively

33:55

choose not to do themselves. And

33:57

there's a persistent reason for that to

33:59

continue. Yeah. And I guess the

34:02

reason You know, like

34:04

an applied intuition would work is I'm

34:06

just, I'm thinking back to you mentioned

34:08

Clay Christensen. I'm sort of like thinking

34:10

about his conservation of attractive profits. Where

34:14

in the early days of new

34:16

technologies, you want one company to

34:18

integrate all the different steps of

34:20

the value chain, basically, because you

34:23

can iterate much quicker. So like

34:25

Tesla. They don't have this

34:27

huge web of different suppliers. They probably

34:29

have a few, but like a lot

34:31

of it, they're just doing themselves. Whereas

34:33

it sounds like, you know, like GM

34:35

or whatever has like thousands of different

34:37

modular manufacturers that they swap in and

34:39

out because like the architecture of the

34:41

car has been around for so long

34:43

that it's not changing. And so it

34:45

doesn't have to be integrated. It can

34:48

just be like, it can be very

34:50

modular, which I guess is a easier

34:52

OEM cell. like because applied can just

34:54

as long as they know that architecture,

34:56

they can sell into it versus like

34:58

a more vertical, more, more integrated company.

35:00

Yeah. Well, and here's how I internalize

35:02

that, Dan. So just to make sure

35:04

that we're on the same page or

35:06

the same language, like what, what I

35:08

understood from Clay, I've kind of got

35:10

a little bit of a crush, an

35:12

intellectual crush on Clay Christensen. I think

35:14

the guy was amazing and a great

35:16

human being. So, so what I understood

35:18

him to say is that In early

35:21

markets, the products are

35:23

never quite good enough. They

35:25

don't perform well enough. And so

35:27

what happens is vendors get rewarded

35:29

for having the integrated system because

35:31

the customers will pay incremental dollars

35:33

for incrementally better performance because they

35:35

value that enhanced performance. But

35:39

then what eventually happens is

35:41

the performance gets mostly good

35:43

enough. and and you know

35:45

what clay christians would call it is

35:47

overshot customers you know now i'm trying

35:49

to cram new features into my product

35:51

to get customers keep buying new things

35:53

that i sell them but now they

35:55

don't want the new things as bad

35:57

and therefore you you get this modularity

35:59

argument that somebody else shows up and

36:01

says look you're being overcharged you don't

36:04

have to have one guy be the

36:06

system integrator anymore in fact you can

36:08

just have a whole bunch of different

36:10

components that you can mix and match

36:12

and swap in and out And so

36:14

then the conservation of attractive profits, it

36:16

goes to the modular suppliers rather than

36:18

the integrated supplier, which I

36:20

think is happening. That was a

36:22

much better summary of conservation of

36:24

attractive profits than I gave you.

36:27

Well, I don't know. But that's the

36:29

brilliance of Tesla. Elon...

36:32

told him you should act like

36:34

a car company acts. You should

36:36

have modular components and suppliers in

36:38

the supply chain. Elon understood, no,

36:41

nobody can make an electric car that's good enough.

36:43

I have to control all the critical technologies because

36:45

I have to have the ability to have something

36:47

that rises to the level of good enough. Nobody's

36:50

ever had that before. So

36:52

like that's another reason, right? Architecturally,

36:55

he's just totally different, right? His

36:57

whole paradigm of how to build

36:59

a car is just different from

37:01

start to finish. So is that

37:03

an argument for AI companies like

37:05

owning the whole stack themselves right

37:07

now as they're sort of innovating

37:09

on what the products even look

37:11

like and customers are willing to

37:13

pay more for incremental value? Yeah,

37:16

what I like about what Clay used

37:19

to say was that what Clay Christensen

37:21

really had was a bunch of mental

37:23

models for innovators. And,

37:25

you know, whenever I think of a mental

37:27

model, I always like to ask under what

37:29

conditions? So under what conditions would

37:31

I want to be the complete integrated solution?

