0:00

As someone who's built a career around

0:02

writing, I think that AI is going

0:04

to shake the foundations of the writing

0:06

world. I got two big thoughts. Number

0:08

one, if you're a writer and you're

0:10

completely ignoring, completely dismissing AI, I think

0:12

you're out of your mind. Okay, but

0:14

also, there's going to be major room

0:16

for writers to succeed. It's not all

0:18

doom and gloom. Really good writers are

0:20

going to be all right. But the

0:22

question is, how do those two things

0:24

square together? And if you're a writer,

0:27

how should you be thinking about AI?

0:29

Well, that's what this episode's all about. What

0:31

I'm gonna do is I'm gonna show you how

0:33

I write with AI, how I think with AI,

0:35

and how to make your writing AI-proof. I got my

0:37

notes here, and I'm just gonna talk about

0:40

it. These are all the things that I've learned

0:42

over the past few months since AI

0:44

really kind of kicked off that existential

0:46

crisis in me. And I'll just start

0:48

off with a little bit of backstory.

0:51

So I ran a writing school called

0:53

Write a passage for like six years.

0:55

taught thousands of students, some of whom

0:57

became really successful, and the whole point

0:59

and purpose of the school was you'd

1:02

come in, and if you wanted to

1:04

build an online audience, you would start

1:06

writing consistently, publish an article, say

1:08

every week or so, build an email

1:10

newsletter, and then become a domain expert

1:12

for whatever it was that you knew a lot

1:14

about or really passionate about. And

1:16

a lot of people did that,

1:19

and then other people. got really

1:21

excited about telling their personal stories.

1:23

There was some idea, some story

1:25

that they had always wanted

1:27

to tell in their life, and

1:30

Write a Passage became a place

1:32

where they could do that. So, ran that,

1:34

thousands of students, and I

1:36

was really focused on the

1:39

curriculum. And what's funny is,

1:41

we ended Write a Passage

1:43

in November. It's now been less

1:45

than six months, and back in

1:47

November, less than 10% of the

1:49

curriculum. Now I can't even imagine

1:52

in not being the linchpin of

1:54

it, both because of how fast

1:56

things have changed, but also because of

1:58

how much people are. talking about AI

2:00

now in a way that they weren't

2:03

just six months ago. That's how fast

2:05

things are changing. And I still remember,

2:07

you know, I still remember being in

2:09

my apartment, it was late at night,

2:12

I had this old couch that I

2:14

bought on Facebook Marketplace, and somebody had

2:16

recommended an article to me from Stephen

2:19

Johnson, an article in the New York

2:21

Times, and it was about GPT3. So

2:23

this is like a long time ago,

2:25

like this is eons ago, like the

2:28

dinosaur age of AI. And the article

2:30

was about. how LLLMs worked and how

2:32

you could basically predict sentences. And actually

2:34

it's one of the best opening paragraphs

2:37

of any article I've ever read. I

2:39

remember reading it, sitting there, and that

2:41

was like when this first, when the

2:43

light bulb sort of came on from

2:46

me, I was like, you know what,

2:48

maybe there's something here. And it was

2:50

the first time that I began to

2:52

understand sort of at a high level,

2:55

how do LLLMs work? What is going

2:57

on here? How are they predicting? words

2:59

and work at a technological level and

3:02

then I had a mentor who runs

3:04

multi-billion dollar company and his whole goal

3:06

for the company in 2023 was to

3:08

get them to be AI first and

3:11

I would sit there at dinner with

3:13

him and I really admire this guy

3:15

from a business perspective and we'd be

3:17

sitting at dinner to say oh yeah

3:20

AI this, AI that and I was

3:22

like dude I don't know I don't

3:24

know I'm kind of a skeptic, like

3:26

if you actually use these things, they're

3:29

not that good, they hallucinate all the

3:31

time. I don't know, it's just no

3:33

better than what a human can do.

3:36

It's not even close. He's like, you

3:38

gotta watch the rate of growth. And

3:40

I really believe that this, this technology

3:42

is gonna start growing really fast. So

3:45

many have thousands of my employees to

3:47

start. really focusing on AI. He put

3:49

in a hiring freeze so that the

3:51

companies, they couldn't even grow, they had

3:54

to get more efficient with AI. And

3:56

I was like, dude, you're crazy. And

3:58

then, fast forward a bit. to see

4:00

glimmers of it in 2024, but it

4:03

was really at the end of that

4:05

year. I did a trip to Argentina,

4:07

and it was sort of my first

4:10

time taking serious time off in a

4:12

while, and when I travel, I just

4:14

love to learn about wherever I am.

4:16

So what I would do is, as

4:19

I would spend the day in Buenos

4:21

Aires, I would, you know, be at

4:23

an art museum or see a statue,

4:25

and I basically have a question, and

4:28

I put it into GPT, and then

4:30

it... Would get the answer in the

4:32

background and then at night I'd come

4:34

home and I would read all the

4:37

answers of questions that I'd asked throughout

4:39

the day So because of that I

4:41

could basically file ideas into chat GPT

4:43

and then at the end of the

4:46

day have like a whole summary on

4:48

the things that I needed to know

4:50

and Man, it was so cool. I

4:53

feel like I learned so much from

4:55

From doing that and I was like

4:57

who I now have a tour guide

4:59

with me at all times, the AI

5:02

answers were more useful for me from

5:04

a learning perspective than hiring actual tour

5:06

guides for basically all the tour guides

5:08

I hired. I probably hired four or

5:11

five tour guides down there for various

5:13

things. There's only one tour guide that

5:15

was better than what the AI would

5:17

have given me, which is pretty crazy.

5:20

But it's not just that. The models

5:22

themselves are getting better and cheaper at

5:24

a really fast rate. So for starters.

5:27

The amount that I'm engaging with AI

5:29

is up like 10x in the past

5:31

year. But there's also a lot of

5:33

competition for who's going to make the

5:36

best model. And you'll notice because there's

5:38

so much competition, prices are falling and

5:40

people are competing like crazy to basically

5:42

try to have the best model that

5:45

you can go to. I remember It

5:47

was 2023, GPT4 came out, and it

5:49

was so much better than the other

5:51

models. It was like, LaWain has this

5:54

line, I'm resting in the lead, need

5:56

a pillow and a cover. It was

5:58

like they are so far ahead. And

6:01

this was wild. Throughout the year 2020,

6:03

18 different companies came out with a

6:05

model that was as good or better

6:07

than chat gPT4 and actually that's not

6:10

obvious I have the Jordan Peterson me

6:12

that's not obvious you know because here's

6:14

the thing it could have been one

6:16

of those things where one company AGI

6:19

they're just gonna hold the keys to

6:21

the kingdom and actually there's a lot

6:23

of different models that are really good

6:25

but all this is to say that

6:28

if you're interested in writing and you're

6:30

just outright ignoring these advancements, or you're

6:32

just dismissing them, I just think you're

6:34

a fool. I think you're a fool.

6:37

And we are on the precipice of

6:39

a new paradigm of writing. Once again,

6:41

that isn't something to be terrified about.

6:44

I really don't think so. At least

6:46

for writing. There's societal issues. That's another

6:48

conversation. But GPT 4.5 is like routinely

6:50

making me laugh out loud. So I

6:53

want to show you this joke that

6:55

I found. This is a very niche

6:57

example, so hard. and it's a description

6:59

of Tyler Cowan's life. So Tyler Cowan

7:02

came on the show recently, he's an

7:04

economist at George Mason, but he's also

7:06

just like a quirky and funny, very

7:08

distinct guy. And there's this meme, I

7:11

think it's like downstream of 4chan or

7:13

something, but it's like be me and

7:15

you'll see it's like a funny formatting.

7:18

And this guy put in, hey, do

7:20

like a be me joke about Tyler

7:22

Cowan. It is hilarious. And now here's

7:24

the thing, if you're looking at it,

7:27

you don't know Tyler Cowan, you're like

7:29

whatever, but if you know Tyler Cowan

7:31

stuff, this is freaking hilarious, and I

7:33

looked at this, this is right after

7:36

ChatGPT 4.5 came out, and I was

7:38

like, you could have given me a

7:40

week, and I could not have come

7:42

out with something that funny. No way.

7:45

And this is a crucial point. AI

7:47

is gonna be really good at super

7:49

niche humor. So if you're interested in

7:52

some strange and esoteric, an esoteric thing.

7:54

And you just want to get like

7:56

funny takes. AI is going to be

7:58

really good at that. because the chances

8:01

that some comedian is also interested

8:03

in that are going to be pretty

8:05

low. Obviously, this is a one-off example,

8:07

but William Gibson, the science fiction writer,

8:09

has this line that I think about

8:11

all the time where he says, the

8:14

future is already here, it's just not

8:16

evenly distributed. And I think this is how

8:18

the future shows up. You see little glimpses

8:20

and little glimmers of the future. And

8:22

then you just have to think in your

8:24

mind, okay, if that becomes higher resolution, what

8:26

is it going to look like? And I

8:28

think this is a good example of something

8:30

that is way funnier, way better than most

8:32

of the things that I see AI doing,

8:34

which still actually aren't that great, they're

8:36

okay, but like I can see the writing on the

8:39

wall when I look at something like this. And the

8:41

point is, I think it's not just coming for

8:43

humor. I think that what it means

8:45

to be a journalist, a journalist, a

8:48

researcherer, an academic, a full-time author, a

8:50

full-time author, is going to be rewritten,

8:52

It actually already is being rewritten a

8:54

bit. Like Ethan Mallock, who is

8:56

a writer, I think he's at

8:59

the University of Pennsylvania, he said

9:01

that the past 18 months have

9:03

seen the most rapid change in

9:06

human written communication ever. As of

9:08

September 24th, 18% of financial consumer

9:11

complaints, 24% of press releases, 15%

9:13

of job postings, and 14% of

9:15

UN press releases, showed signs of

9:18

UN press releases, showed signs of

9:20

LLM writing. These are just the

9:22

official stats, right? Because that's showed

9:25

signs of LLLM writing. Like the

9:27

number of people who are probably

9:29

using LLLMs, but then you don't

9:31

see it showing up in their

9:33

writing, is probably even higher. We're

9:35

seeing a major change, that's just

9:37

in the past six months. We're

9:39

getting 24% of press releases. But here's

9:41

the thing. Absolutely. Humans are still better

9:43

at writing than AIs in a lot

9:46

of ways. But I think that's about

9:48

to change. I'll just about my own

9:50

life. A full-on half of what

9:52

I read probably is basically

9:55

generated by an AI. So like, what

9:57

am I talking about? What am I

9:59

reading? Well, I have a lot of

10:01

conversations with LLLM. So I'll go back

10:04

and forth all the time, you know,

10:06

if I have a question about how

10:08

something works in the world or I'm

10:10

interested in, for example, last night I

10:13

went out for a wahauken Mexican dinner

10:15

and I love wahauka cheese. So like

10:17

something that I might do today is

10:19

like, teach me about wahauken cheese. What

10:21

creates the texture? Why does it come

10:24

from wahaca? What makes it so distinctive

10:26

and unlike other cheese? It used to

10:28

be that I'd probably Google something like

10:30

that now I just talked to an

10:33

LLLM But the thing that I really

10:35

use LLLMs for is deep research reports

10:37

So if you go on chat GBT

10:39

at the bottom if you're on the

10:41

pro or the professional Or the advance

10:44

of the professional version whatever the two

10:46

paid versions are you'll see that there's

10:48

a little deep research button that you

10:50

can press and those deep research reports

10:53

reports are so Good, I'm generating a

10:55

few of them every single day. And

10:57

here's an example of how I use

10:59

it. What it does is it uses

11:02

a kind of more advanced model under

11:04

the hood called O3, and it'll sort

11:06

of scan the internet for information. And

11:08

the way that I use it is

11:10

I live on Lady Bird Lake in

11:13

Austin, so there's this lake in the

11:15

center of the city, and it's March

11:17

now, so... the weather's getting nicer, the

11:19

leaves are beginning to come back, and

11:22

you can sort of feel the world

11:24

changing as we move into springtime. And

11:26

I have this 20, 25-minute walk between

11:28

my apartment and my office where I

11:30

work. And I want to enjoy that

11:33

walk more. So what I did was

11:35

I said, hey, deep research, I want

11:37

you to basically make me a report

11:39

on the flora and the fauna in

11:42

this area. And I'm particularly interested in

11:44

how the nature around me begins to

11:46

springtime in Austin, in Austin, Texas. Super

11:48

specific. And what it does is it'll

11:50

go off and it'll take a bunch

11:53

of really good pieces of information and

11:55

it'll come back and it'll... deliver me

11:57

a full report that's say two or

11:59

three thousand words that I can read

12:02

that's really tailored for my interests, my

12:04

curiosities at the moment, exactly where I

12:06

live, and took me what, 20 seconds

12:08

to produce the prompt, five minutes to

12:10

basically wait for an answer. And now

12:13

I feel like I know, you know

12:15

what I mean? It's like, it's right

12:17

there, and I don't have to go

12:19

scan through a bunch of Google reports

12:22

or read something that's generic. It's really,

12:24

really tailored to my interests. And look,

12:26

this is the first version of the

12:28

software. Deep Research came out only a

12:31

few months ago, and it's already at

12:33

that level. And you'll notice if you

12:35

use it that the writing itself actually

12:37

has some voice and some style in

12:39

a way that ChatGPT definitely didn't have

12:42

a year and a year and a

12:44

year and a year and I read.