37:34

I believe that you want to be

37:37

the complete integrated solution if the customers

37:39

are desperate for more performance and will

37:41

pay for that enhanced performance, right? So

37:44

like before Nvidia, there were Silicon Graphics.

37:46

And like if you wanted to make

37:48

dinosaurs in Jurassic Park, you had to

37:51

buy the most expensive SGI machines, millions

37:53

of dollars worth. And if

37:55

you could make the graphics run twice

37:57

as fast, industrial light magic would pay

37:59

twice as much because it was mission

38:01

critical to render those dinosaurs overnight. But

38:05

now that there's chips

38:07

commoditized, NVIDIA has the

38:09

better model because they say, hey, I'll just

38:11

sell you off the shelf, these GPUs. So

38:14

I think that the question always becomes

38:16

under what conditions are you advantaged by

38:18

being the integrated solution? And under what

38:20

conditions are you advantaged by being a

38:22

modular component of the solution? That's

38:25

interesting. And I guess what's your best guess about

38:27

where we are now in the, in the

38:29

AI landscape overall? Cause I think that there's a

38:31

lot of, there is this common thing. And I

38:34

actually felt this too, like when 01 came out

38:36

where people were like, I feel like my model

38:38

is pretty much good enough. Like, I don't know

38:40

what I would use 01 or 03 for even

38:43

in the demos. Like they, I remember the,

38:45

one of the demos was like, list out the

38:47

13 Roman emperors, you know, and it's like, that's

38:49

not really something that I. I care that

38:51

much about generally, and most people are not doing

38:54

PhD level research. That was my first feeling, but

38:56

to be honest, now I just use O1 all

38:58

the time, and I don't really use any

39:00

other model, or now I use O3. So

39:03

I'm curious what you feel about where we are,

39:06

and how much performance improvements in terms of intelligence

39:08

people are willing to pay for. Well,

39:11

first of all, I'm

39:13

really excited, but I'm probably

39:16

in these tools too much

39:18

now. So I'm probably in

39:20

these tools three, four, five

39:22

hours a day. And

39:25

there's a lot of things that

39:27

I would benefit from in terms

39:30

of enhanced performance. And if that's

39:32

just me, I gotta believe

39:34

there's a lot of other people like that too. So

39:38

the thing that I think

39:40

is so interesting about AI

39:42

is it really rewards the

39:45

system thinker. And so

39:47

I'll give you an example. Uh,

39:49

you know, I have this database of

39:51

what I call a hundred bagger startups

39:53

and I try to understand them all.

39:55

I've got like, I've got the original

39:57

pitch deck for air bed and breakfast

39:59

for it was Airbnb and I've got

40:01

it for Dropbox and Pinterest and all

40:03

these companies, right? And I,

40:06

I track, you know, if you'd bought a

40:08

share in the seed round, what would have

40:10

happened? I run the inflection theory against it.

40:12

I run insights. I try to understand if

40:14

our frameworks would have caused us to decide.

40:17

Well, now that I have that list, I

40:19

could do all kinds of things. Like I

40:21

can say, okay, please consider

40:24

this list of 100 -Bagger startups,

40:27

which of Hamilton -Helmer's seven powers

40:29

were harnessed by each of them

40:32

as their primary power? Which

40:34

Clay Christiansen jobs to be done was the primary

40:36

job that they did to get product market fit?

40:38

How long did it take them to get 10

40:40

million in revenue? How long did it take them

40:43

to get 100 million in revenue? which

40:45

of them had the first time founder

40:47

or CEO, which of them replaced their

40:50

CEO? You know, I mean, if you're

40:52

curious, it's like, it's

40:54

like having an unlimited supply of smart

40:56

people to go do that research for

40:58

you. It's incredible. I feel

41:00

the same way. I can read and think about so

41:02

many more things than I would have been able to

41:04

previously. And it makes it such a pleasure to get

41:06

up every day. It's the best. It's unbelievable. It's just,

41:09

it's a miracle, right? Like, I just wish I was

41:11

in my early twenties again. I'd be,

41:13

I'd be, I'd be dangerous. Me too. Um,

41:16

uh, well, I guess that, that just makes

41:19

me think like why a hundred bagger startups,

41:21

why not a hundred bagger founders, right? Like

41:23

how much is really in the Airbnb deck?

41:25

That's actually that useful. Yeah.

41:27

So I've been, I've been working on that

41:29

question a lot. And so, um, I've

41:32

been, um, applying our

41:34

frameworks and backtesting them to prior

41:36

startups. So I have these things

41:38

that I call atomic eggs and

41:40

we'll probably launch them here pretty

41:42

soon. But, um, What an atomic

41:44

egg lets you do is it

41:47

lets you upload a pitch and

41:49

then it runs a whole bunch

41:51

of different generative models against it.

41:53

So an example would be pattern

41:55

breakers insight stress test. So you

41:57

could upload the Airbnb pitch deck

41:59

and it would spit out, this

42:01

was the fundamental air insight with

42:04

Airbnb or this is like the

42:06

part that was non -consensus or

42:08

these are the inflections that Airbnb

42:10

is harnessing. The

42:13

AI has gotten really good at that. And

42:15

then the other thing that it can do,

42:17

I like the Sequoia Arc framework. They

42:20

talk about, is this idea a hair

42:22

on fire problem type? Is it a

42:25

known problem type or is it a

42:27

future vision problem type? You

42:29

can run that against a hundred bagger startups.

42:32

And then I could say a scale of one

42:34

to 10, how confident are you that that's the

42:36

right way to classify it? And then

42:38

back to your point about founders. You

42:40

can start to say, OK, there's all these founders.

42:43

What jumps out at you as anomalies

42:45

about these founders? What jumps out at

42:48

you as commonalities about these founders? OK,

42:50

now let's group these startups in different clusters

42:53

and run the same experiment again. And

42:55

then once you get some patterns, you say, OK, how

42:57

might those patterns shift in World of AI? How might

42:59

they be the same in World of AI? Like

43:02

you could have just wondered about that as

43:04

you walk down the street in the past,

43:06

but like now you can act on that,

43:08

right? You can act on that curiosity in

43:10

real time. And that's just like, just

43:13

such a game changer, right? If you're

43:15

curious about this stuff. How

43:17

much does it like, because I

43:19

mean, you write a lot about pattern breakers, right?

43:21

So like, I

43:24

guess I'm thinking about

43:26

business theories or strategy

43:29

theories as Pattern

43:32

patterns, right? There are always patterns that work

43:34

under certain conditions. Sometimes they're like more general

43:36

than others, but like they're usually not like

43:38

infinitely general. I don't know what the what

43:41

the right word is perfectly general. And

43:44

like I wonder, you know, for example,

43:47

if you took the let's say let's

43:49

say we like wound back the clock.

43:51

We went back to like the 80s

43:53

and we used all of the like

43:55

frameworks they had in the 80s and

43:57

put them into AI and like gave

43:59

gave them you know, Cisco or whatever,

44:02

pick whatever company you want, Google, like

44:05

would it have been able to

44:07

tell the Google or the Airbnb

44:09

pitch deck that it was a

44:11

good company? I

44:13

don't know that it could have predicted that it

44:15

was going to have the success it had. Yeah.

44:18

And I apply a slightly less stringent

44:20

standard. What I really want to know

44:22

is, should I spend time on this?