12:46

AI generated. And because of that, I'm

12:48

just reading fewer things that humans have

12:51

made because time is finite and more

12:53

of my reading is going to things

12:55

that computers have written. But I don't

12:57

just want to talk about where things

12:59

are now, I want to talk about

13:02

where things are going. Once again, I

13:04

don't believe that all writers are screwed,

13:06

I don't believe that we shouldn't teach

13:08

writing anymore. Actually, as I was prepping

13:11

for this, you know, I obviously used

13:13

AI a bit, but like the process

13:15

of working on this outline and putting

13:17

my fingers on the keyboard and really

13:19

thinking through things, a lot of ideas

13:22

came to my head and were crystallized

13:24

and clarified in a way that they

13:26

never would have been had I not

13:28

sat down to do the writing. So

13:31

it's still the case that writing improved

13:33

your thinking and I still think that'll

13:35

continue to be true. But here's what

13:37

I do believe. At the number of

13:40

people who can gain an an audience

13:42

and make an audience and make money

13:44

and make money. For their non-fiction writing,

13:46

because they're able to just outperform AI,

13:48

is going to fall. It's going to

13:51

fall considerably. It's going to continue to

13:53

do so. If you want to make

13:55

money as a writer, your writing is

13:57

going to have to get better and

14:00

better. So here's my characteristic for what

14:02

kind of non-fiction writing will last. Like

14:04

if you're focused on that, what should

14:06

you do? The more that a piece

14:08

of writing... comes from personal experience, the

14:11

less likely it is to be overtaken

14:13

by AI. So that's personal writing. That's

14:15

things like biographies. That's things like memoirs.

14:17

They ain't gonna go away any time

14:20

soon. And there's a few reasons why.

14:22

One major thing that we get from

14:24

writing is connection to human to human

14:26

connection. Like love is one of those

14:28

things that just has this infinite ceiling.

14:31

Like if you can really connect with

14:33

somebody and people feel like they're there

14:35

with you, they're in your mind as

14:37

they're reading your stuff. Man, you got

14:40

such a bright future as a writer,

14:42

that's not going to go away. Like

14:44

David Foster Walla said that a major

14:46

reason why we read is to countenance

14:49

loneliness, to have the thought that, wow,

14:51

there's somebody else who feels this way

14:53

like I do. I thought I was

14:55

the only one. We really want to

14:57

hear personal narratives. So if an AI,

15:00

and I'm going to talk about this

15:02

more later, but if an AI can

15:04

help you write your personal narrative, great.

15:06

I'm totally a fan of that. That

15:09

sounds cool. I personally have no issues

15:11

with that. I have no issues with

15:13

that. But no one wants to hear

15:15

a personal narrative from a computer. Like

15:17

that is a completely hollow thing. And

15:20

if you're like a personal narrative from

15:22

a computer like why would you ever

15:24

want to read that I don't care

15:26

how good it is you know I

15:29

completely feel you that that that sounds

15:31

like a completely soulless enterprise okay here

15:33

we are coming from there and this

15:35

really became clear to me when I

15:37

read this testimony of how the co-founder

15:40

of Wikipedia became a Christian a few

15:42

a few weeks ago and I was

15:44

reading it and I was just riveted

15:46

I was connecting with this guy and

15:49

I'm reading this piece about The the

15:51

ways that he changed his mind his

15:53

life arc his his emotional journey it

15:55

is the most Moving thing that I've

15:58

read all year and That is exactly

16:00

the kind of writing that just isn't

16:02

gonna go away with AI. Not at

16:04

all. But it's not just the personal

16:06

narratives and you being able to tell

16:09

your own story. It's also, you got

16:11

to be asking, where do I have

16:13

data, facts, information about the world that

16:15

the LLLMs aren't going to have? Because

16:18

if you have that, it means that

16:20

you can write something that LLLMs can't

16:22

possibly replicate. So obviously, you know a

16:24

lot about your life. And that's why

16:26

biographies and memoirs and personal stories will

16:29

continue to work out and not be

16:31

copied by LLMs. But what about things

16:33

that you know about the world, right?

16:35

Because there's a lot of stuff in

16:38

that dimension too. So I'll give you

16:40

some examples for my life. For example,

16:42

like I've lived in Austin for almost

16:44

five years now. So there's things that

16:46

I know about the vibe of Austin

16:49

that an LLM isn't going to be

16:51

able to be able to replicate. The

16:53

culture of Austin the people who live

16:55

here how the vibe is changing and

16:58

I can speak to that with a

17:00

lot of specifics in a way that

17:02

in LLLM There's no way they can

17:04

another thing is like like I was

17:07

saying earlier, right? I ran right a

17:09

passage for For six years and we

17:11

ran almost 200 live sessions like I

17:13

know how to run a zoom a

17:15

live session with a few hundred people

17:18

Very well in a way that there's

17:20

just no way that an LLLM is

17:22

going to have an answer on that

17:24

That I have there's just no way

17:27

I have that experience and then the

17:29

other thing is just things that are

17:31

more cutting edge and up to date

17:33

You know what happens with information is

17:35

it gets shared in small tight tightly

17:38

connected social circles and it gets shared

17:40

through sort of whispers and voice and

17:42

then it ends up getting shared in

17:44

conferences through sort of more formal settings

17:47

and maybe podcast and then later on

17:49

it kind of ends up in books

17:51

and It's only later that the LLLMs

17:53

would end up actually having that information

17:55

and to get really concrete here. There

17:58

are things like I'm really interested in

18:00

the YouTube algorithm because that's what we're

18:02

focused on with how I write. And

18:04

there's things right now of how to

18:07

grow on YouTube and how the YouTube

18:09

algorithm works that I'm just talking to

18:11

friends about at dinner and I know

18:13

those things in ways that the LLM

18:15

just it's not going to know. And

18:17

so all of those examples, Austin, the

18:19

intricacies of a Zoom live session, talking to

18:21

friends about the YouTube algorithm, they come down

18:24

to two E's the two E's of Being

18:26

differentiated from LLLMs in

18:28

your nonfiction writing, which

18:30

are experience and expertise. If you

18:32

have a lot of experience in a

18:34

particular domain, if you have a lot

18:37

of expertise, you just know a lot

18:39

about something. And you write that, you

18:41

put that on to the page and

18:43

you do it well, you're going to be

18:45

all right. So, the question is, okay,

18:47

David, you're talking about what I

18:50

should do, how is your writing going to

18:52

change? And I'll tell you this. I'm

18:54

definitely going to focus on experience and expertise,

18:56

but also my writing is going to become

18:58

more personal, it's going to become more opinionated.

19:00

I want to just do bolder work. And

19:03

the piece that I'm writing right now, it's

19:05

a long piece right now, it's probably at

19:07

about 11,000 words, is the story of how

19:09

I became a Christian. And one thing, actually,

19:11

just to go back, I read that testimony from

19:13

the co-founder of Wikipedia, and I said, if I'm

19:15

feeling this about that, well, well, then I can

19:18

definitely do the same thing in my same thing

19:20

in my own thing in my own writing in

19:22

my own writing, and I can trust that it's

19:24

going to stay on the test of time and

19:26

I encourage you to do the same thing.

19:29

Just start thinking, okay, is I'm reading a

19:31

piece of writing, do I think that this

19:33

is going to continue to last or

19:35

will become obsolete? Am I really moved

19:37

by this style of writing? And if

19:39

so, well, hey, maybe consider doing that

19:41

kind of writing yourself. And for me,

19:43

what I like about the, the sort of how

19:46

I became a Christian is it's that

19:48

personal story, it's deeply emotional, and

19:50

it has a spiky point of view,

19:52

which is How in the world did

19:54

I go from being raised Jewish, basically

19:57

being an atheist, living in New York

19:59

City, which like the Mecca of

20:01

materialism, how did I go

20:03

from there, to then believing

20:05

that Jesus Christ is the

20:07

literal son of God, believing

20:10

in the hystericity and truth

20:12

of that story, and then

20:14

thinking that he's my lord and

20:16

Savior, that to me was

20:18

probably the biggest change of

20:20

mind I've had my entire life,

20:22

and how in the world did I

20:25

get to a place where I thought

20:27

this guy, God became flesh. died on

20:29

a cross and was resurrected three days

20:32

later. I mean, that idea seemed absolutely

20:34

crazy to me 10 years ago, and

20:36

now I actually believe it. Like, that

20:38

is a story worth telling, and it's

20:41

the kind of thing, like, hey, I's

20:43

not gonna make that obsolete, that story,

20:45

if I do a good job,

20:47

there's just no way. There's no way.

20:49

This piece, right, it's deeply personal, it's

20:51

deeply opinionated, it's a story

20:53

about what I've been through,

20:56

how I've changed my mind.

20:58

extreme sorrow and pain and how

21:00

I think the world works. How I

21:02

think, what do I think the story

21:04

of reality is? And it's just about

21:06

the greatest and most intimate story

21:08

that I have to tell. And people will

21:11

often say, you know, if they're critiquing AI,

21:13

they'll say, oh, you know, you know, the

21:15

best human writing is still so much better

21:17

than the best AI writing. It's still like,

21:20

like, the best writers are so much better.

21:22

So like, why in the world would you

21:24

even read AI? Like David why are you

21:26

reading all this AI generated writing you

21:29

could go read a bunch of other

21:31

things and here's the thing they'll

21:33

say it's the best so the best

21:35

it's sort of talking about

21:37

quality and I've realized that

21:39

there's really two kinds of quality

21:42

when you sit down to read

21:44

something okay so there is the

21:46

objective quality of a book that's

21:48

the first one and this is

21:50

what everybody thinks about so we'll

21:52

take a book like a book like Gibbons,

21:55

the decline and fall of the

21:57

Roman Empire. Okay, people tell me

21:59

amazing. book. They'll say the writing

22:01

is really strong. It's been super influential.

22:03

Sure, you know, maybe you got some

22:06

facts wrong, but David, you got to

22:08

read this book. Or they'll talk about

22:10

the power broker. That's one of my

22:13

favorites. The power broker by Robert Carrow.

22:15

It is 1,34 pages, but David, I

22:17

know it's long, but I'll tell you,

22:20

Robert Carrow is such a good writer.

22:22

He does insane research. All that sort

22:24

of stuff. And look, I get it,

22:26

the writing quality is great, but actually

22:29

that kind of objective quality is only

22:31

half of the equation. The other half

22:33

of the equation is how tailored is

22:36

a piece of writing to your interests.

22:38

Like, I read the power broker when

22:40

I was living in New York because

22:43

I was living there to kind of

22:45

help me understand my environment, but if

22:47

I was living in Austin, Texas, there's

22:49

just no way that I would read

22:52

that book. It's not tailored enough to

22:54

like... what I'm focused on right now.

22:56

And with the Gibbons example, I haven't

22:59

read the decline and follow the Roman

23:01

Empire because I'm just not that interested

23:03

in the Roman Empire. You know, it's

23:05

like, I get that the writing is

23:08

good, but like how much the writing

23:10

aligns with like what I'm curious about

23:12

right now, that is the other half

23:15

of quality. So quality is one half

23:17

the objective quality of a piece of

23:19

writing, but the other half is how

23:22

tailored is it to your interest. And

23:24

this is the thing about chat cheap

23:26

cheapity. I agree with the critics that

23:28

the writing quality, the objective writing quality,

23:31

isn't as high. You're absolutely right. You're

23:33

absolutely right. Completely agree. But what I

23:35

love about reading with ChatGBT and having

23:38

AI generate writing for me is that

23:40

it's perfectly tailored to my interests at

23:42

all times. We're talking about the Lady

23:45

Bird Lake thing, right? I could go

23:47

read the best book or article that

23:49

has ever been written. on the flora

23:51

and fauna of the city of Austin,

23:54

I bet there's a good book. Maybe

23:56

it's even one. a pulled surprise. But

23:58

actually, it's going to be overcome. What

24:01

I really just want to know is,

24:03

tell me what I need to know

24:05

about the floor in Fauna between where

24:08

I work and where I live in

24:10

the, in March, this exact moment right

24:12

now, so that my walk can be

24:14

more enjoyable. Right? It's so much more

24:17

specific. It's so much more tailored to

24:19

my interest. It's actually shorter. So we

24:21

have to like sort through a bunch

24:24

of information. So it's on the tailored

24:26

to your interest dimension of quality. that

24:28

AI has massively improved. And here's the

24:30

other thing. I'll give you some more

24:33

hope in terms of why writers will

24:35

continue to be just fine. Whether you're

24:37

writing with AI or you're writing without

24:40

AI, the core skills that you need

24:42

to succeed as a writer are exactly

24:44

the same. They're exactly the same. What

24:47

are those core skills? Taste. The ability

24:49

to discern what's worth keeping and what's

24:51

not worth keeping. Look. Whether you're writing

24:53

on your own and you're generating a

24:56

bunch of words with your own fingertips

24:58

or you're working with AI and the

25:00

AI is generating sentences and paragraphs, whatever,

25:03

the vast majority of what ends up

25:05

being generated, you actually end up removing.

25:07

It's just that AI is going to

25:10

end up generating more stuff and you're

25:12

just going to end up cutting way

25:14

way way way way way way way

25:16

way way more. That's It's true that

25:19

you're going to cut the majority of

25:21

what you write, whether you're writing yourself

25:23

or whether the AI is doing it.

25:26

And a lot of what writing is,

25:28

it's just putting a bunch of words

25:30

on the page and you sort of

25:33

have it there. You sort of have

25:35

the marble and then you begin to

25:37

sculpt. And Michelangelo said that what you

25:39

need to do to make a great

25:42

statue is just remove everything that isn't

25:44

what that final statue should be. And

25:46

that's true whether AI is writing or

25:49

you're writing. That's the first thing, and

25:51

that'll continue to be true. So much

25:53

of what AI will produce will be

25:55

not quite right. It'll be junk. And

25:58

guess what? It's the same thing with

26:00

me. So much of what I produce

26:02

is just absolute junk and nonsense and

26:05

clutter, and I just, I get rid

26:07

of that stuff. So that's the first

26:09

thing. Taste is the first core skill,

26:12

and that'll continue to be important. And

26:14

the second thing is a spiky point

26:16

of view. That is a unique insight,

26:18

a unique belief about what you believe

26:21

to be true about the world. And

26:23

man, if you can have that, something.

26:25

that is like this is my bold

26:28

take you know like my take on

26:30

Austin is that Austin is a mediocre

26:32

city but a good place to live

26:35

or this is the famous Peter Thiel

26:37

interview question right what do you believe

26:39

about the world that's true that very

26:41

few people would agree with you on

26:44

those sort of secrets he calls them

26:46

those will continue to be valuable in

26:48

an AI future because AI isn't going

26:51

to be able to produce those nearly

26:53

as well as human beings at least

26:55

for a long time So

27:00

just like in the past, the past

27:02

20, 30 years, just like in the

27:04

present, just like in the future, if

27:06

you have a distinct, an idiosyncratic take

27:08

about how the world works, where the

27:10

world is going, you're going to be

27:13

completely fine. But that was true, it

27:15

is true, and it will be true,

27:17

you know what I mean? Past present

27:19

future. And I'll give you a concrete

27:21

example. So I have a friend who

27:23

thinks that... The future of education is

27:26

going to be school and there's not

27:28

going to be teachers. Instead of teachers,

27:30

you're going to work with coaches and

27:32

instead of those teachers who would basically

27:34

lecture you with information, those coaches are

27:36

basically going to motivate you. And since

27:38

you're not going to be sitting in

27:41

lectures, what you're going to be learning

27:43

through AI and you're going to be

27:45

learning through apps. And he believes that

27:47

because of that, because it's so much

27:49

more effective, because it's so much more

27:51

efficient. that the kids are going to

27:53

be able to learn everything they need

27:56

to know, everything, from K through 8,

27:58

in two hours of learning per day,

28:00

and then his model of schools, then

28:02

you can spend the rest... the day

28:04

learning life skills public speaking making friends

28:06

all that sort of stuff riding bicycles

28:08

whatever it is you kids can learn

28:11

to ride a bike in second grade

28:13

they learn to swim and they can

28:15

do that with with the rest of

28:17

their day and this is a crazy

28:19

idea like I've seen him pitch it

28:21

and people are like that is the

28:23

most insane thing I've ever heard there's

28:26

no way I'm gonna send my kids

28:28

there but for him it's not some

28:30

like pie in the sky theory that

28:32

he just like made up someday. He's

28:34

been thinking about this for 25 years.

28:36

He runs a school. He already has

28:38

data to show that it works. They

28:41

got the school. It's called Alpha School

28:43

in Austin. And most people think it's

28:45

insane, but I think he makes a

28:47

lot of good points. And he thinks

28:49

that he's right. He's like, this is

28:51

the conviction that I have on the

28:54

world. And look, if you have that

28:56

kind of conviction in anything, There's nothing

28:58

to worry about. That is what great

29:00

writing has always had. And AI is

29:02

just not going to be able to

29:04

do that because AI is, at least

29:06

right now and for the time being,

29:09

very trained on the consensus. So this

29:11

is a major white pill for writers.

29:13

That if you have good taste, if

29:15

you have that spiky point of view,

29:17

you're going to be just fine. And

29:19

look, I would even say that those

29:21

are the skills that used to be

29:24

true. Those are the skills that will

29:26

be true. It hasn't really changed. thinking

29:28

about quality as I've been thinking about

29:30

the kind of writing that is going

29:32

to work in this age of AI.

29:34

I've been asking like, is AI going

29:36

to be more like chess or is

29:39

there going to be like music? And

29:41

what I mean by that is with

29:43

chess, what has happened is the AI,

29:45

the computers, are already better than the

29:47

very best humans. They're really good. But

29:49

you have people like Magnus Carlson. who

29:51

are huge celebrities, huge celebrities, and people

29:54

don't really watch the computers play chess,

29:56

they watch the human beings play chess,

29:58

they want to see the human drama,

30:00

they want to see the rivalry, they

30:02

want to see two people. Duking it

30:04

out, trying to basically checkmate the other

30:06

person, right? So when it comes to

30:09

chess, people care less about perfection, they

30:11

care more about humanity. So that's one

30:13

path for AI. The other path could

30:15

be music. And what I mean is

30:17

like, when you walk into a club,

30:19

right, you go to the bar, you

30:22

know, you step out on the dance

30:24

floor, and there's a song playing, you're

30:26

like, damn, this song, slaps, like you

30:28

just know it's good. You turn on

30:30

the radio, you hear a bang, you

30:32

hear a banger, you hear a banger,

30:34

Stop in that moment and say, huh,

30:37

I wonder how this song was made.

30:39

You know, did they use ableton? Was

30:41

this a was this an electric guitar

30:43

or was this an acoustic guitar? Is

30:45

there sampling? You know, what's going on

30:47

there? You don't do that. You're just

30:49

like, oh, this song is sick. This

30:52

song is absolutely filthy. Let's get down.

30:54

And like if it's a vibe, it's

30:56

a vibe. And you just keep listening,

30:58

right? And I was thinking about this

31:00

and I learned that sampling in the

31:02

past used to be kind of taboo.

31:04

People were like, oh this is stuff,

31:07

this isn't real music. And then what

31:09

happened was you had the Beastie Boys

31:11

and he had Dr. Drey and then

31:13

I remember Kanye West was the guy

31:15

who used a bunch of samples. And

31:17

now sampling is like completely normal. You're

31:19

taking songs that were made in the

31:22

past or sort of... taking that beat

31:24

or a section from that song bringing

31:26

it into your song and it's like

31:28

completely fine right there's actually a bunch

31:30

of songs old songs that I found

31:32

through samples of new ones like there's

31:35

a song called the run-around Sue by

31:37

G easy that I used to listen

31:39

to all the time when I was

31:41

in college like the most like college

31:43

dude song to listen to ever and

31:45

I didn't realize that was like an

31:47

old song you know that you'd play

31:50

on record and what now I was

31:52

like when my mom was growing up

31:54

or whatever and the point is That

31:56

sampling used to be the kind of

31:58

thing that people were like, nah, you

32:00

can't do that. And now it's just

32:02

completely normal, completely normal. And I think

32:05

AI is going to be like music.

32:07

I don't think it's going to be

32:09

like chess. Right now, we're in this

32:11

time of those early days of sampling,

32:13

where if you're using AI for your

32:15

writing to help you with your writing,

32:17

people like, that's not cool. And a

32:20

lot of people are like, whoa, that's

32:22

really not cool. I don't like this.

32:24

And look, I get it. I completely

32:26

get it. This is sort of a

32:28

strange moment right now. But you know

32:30

what? I'm pretty confident. That in 15

32:32

years Of course you used AI for

32:35

your writing. It's just it's just like

32:37

a piece of technology of course you

32:39

used it and Like sampling you know

32:41

sometimes you sample a song sometimes you

32:43

don't but people don't have like a

32:45

moral aversion to sampling now like they

32:47

used to and I think AI is

32:50

the same way right now we have

32:52

that moral aversion, but I don't think

32:54

that'll be true and This is the

32:56

key point that I want to make

32:58

that the only thing that will matter

33:00

will be the objective quality of a

33:03

piece of writing. Let me repeat that.

33:05

The only thing that's going to matter

33:07

is going to be the objective quality

33:09

of a piece of writing. Because look,

33:11

just as when I walk into a

33:13

club or some bar and I hear

33:15

the song, I don't care how it

33:18

was made, I don't care who made

33:20

it, all I care about is that

33:22

the song is a vibe. And I

33:24

think it's going to be the same

33:26

thing with writing. I don't care if

33:28

A. I wrote it for you, I

33:30

don't care if A. I wrote it

33:33

with you, I don't care if you

33:35

wrote it by yourself, I think in

33:37

10 to 15 years, that is the

33:39

only thing that'll matter. See, I'm not

33:41

interested in the best writing that only

33:43

humans can do. I'm not interested in

33:45

that. I'm interested in the best writing,

33:48

period. Whatever produces the highest quality writing,

33:50

go, sign me up for that. Sign

33:52

me up for nonfiction. But, you know

33:54

what's funny. You know what's funny. You

33:56

know what's funny. It kind of reminds

33:58

me of special effects in movies like

34:00

one thing that really really bothers me

34:03

kind of more than anything is If

34:05

I'm watching a movie and there's a

34:07

special effect, and that special effect doesn't

34:09

look real, like it looks fake, it

34:11

looks like a special effect, and now

34:13

I'm like taken out of the fantasy,

34:15

the world of the movie that I'm

34:18

watching, and I'm like, oh, okay, they

34:20

just did special effects there. It completely

34:22

kills the vibe, and it's the same

34:24

thing with AI writing. I already feel

34:26

like an aversion to it now. If

34:28

I read a piece of writing, and

34:31

I'm like, I definitely wrote that. and

34:33

all those cliches that AI writing has,

34:35

it just infuriates me. It gets me

34:37

so mad. Because I'm like, you don't

34:39

have the taste to just do a

34:41

good piece of writing, and now you're

34:43

outsourcing to AI, and you're just like,

34:46

oh, AI can do it better, and

34:48

it kind of just bothers me more

34:50

than anything. But for me, now this

34:52

is my opinion, if I read an

34:54

amazing piece of writing, and it's just,

34:56

it's captivating, it's compelling, the takes are

34:58

strong. And someone's like, yeah, you know,

35:01

I used AI to help me refine

35:03

my ideas. I do not have a

35:05

problem with that. I do not have

35:07

a problem with that. I think I'm

35:09

in the minority in 10 to 15

35:11

years there. Okay? But I do hate

35:13

when writing was just so clearly outsourced

35:16

to AI. And here's the other thing.

35:18

Right now, everyone's talking about AI Slop.