44:25

Right. And so I needed what I what I need

44:27

to know when I look at a pitch like Airbnb

44:29

is Is

44:31

there something that's wacky and good about

44:34

this that I might overlook if I'm

44:36

busy and tired that day? But

44:38

like if I can run a whole bunch

44:40

of different tests against it. So like, you

44:42

know, you talked earlier about these models, like

44:45

Charlie Munger is somebody else who I've always

44:47

respected. And, you know, he had this saying

44:49

the map is not the territory. And what

44:51

he meant by that is that, you know,

44:54

if you and I want to go from

44:56

San Francisco to Cupertino and we use a

44:58

flat map, and let's say we use Google

45:01

Maps or whatever, the odds

45:03

that we will get there if

45:05

we follow the directions are basically

45:07

100%, like 99%, in fact, I

45:09

would argue that that map is

45:11

a better representation of reality than

45:14

all the complexities of all reality.

45:17

You're trying to compress knowledge for

45:19

the decision that matters, right? But

45:21

if you and I want to

45:23

go to Germany, the

45:25

map is not gonna be an accurate

45:28

portrayal of the territory because straight line

45:30

is not the shortest path on a

45:32

flat map that represents a globe, right?

45:34

It would look like a curved line.

45:37

And so like what you learn is that

45:39

it's like we talked about earlier, the

45:42

question is under what conditions is this

45:44

model useful and under what conditions are

45:46

the boundary conditions exceeded? And that's

45:49

why you wanna have a whole bunch of them, right?

45:51

You wanna have the right tool for the right situation.

45:53

And then when it exceeds the scope of

45:56

the boundaries, you want to not use that

45:58

tool because you'll get bad, bad decision making.

46:00

Are there any new things? Because like one

46:02

of the things you talk about in your

46:05

book a lot that I like, I like

46:07

because this is sort of how I work.

46:09

So it's maybe confirmation bias, but I like

46:11

a lot is sort of the idea of

46:13

living in the future, right? Like the best

46:15

way to know what's coming is to just

46:18

be like you're doing in these tools all

46:20

day, every day. And you start to kind

46:22

of like see things that your other people

46:24

maybe won't see because they're just they're living

46:26

in a different reality and your reality is

46:29

going to sort of spread everywhere else eventually

46:31

is the idea. I'm curious if

46:33

there's anything like that that you're feeling

46:35

and seeing right now that you're kind

46:37

of like sensitive to that is new

46:39

and interesting to you. Yeah,

46:42

you know, some of these AI companies

46:44

you'll go to and there'll be somebody

46:46

who's a couple years out of college.

46:48

And they'll be using Devon or

46:50

cursors of these other products and

46:52

they're kind of creating these agentic

46:55

oriented entities that go out and

46:57

get a bunch of stuff for

46:59

them and bring it back. And

47:02

they just almost act like that's normal. So

47:04

they're almost like programming these virtual employees to

47:06

go out and do stuff for them. And

47:08

you'll sit with them and you'll say, well,

47:10

what motivated you to do that and to

47:12

think about solving the problem that way? And

47:15

they look at you funny like, well, how

47:17

else would you do it? You

47:19

want me to Google? Yeah. And

47:21

so the thing that I find

47:24

interesting is, you

47:26

know, and this is like how Zuckerberg was

47:28

with social networking, right? Like, Zuck

47:30

didn't have to unlearn anything. You know,

47:32

he grew up at a time when

47:35

the lamp stack was coming out and

47:37

you could A .B. test things and

47:39

the broadband was everywhere. Before

47:42

Facebook, you know, like in the 90s,

47:44

You had to have products that were

47:46

well engineered because they just weren't scalable

47:49

enough otherwise, right? You had to have

47:51

experts that would architect and instrument the

47:53

system so that it would be somewhat

47:55

performant. Well, by the time Facebook comes

47:57

around, Zuck's like, hey, well, we

48:00

just try it and see what happens by the

48:02

afternoon and decide whether we want to keep with

48:04

this or not. Now, did

48:06

Zuck say? Aha, there's a disruptive trend and

48:08

I'm going to get a leapfrog all these

48:10

companies. No, like Zuckerberg didn't know anything about

48:12

business at the time. It's almost like it's

48:14

like if you and I were raised in

48:16

a world of Cartesian coordinates and now it's

48:18

a world of polar coordinates and somebody's born

48:20

in a world of polar coordinates and they

48:22

don't even have to translate between the two.