35:20

Everyone's talking about it. And Slop is

35:22

a great word, by the way. It's

35:24

just like such a good descriptive descriptive

35:26

word. Everyone says that right now AI

35:28

is the beginning of slop, but in

35:31

some way it's actually the end of

35:33

slop. Let me explain what I mean.

35:35

So, let's just zoom out. For the

35:37

past decade, we'll start with SEO, then

35:39

we'll talk about online writing that kind

35:41

of used to teach. So for the

35:43

past decade, in the world of SEO,

35:46

what you would do, say that you

35:48

wanted to do like a recipe for

35:50

baking cookies, okay? What you would do

35:52

is you would try to get to

35:54

the top of the top of the

35:56

SEO ranking of the SEO ranking, SEO

35:59

ranking. And you can try. What's really

36:01

annoying about looking for a cookie baking

36:03

recipe is that there's... all the story.

36:05

I'm like, I don't need the backstory.

36:07

I don't need all of these photos.

36:09

I don't need to hear that it

36:11

was like your grandma's story or whatever.

36:14

Just like tell me how much sugar

36:16

I need. Tell me how much cookie

36:18

dough, how many chocolate chips. What is

36:20

the ratio? Just give it to me

36:22

straight. But the problem is that the

36:24

incentives of the SEO industry are to

36:26

increase time spent on page because that's

36:29

what Google... search rewards and because that

36:31

you just need to add all of

36:33

this freaking nonsense and the internet is

36:35

just polluted with it like if you

36:37

go search like tell me interesting things

36:39

about London the results are not going

36:41

to be interesting they're just like all

36:44

of these travel sites with like here

36:46

the ten sites that you need to

36:48

see it is to me the very

36:50

definition of slop and it is all

36:52

over the freaking internet it's all over

36:54

the place I would way rather have

36:56

the chat sheet BT output that already

36:59

exists without any of the bells and

37:01

whistles. And here's the point, that the

37:03

incentives of writing were completely misaligned. What

37:05

was best for Google was to basically

37:07

make money from ads. What was best

37:09

for the creators of the recipes or

37:11

the travel sites was, hey, we need

37:14

to rank high on Google, so we

37:16

need to serve Google, and because of

37:18

that, we need to add a lot

37:20

of slop so that the time spent

37:22

on the page, which is like a

37:24

crucial metric for Google. increases. And then

37:27

the people, people like me, people are

37:29

like you, we're just like, guys, can

37:31

you just give us the day answer?

37:33

You know what I mean? So you

37:35

had this total misalignment. And I think

37:37

that is the epitome of Slop. That

37:39

was the SEO world, but how about

37:42

the personal writing world? I used this

37:44

strategy completely. And the strategy was, you're

37:46

going to pick a niche. You're going

37:48

to write consistently about that niche. And

37:50

then under your tweets you're going to

37:52

link to your email newsletter and then

37:54

you're going to send an email newsletter

37:57

consistently and you're going to publish that.

37:59

newsletter every single week no matter what

38:01

you're not going to miss a week

38:03

and even if something isn't like the

38:05

best thing that you can possibly produce

38:07

well hey you know being consistent is

38:09

super important so go publish every single

38:12

week that's what has worked on the

38:14

internet for the past 10 years and

38:16

hey that's what it rewarded I saw

38:18

that strategy I took advantage of it

38:20

it really served me and a lot

38:22

of my students well and you just

38:24

sort of repeat that cycle constantly and

38:27

it worked But here's the thing, distribution

38:29

was king. If you just had distribution,

38:31

you had enough email subscribers, you would

38:33

be fine. And what mattered more was

38:35

publishing consistently than publishing the best quality

38:37

piece of content, and how good your

38:39

distribution was was often more important than

38:42

how good your content was. Like, I

38:44

published a newsletter every single week for

38:46

five years, and I wasn't always entirely

38:48

happy with it. But I just did

38:50

it, because that's what worked on the

38:52

internet. You know you could

38:55

say dude David you were kind of

38:57

contributing to slop and here's the thing

38:59

that age is Gone like my chances

39:02

of succeeding with that strategy in five

39:04

years are so much less than my

39:06

chances to succeeding with that strategy Five

39:08

years in the past Because you're writing

39:11

just has to be really good now

39:13

in the age of AI and I

39:15

get that a lot of slop is

39:18

going to be produced But hopefully and

39:20

this is my hope that We have

39:22

good algorithms and that's basically what algorithms

39:24

do, right? They ignore 99.99999 insert however

39:27

many nines you want. They ignore that

39:29

much content to only give you the

39:31

very best stuff that's tailored to your

39:34

exact interests. Now, quality is king and

39:36

I think content's going to be king

39:38

because if you publish something really good,

39:41

that's the only way to rise up

39:43

above the AI. Here's how I define

39:45

Slop as when simply... publishing or getting

39:47

something done is more important than the

39:50

quality of what you publish. And I

39:52

would say that so much of the

39:54

online writing, going back the last 10

39:57

years, was Slop under the definition I

39:59

just gave, and he ain't succeeding with

40:01

that now. And the reason is that

40:03

the bar for the kind of writing

40:06

that people are going to read is

40:08

already increasing and it's going to continue

40:10

to increase. Because now you're not just

40:13

competing with other human beings, and like

40:15

the scale of the internet and other

40:17

human beings, you're competing with computers and

40:20

computers can produce information so fast. And

40:22

now what I want to do is

40:24

I want to move into how do

40:26

I actually write with LLLMs? Like what

40:29

do I do? Okay, I want to

40:31

just distinguish between two kinds of writing.

40:33

There's AI that writes for you, there's

40:36

AI that does writing with you, and

40:38

I've yet to meet a good writer.

40:40

A good writer that I admire and

40:42

respect that thinks that AI can do

40:45

the writing for you. Not one. Okay.

40:47

But I know a lot of serious

40:49

writers that I admire, they write with

40:52

AI all the time. And that's what

40:54

I do too. So here is. an

40:56

example of what I mean when I

40:59

say writing with not for. So I

41:01

grew up in San Francisco, I now

41:03

live in Austin, Texas, and I love

41:05

that city. It's so beautiful, it has

41:08

so much potential, the city by the

41:10

bay, I mean Tony Bennett has a

41:12

song, he says I left my heart

41:15

in San Francisco, that's how I feel,

41:17

I love that song because I feel

41:19

this like deep emotional connection to that

41:21

city. But the city's been just destroyed

41:24

and tarnished by bad politics. And I

41:26

could say, hey, chat GPT, I want

41:28

you to write me a piece about

41:31

how San Francisco has been destroyed by

41:33

bad politics. Yo, what it's going to

41:35

produce is going to be absolute nonsense.

41:38

It's just not going to be that

41:40

good. Now, it could do a deep

41:42

research report that could be fairly interesting,

41:44

and maybe I would do this like,

41:47

hey, walk me through exactly how that

41:49

happened. Let me understand the history. But

41:51

there is a opportunity. For me, as

41:54

a writer, and this is what I

41:56

mean when I talk about writing with

41:58

AI, Maybe I would team up with

42:00

the deep research to say, hey, can

42:03

you give me some of the historical

42:05

background? But really what I would do

42:07

is I would tell you stories. I

42:10

would tell you stories about, like in

42:12

seventh grade, my mom came home and

42:14

she was covered in blood because she

42:16

had been walking home from the train to

42:19

her car and a guy came up to

42:21

her, threw her purse, and then they left

42:23

and completely drove away. I could tell you

42:25

stories about how when I was in middle

42:28

school those years, I was, you know, this

42:30

kid, a super innocent kid at the time,

42:32

and I'd come out in the morning as

42:34

we'd go to school and the windows would

42:36

just be shattered because people had broken into

42:38

our cars. I could tell you a story

42:40

about how I went to the Orpheum Theater

42:43

on Market Street one year and I picked

42:45

up a syringe as a kid and I

42:47

was like, what is this thing? I was

42:49

like, put that down right now. And I

42:51

didn't learn why there were syringes in the

42:53

middle of San Francisco until I was in

42:55

high school. I could tell you stories about

42:57

how a few of my friends from high

43:00

school in San Francisco, they had to go

43:02

to drug rehab because the drug situation

43:04

moving in through San Francisco was just

43:06

so bad. And my point is this,

43:09

AI is not going to give you that. Those

43:11

are stories that I had to learn from my

43:13

experience in San Francisco that

43:15

I think will give that peace so

43:17

much more life. and help you to see

43:19

the heartbreak that I feel about

43:21

how San Francisco has been destroyed. And

43:24

I can't give you heartbreak. When

43:26

it comes to writing with ChatGPT, here's what

43:28

I do. A lot of the way that

43:30

I start, like I love just speaking out

43:32

ideas. I love doing it. So what I'll

43:34

do is I don't really like typing, it

43:36

hurts my fingers, and you know, I

43:39

feel like I'm gonna get early onset

43:41

arthritis or something like that. What I'll

43:43

do is I'll just go on walks and

43:45

I'll just speak out my ideas and what

43:47

I do is I have a prompt And

43:49

I can share it in the description.

43:52

I have a prompt where what I'll do is

43:54

I'll say hey, I've just spoken something

43:56

out Can you turn that to an outline or

43:59

what I'll do? is I'll say, hey, can

44:01

you turn that into pros? And I'll create

44:03

the ideas and then I'll have AI kind

44:05

of help me out. But then it's not

44:08

just that, I'll say, hey, what ideas that

44:10

I just shared were particularly interesting, what stories

44:12

do you need more information on, what transitions

44:14

were unclear, and I'm sort of instantly getting

44:17

feedback so that I've just finished my version

44:19

one and then. I can get the things

44:21

that are unclear, the things that I need

44:24

to share more about, and then I can

44:26

just do a V2, and the AI, the

44:28

transcription is just so good, is a thousand

44:30

times better than Siri. And then what I'll

44:33

do is I'll ask the AI, based on

44:35

what I've shared, I'll say, what I've shared,

44:37

I'll say, what are the weakest points my

44:40

argument, what are the best points that I

44:42

should double down on, what transitions weren't clear,

44:44

and what more would you need for a

44:46

story to a story to, What was she

44:49

wearing? How old was she at the time?

44:51

What was the night like? Right? It was

44:53

really foggy. Okay, interesting. The fog is like

44:55

a really good motif for that. And maybe

44:58

I wouldn't think about that. AI is really

45:00

good for that. Not because it's like the

45:02

best editor in the world, but because it's

45:05

going to be instant, it's going to be

45:07

free to work with, you know? And I

45:09

can just start being in dialogue. I'll start

45:11

to actually sit down, I'll begin to write

45:14

the piece with with with AI. I've

45:16

been doing this for the piece that

45:18

piece I've been writing about Christianity, really

45:21

thinking through character development. That's something I

45:23

want to get better at. And I'll

45:25

say, okay, what do I need to

45:27

know for what makes a good character?

45:29

So tell me from like the theory

45:31

of writing of Hollywood and movies and

45:34

literature, what makes for good characters? And

45:36

so I might get, okay, these are

45:38

the four things. And then I'll say,

45:40

okay, based on these four things, I

45:42

want you to interview, I want you

45:44

to interview. it'll kind of get my

45:46

brain going and get words on the

45:49

page. And I find that the back

45:51

and forth is so much more generative

45:53

than me trying to do it all

45:55

on my own. So now I have

45:57

AI as a thinking partner. Now this

45:59

is a prediction, and I'm not quite

46:02

sure what's going to come from it,

46:04

but I do. think that AI is

46:06

going to breed new kinds of writing

46:08

that we've never had before. And I

46:10

credit my friend Justin Murphy with this

46:12

idea. But we were talking about the

46:15

renaissance and how new technologies that had

46:17

come out at the time actually led

46:19

to the change in art. So like

46:21

you ever go to a museum. And

46:23

you're in the medieval painting section and

46:25

like everything looks super flat like there's

46:28

no depth and perspective and those are

46:30

like 13th century paintings and then you

46:32

look at like a 15th century Renaissance

46:34

painting that was made in Florence and

46:36

you all of a sudden you see

46:38

that perspective shift and clearly something happened

46:41

to change how people view the world

46:43

and at least we're able to paint.

46:45

What happened is there were a few

46:47

technological innovations that actually led to that.

46:49

So the first was the camera obscure.

46:51

And that allowed artists to basically trace

46:53

images that they would project, and then

46:56

they were tracing, and then that led

46:58

to what they created. So that was

47:00

the first one. And then the second

47:02

was the prospective grids. So there was

47:04

a guy named Leon Alberti. And what...

47:06

He figured out he was an architect

47:09

who's sort of this polymathic guy, but

47:11

he figured out that you could use

47:13

these perspective grids and you could use

47:15

them to draw and then you could

47:17

show perspective in the painting. And now

47:19

you could ask, man, you know, is

47:22

this guy cheating? Like we used to

47:24

paint all by ourselves. Dude, you are

47:26

completely cheating. But like now we look

47:28

at a painting, we're like, I'm just

47:30

happy to have perspective. I'm really glad

47:32

that we had technology that allowed us

47:35

to... make that development. And I think

47:37

that something similar is going to happen

47:39

with AI, where right now we're going

47:41

to end up having these technologies that

47:43

change how we think about writing, how

47:45

we see writing, and it might even

47:48

lead to changes in the kinds of

47:50

writing that we produce. I don't know

47:52

what those are going to be, but

47:54

I think there's going to be some

47:56

interesting things. Actually, one just came to

47:58

mind, which is some stories that you

48:00

can tell that are almost like mad

48:03

libs where you can change different things

48:05

in the stories and then you can

48:07

personalize those stories might be like kids

48:09

stories you can personalize those stories for

48:11

the interest of the kid so say

48:13

that one kid is Really interested in

48:16

the Denver Broncos football team and another

48:18

kid is like really interested in Ballet

48:20

and she like loves Ballet and she's

48:22

from St. Petersburg Russia and she wants

48:24

the setting to take place there The

48:26

little boy wants the setting to take

48:29

place in Denver and you could have

48:31

the same general story that all the

48:33

kids in the class read But then

48:35

that story is tailored for all the

48:37

kids interests. So the idea that you

48:39

produce one piece of writing that everybody

48:42

reads might change where you can change

48:44

different things inside the piece of writing

48:46

to really tailor it for what people

48:48

are interested in. That's what I mean.

48:50

I think we're gonna get some changes

48:52

in the kinds of writing that become

48:55

popular. And so what I'm working on

48:57

when it comes to writing with ChatGPT

48:59

is I'm working on a custom project

49:01

folder for my writing style. And I'm

49:03

basically trying to say, here's how it's

49:05

structured. The first section is sort of

49:07

at the top, it's what do I

49:10

want my writing to be like? That's

49:12

just. Really describe it super clearly. That's

49:14

the first section layout say 10 bullet

49:16

points for each right and What I've

49:18

done is I haven't just thought about

49:20

this But what I've done is I've

49:23

taken my best writing and I've said

49:25

Hey, I want you to actually analyze

49:27

this for me AI and tell me

49:29

what are the things that are going

49:31

on in the writing? And I want

49:33

you to describe it and then I'll

49:36

take the best descriptions that it gives

49:38

me and feed that back into it.

49:40

And I'll use AI and work with

49:42

it to really compress it so that

49:44

I have the clearest, most succinct descriptions

49:46

of what I am and I'm not

49:49

going for. So that's the first thing.

49:51

And then the second thing is, from

49:53

there, training data on all the things

49:55

that I do and don't want. So

49:57

to say that I want my writing

49:59

to be interesting, right? Well, I'm going

50:02

to put in a lot of paragraphs

50:04

that I think are particularly interesting, and

50:06

then I'm going to actually describe why

50:08

I think that paragraph is interesting and

50:10

say, hey, I want your help getting

50:12

me to write like that as you

50:14

give me feedback. That's exactly what I'm

50:17

going for. And training data is going

50:19

to be more and more important. The

50:21

way that I'm taking notes is beginning

50:23

to change because of AI. Right now.

50:25

It used to be that I would

50:27

take notes for me to read. Now

50:30

I'm increasingly taking notes for AI to

50:32

read. And what that means is that

50:34

when I just had my own notes,

50:36

I wanted a bunch of different small

50:38

notes and then I could search and

50:40

I didn't have to spend a bunch

50:43

of time scrolling. With AI, it's very

50:45

different. Rather than having a bunch of

50:47

pages with a little bit of information,

50:49

I want a few pages with a

50:51

lot of information on them. And here's

50:53

why. AI can do the searching, it

50:56

can do the scrolling, no problem, and

50:58

the context windows for AI are getting

51:00

bigger and bigger and bigger. And here's

51:02

what I mean by that. So if

51:04

you say, hey, I want your help

51:06

with my writing, I might now be

51:09

able to put in 30, 40,000 words

51:11

of training data. In two years, I

51:13

think I'll be able to put in.

51:15

two million words worth of training data

51:17

and then ten million words worth of

51:19

training data and it'll easily be able

51:21

to parse that so I'm just gonna

51:24

have these giant documents for different parts

51:26

of my life that I'll just be

51:28

able to plug into the AI and

51:30

it'll be able to to read the

51:32

whole thing for me and if I

51:34

have a bunch of different documents then

51:37

I always have to copy and paste

51:39

that so the way that I'm taking

51:41

notes is beginning to change and I'm

51:43

starting to think about how do I

51:45

write for LLLMs as much as my

51:47

future self when it comes to taking

51:50

notes so that I can really begin

51:52

to work with these LLLMs and give

51:54

them context on how I want to

51:56

write, what's going on in my life.

51:58

meeting notes like another thing is maybe

52:00

I put all my emails that I've

52:03

really thought were good into a single

52:05

document and then I say hey I

52:07

want your hope writing emails do it

52:09

in the style that I've shared that's

52:11

you know a hundred thousand words worth

52:13

of emails that I've written and that

52:16

leads me to how I think with

52:18

LLLMs so when I'm discovering new ideas

52:20

I found that jamming with LLLMs is

52:22

more useful than talking to basically any

52:24

person in my person who I'm close

52:26

to who give really good feedback. And

52:28

look, I'm not the only one. Microsoft

52:31

CEO Satya Nadella, he is no schmuck,

52:33

he's a smart guy, here's what he

52:35

said. The new workflow for me is

52:37

that I think with AI and work

52:39

with my colleagues. That gets to the

52:41

heart of it. You think with AI,

52:44

you work with your colleagues. You're going

52:46

back and forth, you're working through problems,

52:48

strategy with the different AI. And on

52:50

the same wavelength, I have a friend

52:52

in town, have a friend in town,

52:54

Probably as 2,000 employees working for him.

52:57

The guy was talking about earlier, right?

52:59

And you know what he said to

53:01

his executive team? He said, it's got

53:03

to the point where talking to an

53:05

LLLM for an hour is more useful

53:07

than about 70% of the conversations I

53:10

have with you. And that's his executive

53:12

team. These are talented freaking people. This

53:14

isn't to say that his executives aren't

53:16

competent. It is to say that the

53:18

LLLMs are ready at the point. Where

53:20

if you're really good at prompting and

53:23

it has a lot of context, it's

53:25

really good at bringing in strategic context

53:27

into the chat window. These LLMs are

53:29

really good. And you know what I've

53:31

been having a blast with. Oh my

53:33

goodness, it's so fun. So Grock is

53:35

the Twitter AI. And if you go

53:38

into Grock, it has all of these

53:40

modes at the top. Like it has

53:42

assistant and it has storyteller, meditation, unhing

53:44

sexy, sexy, sexy, whatever. But my favorite

53:46

one for thinking is argumentative mode. I've

53:48

started playing with Grock to really think

53:51

through ideas with and what'll. is I'll

53:53

take something of that very high conviction

53:55

and I will say hey I want

53:57

you to challenge this I want you

53:59

to argue with me and I want

54:01

you to point out my core thesis

54:04

and tell me why it's wrong and

54:06

we'll get on voice mode and we

54:08

will just start arguing like the other

54:10

day we were arguing about something here

54:12

and she was like What? So you're

54:14

just going to get LLLM's to do

54:17

all your thinking for you because you're

54:19

talking to LLLM's like you are so

54:21

freaking lazy? What the heck is wrong

54:23

with you? And I'm like, how dare

54:25

you call me lazy? It's not that

54:27

LLLM's giving me the answers, it's that

54:30

LLLM is helping me ask good questions

54:32

and then I can find answers in

54:34

my mind. And the core thing is

54:36

that the answers are like these little

54:38

chambers in my brain and the I

54:40

would have never found on my own...

54:42

What the heck is wrong with you?

54:45

And so we go back and forth.

54:47

It's so fun. And that is what

54:49

I'm using LLLMs for. And a lot

54:51

of the people who you end up

54:53

talking to, they'll be too agreeable or

54:55

too disagreeable. And they'll sort of stay

54:58

in that lane. What's really fun about

55:00

arguing with the LLLMs is you can

55:02

get it to act exactly how you

55:04

want it to act. So like, the

55:06

other day I was getting really annoyed

55:08

with it because it was getting really

55:11

annoyed. tone it down a bit and

55:13

actually just be more supportive here. And

55:15

I'm giving you these ideas. Can you

55:17

just help me find and shape the

55:19

best ideas that I have? Let's work

55:21

together. And then a few minutes, I'm

55:24

going to ask you to argue with

55:26

me. And so that's what we did.

55:28

And then what I do is I

55:30

get to the very end and I

55:32

say, okay, based on this entire conversation,

55:34

I want you to summarize. What are

55:37

the key points of pushback that you

55:39

gave me and what questions. Should I

55:41

think about for next time? And I'll

55:43

just scroll to the bottom of the

55:45

chat window and I will just have

55:47

a full summary of what we spoke

55:49

about. And look, this might sound weird,

55:52

but it's actually kind of refreshing to

55:54

be in like an extreme argument with

55:56

somebody when I just know it's not

55:58

going to impact the relationship. Right? Like

56:00

if I did this with a friend,

56:02

we're just fighting with... each other like

56:05

that, it'd be kind of a thing,

56:07

right? But with the AI, it's like,

56:09

whatever. So I can go harder, they

56:11

can go harder, it's good fun. So

56:13

where is all this going? We've been

56:15

talking about the Satya Nadella, thinking with

56:18

AIs, working with human beings. I was

56:20

talking about those long context windows. Now

56:22

we're talking about arguing with the AIs.

56:24

And one killer app of Lelems is

56:26

going to be memory. Humans are really

56:28

bad of memorizing things. from last weekend.

56:31

So I've been, you know, seeing this

56:33

girl and it's one of our, you

56:35

know, fourth, fifth date and she had

56:37

a tough week. So I was like,

56:39

hey, you know, what would make your

56:41

week really good? She's like, hey, you

56:44

know, I just want to get dressed

56:46

up and go to dinner. I said,

56:48

great. We'll go to a steak house

56:50

in town. So I'm like, all right,

56:52

this could be great night, you know,

56:54

pick her up in the Uber. I

56:56

finally tell her, hey, you know, this

56:59

is where we're going. I say, hey,

57:01

this is my favorite steak, steakhouse in

57:03

town. She goes, uh, what, you took

57:05

me to a steakhouse? I told you

57:07

twice, I don't eat red meat. And

57:09

it was just like, instant, just like

57:12

change in vibe. And I was just

57:14

like, oh my goodness, you're so stern,

57:16

I was super embarrassed, felt so much

57:18

shame. I was like, dude, you've got

57:20

to be kidding me kidding me kidding

57:22

me, and... All this is to say,

57:25

humans don't have good memories. Humans don't

57:27

have good memories. The rest of it

57:29

ended up going fine. Took us like

57:31

20 minutes to get over it. But

57:33

humans don't have good memories. And we

57:35

forget things all the time. And AI

57:38

is going to be really good at

57:40

remembering conversations from weeks ago, from months

57:42

ago, from years ago, and you don't

57:44

just have that context at all times.

57:46

If you're comfortable doing this, I recommend

57:48

using Granola AI to basically record your

57:51

work meetings. And what it'll do is

57:53

it doesn't record the audio to just

57:55

get a transcript, but what it'll do

57:57

is it'll take that transcript and give

57:59

you a summary at the end. which

58:01

then allows you to, right afterwards, you

58:03

can ask questions about what it is

58:06

that you spoke about. So for example,

58:08

we're in a meeting on Tuesday of

58:10

this week, and I spoke out the

58:12

intro for this video in that meeting,

58:14

but then me and a guy I

58:16

work with, we sort of forgot exactly

58:19

what the intro was, but I went

58:21

right into Grado, I said, hey, can

58:23

you tell me, what was the intro

58:25

that I said that we should do

58:27

during the meeting, and it just zip,

58:29

zip, zip, zip, zip, zip. spelled it

58:32

out for me. But what's going to

58:34

end up happening is you're going to

58:36

be able to say, hey, I had

58:38

lunch with my friend Sarah two years

58:40

ago, I haven't seen her in some

58:42

time, can you remind me what it

58:45

is that we spoke about at lunch?