48:24

They're like, what else is there? That's the

48:26

only thing there is. I think that some

48:29

of these AI natives are like that. And

48:31

so I really want to spend time with

48:33

them. I want to spend time with anybody

48:35

who says, My entire lived

48:37

experience in business is a world where

48:39

you're programming some form of AI assistance

48:41

as a core function of the job.

48:44

I love that. I mean, I see

48:46

this all the time. We have a

48:48

writer who started working with us probably,

48:51

I would say two months ago. He's

48:53

had a very successful career, not as

48:56

a professional writer, just working in AI

48:58

at various tech companies and startups and

49:00

has founded his own startups. But

49:03

he's working for us mostly as a writer. And

49:07

he writes our Sunday email where we talk

49:09

about all the new model releases. He's such

49:12

a nerd for new stuff that comes out,

49:14

which is amazing. That's the kind of person

49:16

you want writing. And

49:19

he also, when a new

49:21

model comes out, I'll often get early access.

49:23

So we'll get on the phone together. He'll

49:26

write a first take of all the things that

49:28

we saw. And then I'll go through and put

49:30

my own take on it and whatever. So we

49:32

co -write things together. And

49:35

the first one that he did it like that,

49:37

I got the draft and I was like, he's

49:40

smart. He's excited about this stuff, but he's

49:42

not a professional writer, I can tell. It

49:44

wasn't something that I just punch up and

49:46

I can just publish. It was like I

49:49

had to rewrite the whole thing. And

49:52

what was crazy is after we did that,

49:54

I was just like, okay, I want you

49:57

to take my draft and then your draft

49:59

and I want you to put it into

50:01

01 and pull out what changed. And

50:04

he did that. And we did that a couple of

50:06

times. And we just

50:08

covered the launch of DeepSeek

50:10

together. And the first

50:12

draft he did, it was like he made

50:14

a year's worth of progress in a month.

50:16

Like I've seen, I've worked with so many

50:19

writers in my career at this point. And

50:21

I've seen where people are at like when

50:23

I first started working with him, it takes

50:25

them like a thousand drafts to make the

50:27

amount of progress that he made in a

50:30

month. It's crazy. Yes. Yeah,

50:33

it's so interesting, right? And I'm finding the

50:35

same thing, Dan. So like, as

50:37

I started working on these mental models for

50:40

seed and these generative models, I started to

50:42

say to myself, what is a

50:44

good mental model in the first place?

50:46

Like, has anybody ever defined what one

50:48

is? What should it contain? What

50:51

makes it good versus bad? Under what conditions

50:53

is it good or bad? And

50:55

there wasn't a whole lot about it. You know,

50:57

there's a couple of books on mental models, but

50:59

not a whole lot. So

51:01

I said, you know, before I start

51:03

just saying, here's a mental model, jobs

51:06

to be done, I should create a

51:08

foundational document that's the taxonomy of a

51:10

good mental model and the questions it

51:12

should answer and the flow that it

51:14

should take. So I did that. Now

51:17

I can just say, I'm

51:19

going to I'm just going to write about jobs we done

51:21

for what it is. And then

51:23

I run it against this framework and it

51:25

says you're missing A, B and C. And

51:28

I'm like, hey, well, can you elaborate on

51:30

that? And it just, it just adds it,

51:32

you know, and, and, you know, within 30

51:34

minutes, you have something that's just off the

51:36

hook, right? It's just so good. That's great.

51:39

And it's like, you just look at that

51:41

and you're just like, it

51:43

just feels like magic. It feels like

51:45

you put on some cape and just

51:47

learned how to fly all of a

51:50

sudden, you know, and I'm just like,

51:52

but like, it goes back to reward

51:54

system level thinking, right? You had to,

51:56

you had to zoom out and say,

51:58

wait, you know, If i'm gonna someday

52:00

have a hundred mental models i ought

52:02

to define a connect canonical baseline good

52:04

one. And i ought to have

52:06

a theory about what makes it good and i ought

52:09

to apply that theory to everyone that i do because

52:11

i'm gonna get leverage if i do that but now

52:13

i'm gonna make the i do the work for me.

52:16

And it teaches you stuff right like now you

52:18

say oh i thought i knew jobs we don't

52:20

was a mental model but there are. boundary conditions

52:22

I hadn't thought about before that are kind of

52:24

interesting. And so, yeah, it's, you

52:26

know, it's just such a great time to be

52:29

alive with this stuff. I agree. I

52:31

want to go back to the question,

52:33

the original question I asked you because

52:35

it's still on my mind, which is

52:37

software is getting so much cheaper to

52:40

make. The VC

52:42

model, even the seed model,

52:44

which you pioneered is Predicated

52:47

on a different world where it was

52:49

expensive to make software at first and

52:51

then it was free to distribute and

52:53

And I'm curious how you think that

52:56

that might change the VC model if

52:58

at all and I'll preface this by

53:00

saying this is a selfish selfish question

53:02

because I run every we've got I

53:05

can't even it's like I don't really

53:07

have words for the kind of company

53:09

We are we have a newsletter with

53:11

a hundred thousand subscribers and then we

53:14

have three different software products and we're

53:16

10 people. It's like a whole different

53:18

thing. Yeah. And

53:20

I use that sparkle thing, by the way.

53:22

It's cool. You do? Oh, I love that.

53:24

That's great. Love to hear

53:27

that. And I

53:29

feel like I want a

53:31

different funding model. And

53:34

I'm working through different options, but I'm kind of

53:36

curious how you think that that might change. Yeah.

53:38

So I've been thinking about it a lot.

53:41

So there's two different angles, and there's the

53:43

There's the angle that you're describing and

53:46

another person that I respect who thinks

53:48

a lot about this is a guy

53:50

named Greg Eisenberg right on on Twitter

53:52

and so like Let me see if

53:55

I can capture what I think it

53:57

is. It's that You you have a

53:59

situation where what it takes to build

54:01

a product has collapsed yet again Just

54:04

like it did with a lamp stack

54:06

and it's it's profound in a lot

54:08

of ways It's not just that it

54:10

costs you less money to build a

54:13

product, but like You had the chat

54:15

PRD on a few episodes ago. Chat

54:18

PRD lets one person have the entire

54:20

idea premise of the product in their

54:22

own mind and doesn't require them to

54:25

therefore have a giant team of other

54:27

people. So it changes the

54:29

dynamics of who can build software and

54:31

what it takes to build it. And

54:34

so you start to say, OK, well, Are

54:36

you gonna have these tiny little companies that

54:39

generate a ton of revenue and they don't

54:41

even have to generate that much to be

54:43

wildly efficient and profitable? Why

54:45

would you need VC money at all? And

54:48

I'm pretty sympathetic to that point of

54:50

view, although I tend to go to

54:52

the founders and say, look, I'm not

54:54

under pressure to put a lot of

54:56

money into you. Our funds are small

54:58

and all things being equal, I'd rather

55:00

have it be one and done and

55:02

we try a few things. Here's

55:05

the other thing, though, that I

55:07

think is really interesting, that I'm

55:09

trying to find kindred spirits around.

55:12

The lamp stack didn't just

55:14

collapse the cost of startups.

55:17

It created a new way of building. It

55:20

created a new model of building, right? So

55:22

you used to have waterfall development, and

55:25

you had to define everything that's in the

55:27

release upfront, and then you go on a

55:29

death march for a year, and you ship

55:31

it and it either succeeds or it bombs.