58:47

And look, certain people are going to

58:49

be really comfortable with this, certain people

58:51

are going to be not comfortable, that's

58:53

a personal preference thing, but the point

58:55

is AI is going to be able

58:58

to look across a wide swath of

59:00

people and ideas. and give you context

59:02

on what's going on. So it's not

59:04

just going to look at your own

59:06

goals and your own notes, but it's

59:08

going to be able to look at,

59:11

hey, what are all the memos that

59:13

have been written inside of a company?

59:15

What are all the emails that have

59:17

been sent? Say that you run a

59:19

company of thousands of people. And here's

59:21

why. So... AI is going

59:23

to be able to basically scan every

59:26

single email and see exactly what's going

59:28

on and give instant feedback on what

59:30

people are thinking, what they're writing, from

59:33

the perspective of the CEO. So I

59:35

got a friend named Warkesh Patel, he's

59:37

got a great podcast about AI, and

59:39

he was writing about Google and Sundar

59:42

Pinchai, their CEO, and this was his

59:44

prediction. I'm going to redo this quote.

59:46

Human Sundar simply doesn't have the bandwidth

59:49

to directly oversee 200,000 Google employees. hundreds

59:51

of products and millions of customers. But

59:53

AI Sundar's bandwidth is capped only by

59:56

the number of TBUs you give him

59:58

to run on. And I'm. but he

1:00:00

basically means like computer processing power. All

1:00:02

of Google's 30,000 middle managers can be

1:00:05

replaced with AI Sundar copies. Copies of

1:00:07

AI Sundar can craft every product strategy,

1:00:09

review every pull request, answer every customer

1:00:12

service message, and handle all negotiations. Everything

1:00:14

flowing from a single coherent vision. A

1:00:16

company of Google scale can run much

1:00:19

more as the product of a single

1:00:21

mind, the articulation of one thesis than

1:00:23

is possible now. Now that's a crazy

1:00:25

prediction of the future, but I can

1:00:28

sort of see it happening, right? Where

1:00:30

your job as the leader is to

1:00:32

write and write and write and make

1:00:35

your thinking and strategy legible, and then

1:00:37

you have basically an army of computers

1:00:39

who are reviewing everything and making sure

1:00:41

that the entire company is aligned. I

1:00:44

texted that to one guy who runs

1:00:46

a big company said, man, that would

1:00:48

be my dream. And other people are

1:00:51

like, wait, what, like, that's crazy. And

1:00:53

I think it speaks to something bigger,

1:00:55

where AI is this unique technology in

1:00:58

that people who are managers, they have

1:01:00

totally said yes to this technology, much

1:01:02

more than kind of rank and file

1:01:04

people who just do the work. And

1:01:07

I think that part of the reason

1:01:09

is that working with AI, if you're

1:01:11

a manager, is actually super similar to

1:01:14

the way that you've always done the

1:01:16

work. but working with AI if you're

1:01:18

kind of a rank-and-file person is completely

1:01:21

different. So here's what I mean. If

1:01:23

you're a manager, what is your job?

1:01:25

Your job is to basically set a

1:01:27

vision, describe that vision, delegate that vision,

1:01:30

get a response, have that response, not

1:01:32

be what you want, give feedback, go

1:01:34

through cycles of iteration, and then eventually

1:01:37

get something that's pretty good, and then

1:01:39

pass it on. That's exactly what you

1:01:41

do with LLMs. It is the exact

1:01:43

same... motion and that's why I think

1:01:46

managers are like, oh yeah, I've been

1:01:48

doing this for you. But the other

1:01:50

thing is, when you're actually working with

1:01:53

people, there's going to be drama, and

1:01:55

you've got to manage a large team,

1:01:57

you've got to do one-on-one. For people

1:02:00

who don't like doing those things, they're

1:02:02

like, oh my goodness, this is amazing.

1:02:04

I get to delegate a lot of

1:02:06

my work in the same way. I

1:02:09

get to delegate a lot of my

1:02:11

work in the same way I've always

1:02:13

done it, but I don't have to

1:02:16

do with the one-on-one. And I think

1:02:18

that's like rock, chatchachchach-chach-chach-chach-chach-chach-chach-h-chach-chach-h-chach-chach-h-h-h-chach-h-chach-h-h-h-chach-h-h-h-h-h-h-h-h-h-h-h-h-h-h-h-h-h-h-h-h-h Claude, whatever it

1:02:20

is, they sort of assume that they're

1:02:23

all the same. And six months ago,

1:02:25

that was more true than they are

1:02:27

today, and I think that the models

1:02:29

are going to begin to diverge. Here's

1:02:32

why. What's happening is there's crazy competition

1:02:34

at the model layer. And when you

1:02:36

get crazy competition, one way that people

1:02:39

respond is differentiation. So what you had

1:02:41

with, for example, Anthropic, the company that

1:02:43

owns Claude, when they released their 3.7

1:02:45

model, which came out in February of

1:02:48

March of 2025, They said, hey, we're

1:02:50

really going to focus on coding for

1:02:52

this model, and that's really where the

1:02:55

improvements are going to be. GPT 4.5

1:02:57

was really focused on more qualitative parts

1:02:59

of writing. Grock, Grock is really focused

1:03:02

on free speech and even up-to-date stuff.

1:03:04

The ChatGPT cutoff for knowledge is like

1:03:06

September 23 or something like that and

1:03:08

then it sort of uses search for

1:03:11

more recent stuff. Grock is much more

1:03:13

up to date. So like if I

1:03:15

need something up to date, I go

1:03:18

to Grock, not ChatGPT and I think

1:03:20

that the models are going to get

1:03:22

more and more distinct over time. And

1:03:25

when it comes to tracking what's going

1:03:27

on with AI, talk to your friends,

1:03:29

especially ones who work at big companies

1:03:31

about how they're using AI. A lot

1:03:34

of the coolest things are happening in

1:03:36

major companies and they're staying kind of

1:03:38

in private. Like I heard this through

1:03:41

the grapevine that Google allows anyone who

1:03:43

works there to basically do searches with

1:03:45

an unlimited size context window and I'm

1:03:47

sure that there's some really cool things

1:03:50

that... that come from that, and I'm

1:03:52

just talking to different friends, they're building

1:03:54

internal AI tooling that I'm sworn not

1:03:57

to talk about, but a lot of

1:03:59

the AI progress that's currently happening is

1:04:01

happening in back channels, so just ask

1:04:04

your friends what's going on, and just

1:04:06

be like, yo, shh, I won't tell

1:04:08

me, but if you want to email

1:04:10

me, that's fine, but don't tell anyone

1:04:13

else. And look, a lot of this

1:04:15

futures are ready here, actually this morning,

1:04:17

a friend sent me an article that

1:04:20

said me an article that said this,

1:04:22

a friend sent me an article that said,

1:04:24

a, a, a, a, a, a, a, a, a, a, a, a,

1:04:26

a, a, a, a, a, a, a, a, a, a, a, a,

1:04:28

a, a, a, a, a, a, AI generated poetry is indistinguishable from

1:04:30

human written poetry and is rated more favorably. Now

1:04:32

this sort of a crazy prediction

1:04:35

I got no scientific study to back

1:04:37

this up, it's just something I'm kind

1:04:39

of feeling right now, but I think

1:04:41

that AI is going to actually allow

1:04:43

us to communicate with animals. I think

1:04:45

that there's all of these deeper forms

1:04:47

of communication that are happening in the

1:04:49

world and we're going to really

1:04:51

understand that at another level and

1:04:53

we might be able to... Chirp

1:04:55

with the birds and stuff, like

1:04:57

the raven's outside, a little cardinal,

1:04:59

let's buy her window every single

1:05:01

morning. I think we might be

1:05:03

able to start talking to him.

1:05:05

And it's one of those weirder,

1:05:07

second or third order effects that

1:05:09

might come from AI. And so

1:05:11

it's not just that these models

1:05:13

are going to have increasingly rising

1:05:16

IQ, but they're also going

1:05:18

to have different personalities, which

1:05:20

is going to be as diverse as people

1:05:22

are. You know what I mean? And when

1:05:25

it comes to thinking with AI,

1:05:27

I see it as already useful in

1:05:29

a bunch of different ways. So at

1:05:31

this point, I think you'd be insane

1:05:33

if you have a medical issue to only

1:05:35

talk to a doctor, I think talking

1:05:38

to chat GPT or some LLLM is

1:05:40

really a no-brainer thing to

1:05:42

do, at least to get a second

1:05:44

opinion. And once again, that is

1:05:46

a great place where we'd want

1:05:48

to talk to a bunch of

1:05:50

different models to see, hey, are

1:05:52

you getting variant? different perspectives or

1:05:54

the same perspectives and how does that compare

1:05:56

to what your doctor is saying at the

1:05:58

very least you can ask your doctor way

1:06:01

better questions when you show up, which

1:06:03

will help you get a better diagnosis.

1:06:05

I think that's a no-brainer. When I

1:06:07

was in Buenos Aires, I was trying

1:06:09

to understand sort of the immigration patterns

1:06:12

of the city. And this was one

1:06:14

of those moments for me when I

1:06:16

was like, whoa, you can do this.

1:06:18

And what I did is I asked

1:06:20

it to make me a table from

1:06:22

1890 to 1920 in Buenos Aires during

1:06:25

the early boom. how much immigration came

1:06:27

in all those decades, but then also

1:06:29

I want a table of all the

1:06:31

different places that people came from. So

1:06:33

Italy and Spain and then when it

1:06:35

were in Italy, southern Italy, northern Italy,

1:06:38

and I couldn't believe it. I couldn't

1:06:40

believe how much easier it was for

1:06:42

me to just look at that data.

1:06:44

Google wasn't giving me good information. And

1:06:46

if I wanted a book on that,

1:06:48

I just don't, in Buenos Aires, I

1:06:51

don't know how I would have gotten

1:06:53

that. It was so easy with, with

1:06:55

chat GPDPD. It could have hallucinated there.

1:06:57

I will completely acknowledge that. It definitely

1:06:59

could have hallucinated. So here's the thing.

1:07:01

I wasn't doing this as like a

1:07:04

scholar, right? I wasn't taking what ChatGBT

1:07:06

gave me and then instantly, you know,

1:07:08

writing that and sharing it. Hey, this

1:07:10

is like the gospel truth or something

1:07:12

like that. No, it's just trying to

1:07:14

get a general sense for what's going

1:07:17

on and I had enough context where

1:07:19

I was like, okay, if... It was

1:07:21

telling me that a bunch of people

1:07:23

came to Argentina came from China. Like

1:07:25

I know that that's obviously not true.

1:07:27

And I was just kind of trying

1:07:30

to get a general picture of what

1:07:32

was going on. And AI is really

1:07:34

good at that. But look, hallucinations are

1:07:36

definitely a thing. I don't think that

1:07:38

they're nearly as big of a problem

1:07:40

as they used to be. They were

1:07:43

a big problem two years ago. Now

1:07:45

they're less of a problem. And still

1:07:47

sort of in the cultural awareness, I

1:07:49

think that people think that they're a

1:07:51

way bigger problem than they actually are.

1:07:53

Do not Do not just take something

1:07:56

that an LM gives you and pass

1:07:58

it along for fact like you really

1:08:00

need to be careful there Okay. And

1:08:02

Benedict Evans, he's a, he's a, he's

1:08:04

a technological analyst. He speaks about this

1:08:07

really well. I love his synopsis of

1:08:09

LLLMs. He says that LLLMs are really

1:08:11

good at things that don't have wrong

1:08:13

answers, but really bad at precise information

1:08:15

retrieval. And I think that's right. So

1:08:17

like, if you're stuck on a sentence

1:08:20

in your writing and you're like, the

1:08:22

sentence just isn't quite right, can you.

1:08:24

Give me like 10 ideas for how

1:08:26

to improve this sentence. Like there's no

1:08:28

right answer there, right? That's sort of

1:08:30

a taste thing. It's a felt sense

1:08:33

thing. Or like, hey, next weekend I'm

1:08:35

planning a birthday party for a friend

1:08:37

in upstate New York. Hey, can you

1:08:39

give me some ideas for for what

1:08:41

to do? These are the sorts of

1:08:43

things he's interested in. Can you sort

1:08:46

of help me with an agenda and

1:08:48

an itinerary? It'll be really good at

1:08:50

because there's not like a right answer.

1:08:52

Yo for I need some really good

1:08:54

quotes from John Steinbeck. It's just not

1:08:56

going to be helpful with that and

1:08:59

you can make you can make some

1:09:01

mistakes. So like hard-earned mistake here. Okay,

1:09:03

paid the cost for this one. So

1:09:05

I got this project called writing examples

1:09:07

and early on I was like, hey,

1:09:09

you know, we'll write an article about

1:09:12

John Steinbeck. So what I did, I

1:09:14

said, all right, John Steinbeck. Has he

1:09:16

ever written something about food? And so

1:09:18

I'm in GPT-4 and we're going back

1:09:20

and forth and It gives me this

1:09:22

quote and it's just like the perfect

1:09:25

quotes the perfect quote so I spend

1:09:27

like two days Working on this quote

1:09:29

about you're taking the quote and sort

1:09:31

of analyzing it. How does John Steinbeck

1:09:33

write about food? Then we're like man

1:09:35

this quote is so good this information

1:09:38

is great. Let's go make a video

1:09:40

about this. So we spend a bunch

1:09:42

of time recording We you know we

1:09:44

get our editors on it and we're

1:09:46

like We're so happy with it. We're

1:09:48

so happy with it. And so we're

1:09:51

like, yes, we're going to publish the

1:09:53

article. We're going to publish the video.

1:09:55

We send it out. to like 30,000

1:09:57

people. And we get some emails like,

1:09:59

you know, John Steinbeck never, never wrote

1:10:01

this quote, needs to have eaten it.

1:10:04

And I was like, it literally hallucinated

1:10:06

the dang quote. And, ah, you know,

1:10:08

I learned that one the hard way.

1:10:10

I learned that one the hard way.

1:10:12

And that's just my point. When it

1:10:14

comes to quotes, it'll just make stuff

1:10:17

up and just completely BS and answer.

1:10:19

And you got to be careful there,

1:10:21

okay? So LLLMs tend to be good

1:10:23

at the things that computers are bad

1:10:25

at, and they tend to be bad

1:10:27

at the things that computers are good

1:10:30

at. Computers are really good at just

1:10:32

like precise information retrieval. Like if you

1:10:34

need this exact thing, LLLMs aren't precise

1:10:36

in that way. You know, they'll make

1:10:38

stuff up all the time. And there's

1:10:40

good things about that. And there's bad

1:10:43

things about that. So be careful with

1:10:45

hallucinations and don't make the same mistake

1:10:47

I did. Real embarrassing. And then also,

1:10:49

like, when it comes to how I

1:10:51

use AI, you know what it's really

1:10:53

good for? It's really good for meeting.

1:10:56

So if you're meeting somebody who, especially

1:10:58

somebody who has some information about them

1:11:00

on the internet, if you put in

1:11:02

a deep research report and you say,

1:11:04

I'm meeting this person. I want your

1:11:06

help getting background on them. This is

1:11:09

what I'm working on and these are

1:11:11

my goals for the meeting. Deep research,

1:11:13

it'll give you a really good answer.

1:11:15

And the advice they don't give you

1:11:17

will be solid. Like I've been meeting

1:11:19

with some publishers for how I write

1:11:22

to help me get guests. And before

1:11:24

the meeting, I'll have deep research, create

1:11:26

a whole report, and it'll do a

1:11:28

really good job of saying, okay, if

1:11:30

you want to pitch how I write

1:11:32

and I'll tell. Tell them about the

1:11:35

show, what I'm trying to do. And

1:11:37

they'll say, if you want to pitch

1:11:39

how I write, here's my advice. And

1:11:41

the advice is pretty darn good. It's

1:11:43

as good as any advice that I'll

1:11:46

get from a human being. Okay, so

1:11:48

that's how I think with LLLMs. And

1:11:50

I just want to begin to wrap

1:11:52

here and talk about. how to follow

1:11:54

what's going on in AI and then

1:11:56

I'm gonna get super concrete and specific

1:11:59

about exactly how I use every single

1:12:01

model because by far the craziest thing

1:12:03

about this clearly if these companies are

1:12:05

using AI to name their models it

1:12:07

ain't working well because it's so darn

1:12:09

confusing so I'm gonna talk through exactly

1:12:12

what I do but here's the first

1:12:14

thing I got this from Tyler Cowan

1:12:16

on our recent how I write interview

1:12:18

and He made a great point. He

1:12:20

said, if you want to understand what's

1:12:22

going on in AI, you got to

1:12:25

be using the latest models, and that

1:12:27

means paying for them. Okay, the free

1:12:29

models are currently like six months behind

1:12:31

the cutting-edge models, and the cutting-edge models

1:12:33

are just so much better. Okay, deep

1:12:35

research was the iPhone moment for me

1:12:38

with AI. It was that moment when...

1:12:40

I was actually there, I was at

1:12:42

Macworld in 2007, I think it was,

1:12:44

when Steve Jobs announced the first iPhone.

1:12:46

And I remember, like, I had some

1:12:48

flip phones before, and I remember seeing

1:12:51

that thing, it was in this glass

1:12:53

spiral box, rotating slowly. And I was

1:12:55

at seventh grade, and I knew that

1:12:57

the world had changed, that something, something

1:12:59

was going to be different. And that

1:13:01

first iPhone, it didn't even have the

1:13:04

App Store, I think. It sort of

1:13:06

ended up changing the world changing the

1:13:08

world. But you kind of just knew

1:13:10

it when the V1 came out. That's

1:13:12

how I felt with deep research. And

1:13:14

if you're on a plan that doesn't

1:13:17

allow you to use deep research, and

1:13:19

you're like, ah, AI isn't there yet,

1:13:21

you know, get on the, get on

1:13:23

the bigger plan. I think you have

1:13:25

a lot more credibility in terms of

1:13:27

your critique if you're using the latest

1:13:30

models. And it's also going to help

1:13:32

you see how things are changing. So

1:13:34

be on the latest models and use

1:13:36

different models so that you can see

1:13:38

what's going on. Claude 3.7, Grock 3,

1:13:40

ChatGBT 4.5, Open AIs Deep Research Tool,

1:13:43

that runs on the unreleased O3 model,

1:13:45

like I know it's super confusing, but

1:13:47

those are the kinds of models. that

1:13:49

you should be using if you want

1:13:51

to be on the cutting edge. You're

1:13:53

going to have to pay for them,

1:13:56

but they are so much better than

1:13:58

the free model. And what always happens

1:14:00

to me with AI skeptics, this is

1:14:02

the number one thing that annoys me

1:14:04

about AI skeptics, is they'll be like,

1:14:06

AI is not going to do this,

1:14:09

AI isn't doing that, and then you'll

1:14:11

talk to them and you'll be like,

1:14:13

okay, tell me about how you use

1:14:15

GPT. and they're like well you know

1:14:17

I used it for this and it

1:14:19

didn't give me a good output at

1:14:22

the beginning so like it's not that

1:14:24

good so I'm like okay so if

1:14:26

you worked with someone and the first

1:14:28

thing they produced wasn't very good and

1:14:30

you just stopped working with them like

1:14:32

you wouldn't do that you would kind

1:14:35

of do that you would kind of

1:14:37

work with them you would kind of

1:14:39

work with them and that you would

1:14:41

kind of work with them and that

1:14:43

you would kind of work with them

1:14:45

and the second thing is they'll be

1:14:48

like No, it's not good because you

1:14:50

don't pay for it. You know what

1:14:52

I mean? So what ends up happening

1:14:54

is they reinforce their own skepticism. Okay?

1:14:56

And I think that the mark of

1:14:58

good thinking is that whatever you believe

1:15:01

you're trying to challenge it. So if

1:15:03

you're like super bearish and skeptical of

1:15:05

AI, you're going to have a lot

1:15:07

more credibility with me. If you're really

1:15:09

trying to use it well, and if

1:15:11

you're using the latest models, but who

1:15:14

cares about me? Do that favor to

1:15:16

yourself so that you can see what's

1:15:18

going on. It's just going to give

1:15:20

you a much better sense. Now I

1:15:22

want to talk about what models do

1:15:25

I use for what, and I'm going

1:15:27

to get super concrete and specific. And

1:15:29

I'll start off with Open AI and

1:15:31

ChatGPT. Their 4.5 model is the core

1:15:33

model that I make things with. Okay?

1:15:35

So it's pretty funny with Nish humor.

1:15:38

It'll make me laugh every now and

1:15:40

then. And then also, it's pretty good

1:15:42

at writing with voice. It's like a

1:15:44

six and a half, seven out of

1:15:46

ten, but it's way better than chat

1:15:48

TVT4 was about a year and a

1:15:51

half, like way better. But also it's

1:15:53

kind of annoyingly corporate and sterile in

1:15:55

its output, and that's my least favorite

1:15:57

thing about writing with AI, the way

1:15:59

it sort of like sucks up to.

1:16:01

and just that annoying voice that it

1:16:04

uses, oh, tactically, what I'll do is

1:16:06

I'll talk into my phone, I'll ask

1:16:08

GPT 4.5 to clean it up as

1:16:10

pros, I'll do a review on what

1:16:12

it's written, and then I'll send it.

1:16:14

And now get this. Recently I did

1:16:17

this for a team write-up, so I

1:16:19

had something that I want to share,

1:16:21

sort of some feedback to give, so

1:16:23

I went on a walk, and I

1:16:25

did a voice transcription. I said, this

1:16:27

is what I'm thinking about. You know,

1:16:30

I repeat myself, I'll stutter, do whatever,

1:16:32

things that all humans do. And then

1:16:34

I'll say, hey, just clean it up

1:16:36

and turn it to a piece of

1:16:38

writing for me and then be really

1:16:40

clear about the thesis and the main

1:16:43

point. And it'll just turn it right

1:16:45

into pros and I copied and pasted

1:16:47

that in the slack. And then at

1:16:49

the bottom, I had sort of in

1:16:51

italics these parentheses and everyone said no.

1:16:54

Everyone said no. Like the output is

1:16:56

pretty darn good. It's pretty darn good.

1:16:58

Sometimes I need to do some editing,

1:17:01

but it's pretty darn good. And the

1:17:03

same thing happened to me last November.

1:17:05

This was another like, whoa moment for

1:17:07

me. So there was a friend who

1:17:09

was in San Francisco, I went out

1:17:11

to dinner, and he took a few

1:17:14

bullet point notes at dinner, and then

1:17:16

he fed those bullet points into the

1:17:18

AI, and he shared a slack message

1:17:20

with me. And I said, dude, what's

1:17:22

going on? This is... This is the

1:17:24

best voice I've ever seen your writing

1:17:27

hat. And he goes, A.I. wrote it

1:17:29

for me. And I was like, no

1:17:31

way. And he's a good writer. You

1:17:33

know, he probably has 5,000 email subscribers?

1:17:35

Like, he's not Robert Carrow, but like,

1:17:37

he's a good writer. And I said

1:17:40

to him, this is the best thing

1:17:42

you've ever seen. I've ever seen you

1:17:44

right. Okay, so to get back to

1:17:46

this, I use Chatty BT 4.5, whenever

1:17:48

I need to create something, and then

1:17:50

whenever I need to consume something, I'll

1:17:53

use O1Pro or Deep Research. And that's

1:17:55

because Chatty BT 4.5, it's pretty fast.

1:17:57

and it's relatively good at writing and

1:17:59

organizing information when I'm creating something, but

1:18:01

whenever I need to consume something, I'm

1:18:03

fine with it being slower. An O1-Pro

1:18:06

will take two or three minutes to

1:18:08

give me an output, deep research will

1:18:10

sometimes take 20 minutes to give me

1:18:12

an output, but I'm happy to wait

1:18:14

some time for something that's really worth

1:18:16

reading. You know what I'm making something,

1:18:19

I want it to be, the feedback

1:18:21

cycles to be fast, but if I'm

1:18:23

reading something, I'm happy to put in

1:18:25

a prompt. Go make some dinner, come

1:18:27

back, and I have some to read

1:18:29

while I eat. And then there's Claude

1:18:32

3.5 and 3.7. So these are really

1:18:34

good for writing with voice. They sound

1:18:36

the most human. Now people tell me

1:18:38

that's really good for coding and I

1:18:40

don't really use it for coding and

1:18:43

I don't really use it for that.