55:33

And that was just how products were. And then the

55:35

lamp stack comes out and you have lean startups and

55:38

agile. And what I'm

55:40

seeing happen now, and I'm

55:42

not sure what to call it. So

55:44

right now I'm calling it

55:47

Darwinian engineering or digital Darwinism.

55:49

So like if you think

55:51

about it like in an

55:53

ecosystem, you

55:55

don't have the individual elements

55:58

and players in the ecosystem

56:00

be programmed in a

56:03

literal way. What you

56:05

have is a system designer, if you will.

56:08

And then the system gets to

56:10

operate autonomously from the designer. And

56:13

so I sit there and I think, man, that

56:16

kind of rhymes for me. So

56:18

I think about it like

56:20

it's like natural evolution rather

56:22

than traditional development. And that

56:25

you're going to have AI

56:27

tools that shift from

56:29

agile to continuous adaptation. And

56:32

you're going to build software elements

56:34

and components that are adaptive by

56:36

design and that can sense and

56:38

respond to the inputs that they

56:40

get in the real world independent

56:43

of the program. So rather than

56:45

have a business model canvas, you

56:47

have a business model dashboard that's

56:49

live status of what's happening. And

56:52

so, you know, if you're a gaming

56:54

company, you're going to shift from iterating

56:56

games to creating living worlds. you know,

56:59

and you know that kind of stuff. So

57:01

I'm really interested in like what does that

57:03

mean for what a product manager is? What

57:06

does that mean for dashboards of the

57:08

future? What does it mean for how

57:10

QA happens? You

57:13

know, all that stuff. I thought about this

57:15

too a lot because I think we actually

57:17

met originally because you read my article on

57:19

the allocation economy. Yes. And I sort of

57:21

started to think a lot about like what

57:23

is the role of someone who's working in

57:25

the allocation economy and how is that different

57:27

from someone in an analogy economy? And

57:30

a way that I've been thinking about

57:33

it is in an analogy economy or

57:35

just any previous economy, the

57:38

work you're doing, especially as an

57:40

IC, a little bit more,

57:42

still a little bit as a middle

57:44

manager or an executive, but like a

57:46

lot of this is as an IC

57:48

is you're kind of like a sculptor.

57:50

Like everything that happens happens because you

57:52

did it with your hands. You

57:54

have your hands on every little piece of it. And

57:58

I think working with AI models is

58:00

a lot more like being a gardener.

58:03

You're like setting the conditions for the thing

58:05

to grow, and then it just sort of

58:07

grows. And the conditions are like hyperparameters. It's

58:10

like the sun and the soil and the

58:12

water and whatever, and that's going to change

58:14

what comes out. And you know, like opening

58:16

an eye like doesn't, when

58:19

it, when Chagabee responds to a prompt, like no

58:21

one, at OpenAI, like, decided that it was going

58:23

to say that, which is

58:25

totally different from Facebook or

58:28

whatever. Like, someone decided what

58:30

you were going to see

58:32

on Facebook. Or maybe,

58:34

if Facebook's maybe a little bit, they have AI

58:36

too. But like, let's just say, the New York

58:38

Times, someone decided what's on the homepage. And

58:41

it's totally different. And you're

58:44

right. You

58:46

can tune stuff, but it's

58:49

like... It's much squishier because

58:51

you're kind of tuning the

58:53

like environmental conditions rather than

58:55

the specific thing that happens

58:57

and Yeah, I think that's

58:59

such it's such a different

59:02

way of Working it's such

59:04

a different way of building

59:06

products. I don't think like

59:08

if I think about what we're building at every

59:10

like I don't think we're quite there yet. What

59:12

I see is like I

59:16

mean obviously like building an organization you are

59:18

kind of like doing that but like for

59:20

individuals who are building products like. One

59:23

of the things I see is like it's so easy to build a

59:25

feature you can just build it in an hour so it's like. Sometimes

59:28

you just build a lot of features and you're like

59:30

others it's kind of now the products kind of noisy

59:32

it's kind of messy you know. And

59:35

also, it's like the hard thing is

59:38

figuring out what to build, not actually

59:40

building it, which is a different thing.

59:42

But we're not yet in a world

59:44

where it's fully adaptive. But I do

59:46

think you're right. You

59:49

can see that with Chatchabee to Canvas or

59:51

Artifacts or whatever, where it's starting to build

59:53

its own UI and stuff. And I think

59:55

that's where we're going. Yeah.

59:58

And it's just interesting, right? Because it

1:00:00

kind of goes back to systems -level

1:00:02

thinking. It's one thing to

1:00:04

think of yourself as building components

1:00:06

or building tools or building the

1:00:08

end thing. It's another thing to

1:00:10

say, I'm building an

1:00:13

ecosystem and the elements of the

1:00:15

ecosystem operate under certain first principles.

1:00:19

But there's a lot of emergent properties

1:00:21

that are going to occur in that

1:00:23

ecosystem that are a function of the

1:00:25

dynamism of the system and how it

1:00:27

interacts with people. I think

1:00:30

that that's just a fundamentally different world

1:00:32

view about how you architect products. And

1:00:35

so I think that that's another, you know,

1:00:37

there's the what we said earlier, very low

1:00:39

cost, low end disruptive

1:00:41

innovation ideas. But I think

1:00:43

there's also this, hey, the way software

1:00:45

ought to be built in the first

1:00:48

place ideas is interesting as well. Yeah,

1:00:50

it reminds me of like notion, for

1:00:52

example, you know, it's like notion. It

1:00:56

has a block system. It has these atomic

1:00:58

elements that you can build anything with rather

1:01:00

than like they built a specific feature to

1:01:02

do a specific job, which is it's a

1:01:04

different way of thinking about products. It's like

1:01:06

making a language versus like making a hammer.

1:01:09

That's right. That's right. Yeah. Yeah. And so

1:01:11

I think that that's going to be really

1:01:13

interesting. And I think that it, but it's

1:01:15

like, you know, we used my example earlier,

1:01:18

if I want to have mental models for

1:01:20

investing, rather than just jumping straight to it,

1:01:22

what I need to do is I need

1:01:24

to like zoom out a little bit and

1:01:26

say, okay, let me think about this in

1:01:28

a systems level way. What makes

1:01:31

a good mental model in the first

1:01:33

place? Like what, how do I, how

1:01:35

to make sure that I have a

1:01:37

foundation built on something really powerful so

1:01:39

that every subsequent piece of activity or

1:01:42

thinking that I do is a multiplier

1:01:44

effect on what's come before. Totally.

1:01:47

Well, Mike, this is

1:01:50

a pleasure. I

1:01:52

feel like I learned a lot. Me too. I'm really

1:01:54

glad we got the chance to hang out. Thanks for

1:01:56

coming on the show. Yeah, thanks, Dan. It was great

1:01:58

to see you. Oh

1:02:07

my gosh, folks. You absolutely,

1:02:09

positively have to smash that like

1:02:11

button and subscribe to AI and I. Why?

1:02:13

Because this show is the

1:02:15

epitome of awesomeness. It's like finding

1:02:17

a treasure chest in your backyard,

1:02:19

But instead of gold, it's

1:02:22

filled with pure unadulterated knowledge bombs about

1:02:24

chat GPT. Every episode is a

1:02:26

roller coaster of emotions, insights and

1:02:28

laughter that will leave you on the

1:02:30

edge of your seat, craving for more.

1:02:32

It's not just a show, it's a

1:02:34

journey into the future with Dan Shipper as

1:02:36

the captain of the spaceship. So...

1:02:39

yourself of a favor, hit like,

1:02:41

smash subscribe, strap in for

1:02:43

the ride of your life. And now,

1:02:45

without any further ado, let let me just

1:02:47

say, Dan, I'm absolutely hopelessly in love

1:02:49

with you.