1:18:45

But what's funny is that I really

1:18:47

like the tables that ChatGPT produces and

1:18:49

then I like the charts that Claude

1:18:51

produces. So what I want to figure

1:18:53

out how to figure out how to

1:18:56

read that you're working on. A lot

1:18:58

of times, a chart will be super

1:19:00

useful. And like, I guess you could

1:19:02

kind of do it in Excel, but

1:19:04

you can't really use natural language in

1:19:06

the same way. You can do that

1:19:09

in Claude and have Claude make the

1:19:11

chart for you, and it's really worth

1:19:13

playing around with, because a chart can

1:19:15

just in the snap of a finger

1:19:17

make your argument for you, and you

1:19:19

can show something really clearly. So that's

1:19:22

something to think about as you're writing

1:19:24

and working on pieces. And then there's

1:19:26

deep research. Like I said, this was

1:19:28

the iPhone moment for me with AI.

1:19:30

Whenever I want an in-depth explanation of

1:19:32

something. So I was talking about the

1:19:35

flora and the fauna and Austin at

1:19:37

the beginning of spring between my office

1:19:39

and where I live. And then the

1:19:41

other day, you know, I was driving

1:19:43

on I-35, which is the highway that

1:19:45

goes through Austin. Here's exactly what I

1:19:48

prompted it with. I said, it seems

1:19:50

like there's traffic. I want a comprehensive

1:19:52

social and technological history of this road

1:19:54

and I want to know why the

1:19:56

traffic jams are so uniquely bad according

1:19:58

to data and theories of engineering and

1:20:01

road design. Now I do wait 12

1:20:03

minutes and 32 seconds for an answer,

1:20:05

but look, the writing, remember I was

1:20:07

talking about the quality, the absolute writing

1:20:09

quality was probably like a seven out

1:20:11

of 10, but it was just 10

1:20:14

out of 10 personalized. The answer that

1:20:16

it gave me was super clear, super

1:20:18

specific. And in a situation like that.

1:20:20

I'm a lot happier just reading pretty

1:20:22

good writing about a topic that's then

1:20:24

tailored to my exact interest in that

1:20:27

moment, the exact question I had at

1:20:29

that time. I bet there's a legit

1:20:31

incredible book about like Texas highway design,

1:20:33

but I don't need that. It's complete

1:20:35

overkill. You know, I just want the

1:20:37

exact answer. And also I'd have to

1:20:40

do a bunch of searching for information,

1:20:42

so I'm happy to wait 12 minutes

1:20:44

because it's 10 out of 10 with

1:20:46

being tailored to exactly what I want.

1:20:48

And then there's Grock. Grock has the

1:20:50

most personality. Grok is your crazy friend.

1:20:53

I mean look it's like Elon Musk

1:20:55

in a box, right? It's like It's

1:20:57

not as bland as the other LLLMs

1:20:59

You can that's what I used to

1:21:01

explain things to me So if I'll

1:21:03

read something. There's like a technical thing

1:21:06

that I don't quite understand I'll ask

1:21:08

it to give me a funny analogy

1:21:10

or simple explanation Just hey help me

1:21:12

understand this and then like I was

1:21:14

talking about earlier I love to argue

1:21:16

about it in voice mode. It's so

1:21:19

fun to do You have high conviction

1:21:21

on something, hey, let's just start arguing

1:21:23

with each other. And it'll just give

1:21:25

me a transcript of the entire conversation.

1:21:27

And then right at the end, this

1:21:29

is the thing you got to remember.

1:21:32

At the end, ask for that summary

1:21:34

of the best ideas from the conversation.

1:21:36

And now, I haven't tried this yet.

1:21:38

I want to do it, but Patrick

1:21:40

Hall's and the CEO of Stripe, what

1:21:42

he does is he uses Grorockwell Read,

1:21:45

so he'll turn on the voice mode

1:21:47

and say that he's reading on the

1:21:49

and there's context that he doesn't really

1:21:51

understand rather than like opening your phone

1:21:53

and then you get distracted you can

1:21:55

just ask Rock the question it'll give

1:21:58

you the answer and then you sort

1:22:00

of have a tutor by your side

1:22:02

whenever you read I really want to

1:22:04

give that a try and the thing

1:22:06

is to keep it open while you're

1:22:08

reading, Grock is really good in the

1:22:11

background. Now, I was talking about hallucinations

1:22:13

with LLLMs earlier, and that's a problem,

1:22:15

but there's a solution. And the solution

1:22:17

is perplexity. Like if you want facts...

1:22:19

and you want a quick answer and

1:22:21

then really clear and concrete sources for

1:22:24

what you're looking for, then use perplexity.

1:22:26

It's good, but their deep research tool

1:22:28

isn't really that good. ChatGPTs is the

1:22:30

best by far. Grock has a deep

1:22:32

think model, but it's only okay. Use

1:22:34

ChatGPT for the deep research. Then there's

1:22:37

11 labs. So we've done two things

1:22:39

with 11 labs. The first thing, is

1:22:41

there speech to text? Model is great.

1:22:43

So what I'll do is I'll even

1:22:45

take an MP3 file like this podcast

1:22:47

and I will put it in and

1:22:50

I'll instantly get a transcript of the

1:22:52

entire thing and it's really good. So

1:22:54

that's how I'm getting my transcripts now

1:22:56

and they're way better than the kinds

1:22:58

of transcripts that I'd get from like

1:23:00

a descriptive or a rev computer generated

1:23:03

going a year and a year and

1:23:05

a half ago. I mean it used

1:23:07

to cost me something like this would

1:23:09

cost me 150 bucks for a transcript

1:23:11

in a 24-hour turnaround. Like serious advancement

1:23:13

there and then the other thing is

1:23:16

we clone my voice We clone my

1:23:18

voice so what we did is we

1:23:20

put a bunch of training data and

1:23:22

if we ever need to We're working

1:23:24

on a video or something we ever

1:23:26

need to change a few things in

1:23:29

the video We now have cloned my

1:23:31

voice so that we can type what

1:23:33

it is that we want to say

1:23:35

and then 11 labs will just make

1:23:37

it sound like me And that's going

1:23:39

to end up changing audio books, you

1:23:42

know, if you want to change a

1:23:44

fact, you've got something wrong in an

1:23:46

audio book on page 320, you're not

1:23:48

going to have to go back to

1:23:50

the studio. You're just going to be

1:23:52

able to make a small insertion. And

1:23:55

look, it's not perfect. It's definitely not

1:23:57

perfect. Over time, like if you were

1:23:59

to listen to me speak for 10

1:24:01

minutes, but at 10 seconds, I don't

1:24:03

think you can tell. Actually, we just

1:24:05

did it. couldn't tell. Go back. See

1:24:08

if you can tell. And then there's

1:24:10

granola. I was talking about granola for

1:24:12

meeting notes. It's the first AI note

1:24:14

taker to just absolutely nail it. It's

1:24:16

super unintrusive in meetings. It auto generates

1:24:18

meeting notes based on the conversation that

1:24:21

you had. It has a transcript of

1:24:23

the entire conversation and my favorite thing

1:24:25

is that you can search that transcript.

1:24:27

And so it's not just telling you

1:24:29

what you said, but it's almost giving

1:24:31

you quick summaries of what you said.

1:24:34

So if you're We were talking about

1:24:36

some plans that we wanted to do

1:24:38

in Austin. It would basically summarize those

1:24:40

plans and turn it to an output

1:24:42

for you. And then there's Whisper Flow

1:24:44

and Super Whisper. So I don't like

1:24:47

typing. I love being able to go

1:24:49

for walks or just like walk around

1:24:51

my office as I'm writing and I'll

1:24:53

just speak things out. And you know,

1:24:55

Siri is just really annoying to use

1:24:57

because it gets so many things wrong.

1:25:00

Whisper flow and super whisper are pretty

1:25:02

darn accurate when you're just trying to

1:25:04

share things. And what's really nice is

1:25:06

I can just walk around for like

1:25:08

five minutes as I think through something.

1:25:10

I'll say 800 words worth of stuff.

1:25:13

It'll just instantly put that on the

1:25:15

screen. I can start writing from there

1:25:17

and what it's using is it's using

1:25:19

not exactly sure how that technology works,

1:25:21

but pretty advanced speech to text models

1:25:23

to basically. assume things around punctuation and

1:25:26

capitalization and it gets to know you

1:25:28

over time. For example, whenever I say

1:25:30

how I write, it capitalizes how and

1:25:32

right, whereas like Siri would make that

1:25:34

lower case and it's like 10,000 little

1:25:36

things like that that make it really

1:25:39

good. So that's how I use it.

1:25:41

That's how I use AI. And look,

1:25:43

it's just gone to the point where

1:25:45

if you were to look at my

1:25:47

life, the tower of my life, and

1:25:49

a janga block was pulled from it.

1:25:52

No more AI for you. I would

1:25:54

actually be really, really bumped. I'd be

1:25:56

really bummed at this point. AGI isn't

1:25:58

here. We haven't reached the singularity. No,

1:26:00

none of that. Like, just very pregnant.

1:26:02

I'd just be really bummed. Like this

1:26:05

stuff is fun to use. It's really

1:26:07

helpful. It's integrated into my life all

1:26:09

over the place. I'm probably doing 10

1:26:11

to 20 AI interactions per day. You

1:26:13

know, it's pretty crazy. I'm talking to

1:26:15

AI more than any single person in

1:26:18

my life. It's pretty wild to think

1:26:20

about. And, you know, here are the

1:26:22

things I'd miss. I'd miss going back

1:26:24

and forth on an LLLM whenever I

1:26:26

get stuck on a piece of writing.

1:26:28

I'd miss getting instant 80th percentile feedback

1:26:31

on my editing. You know I have

1:26:33

pretty darn good editor right now who's

1:26:35

better than the LLLM still are but

1:26:37

you know I always got to wait

1:26:39

for his response and I get instant

1:26:41

feedback with the LLLM's it can interview

1:26:44

me which is fun it has context

1:26:46

on what I'm writing about. I miss

1:26:48

totally miss being able to speak out

1:26:50

all my ideas into the chat GPT

1:26:52

app and just ask it to write

1:26:54

me. an outline and structure the ideas

1:26:57

and tell me hey you're weaker in

1:26:59

these places you're stronger in these places

1:27:01

when I hear more want you to

1:27:03

clarify this give me more about that

1:27:05

and man the thing I'd miss the

1:27:07

most is instant deep research reports on

1:27:10

whatever I'm interested on it's so cool

1:27:12

being able to do that it's so

1:27:14

cool being able to do this and

1:27:16

so if you've you've listened all this

1:27:18

you're still an AI skeptic that's my

1:27:20

number one recommendation is just Try

1:27:23

deep research and give it an earnest try.

1:27:25

Okay, give it an earnest try go in

1:27:27

and Try to write a really good prompt

1:27:30

about something specific that you're curious about where

1:27:32

you know enough about the topic that you

1:27:34

can Ask a good question and then see

1:27:36

how good the answer is I'm not tell

1:27:39

I'm not saying that it's gonna be the

1:27:41

greatest thing you've ever read But I am

1:27:43

saying it's gonna be super Taylor to your

1:27:45

interest and I think it's gonna be pretty

1:27:48

darn good. All right. I know that was

1:27:50

a lot and if you just want all

1:27:52

of that information in one place that you

1:27:54

can easily read, well I put it on

1:27:56

a PDF for you. And it's not just

1:27:59

the stuff I just shared, but I... I

1:28:01

have all of these prompts that I use

1:28:03

throughout my day, I'll give them to chat,

1:28:05

GPT, and everything is on that one PDF.

1:28:08

So if you want me to just email

1:28:10

it to you, then what you can do

1:28:12

is you can go to peril.com/AI, or you

1:28:14

can just pop down to the description. I

1:28:17

need a way to send it to you.

1:28:19

So go to that page, enter email, and

1:28:21

I'll just flip it all and over to

1:28:23

you. All right, so if you just watch

1:28:26

this, you just listen to this, you're like,

1:28:28

okay, I want more of this, I want

1:28:30

to keep going down this rabbit hole. My

1:28:32

number one recommendation is the episode I just

1:28:35

recorded with Tyler Cowan. And it's all about

1:28:37

writing with AI, but there's a few things

1:28:39

that we talked about in that episode that

1:28:41

I definitely wasn't able to talk about here.

1:28:43

How to read with AI, he speaks really

1:28:46

well about that. How AI is influencing academia,

1:28:48

why secrets are going to be more valuable

1:28:50

in an AI driven world and look. The

1:28:52

title of the episode explains it pretty well.

1:28:55

It's something like how to write with AI

1:28:57

in 68 minutes. So if you just want

1:28:59

to go check that out, you like this,

1:29:01

then I recommend going to watch that episode.

1:29:04

But also, look, there's probably.