0:00

In a world of economic

0:02

uncertainty and workplace transformation, learn

0:05

to lead by example from

0:07

visionary C-sweet executives like Shannon

0:09

Schuyler of PWC and Will

0:11

Pearson of Ihart Media. The

0:14

good teacher explains, the great

0:16

teacher inspires. Don't always leave your

0:18

team to do the work. That's

0:20

been the most important part of

0:22

how to lead by example. Listen

0:24

to Leading by Example, executives making an

0:27

impact on the I Heart Radio app,

0:29

Apple Podcasts, or wherever you get your

0:31

podcasts. Hi, I'm Bob Pittman, Chairman

0:35

Bob Pittman, chairman and CEO of I Heart

0:37

Media. I'm excited to introduce a brand new

0:39

season of my podcast, Math and Magic, stories

0:41

from the frontiers of marketing. I'm having conversations

0:44

with some folks across a wide range of

0:46

industries to hear how they reach the top

0:48

of their fields and the a

1:00

rock star is very fun, but helping people is

1:02

way more fun. And Damien Maldonado, CEO

1:04

of American financing. I figured out the

1:07

formalize have to work hard, then that's

1:09

magic. Join me as we uncover innovations

1:11

and data and analytics, the math, and

1:13

the ever important creative spark, the magic.

1:16

Listen to math and magic on the

1:18

I heart radio app, Apple podcast, or

1:20

wherever you get your podcast. I'm ready

1:22

to fight. Oh, this is fighting words.

1:24

Okay, I'll put the hammer back. Hi,

1:27

I'm George M. Johnson, a best-selling author

1:29

with the second most banned book in

1:31

America. Now more than ever, we need

1:33

to use our voices to fight back.

1:36

Part of the power of black, queer

1:38

creativity is the fact that we got

1:40

us, you know? We are the

1:42

greatest culture makers in world history.

1:45

Listen to fighting words on the

1:47

I-heart radio app, Apple podcast, or

1:49

wherever you get your podcast.

1:55

Hey y'all, it's your girl Cheekies, and I'm

1:57

back with a brand new season of your

1:59

favorite party Cheekies and Chew. I'll be

2:02

sharing even more personal stories with

2:04

you guys and as always you'll

2:06

get my exclusive take on topics

2:08

like love, personal growth, health, family

2:10

ties, and more. And don't forget

2:12

I'll also be dishing out my

2:14

best advice to you on episodes

2:16

of Dear Cheekies. It's going to

2:18

be an exciting year and I

2:20

hope that you can join me.

2:22

Listen to Cheekies and Chell season

2:24

4 on the iHeart Radio app,

2:26

Apple Podcast or wherever you get

2:28

your podcasts. Soul of an angel,

2:30

body of a devil, chosen by

2:32

God and perfected by science, this

2:34

is better offline and I'm your

2:36

host, etc. Now we're working on

2:38

my newsletter last week. Now we're

2:40

working on my newsletter last week,

2:42

I was chatting with my friend,

2:44

friend of the show, Casey Cagawa,

2:46

about generative AI, and we kept

2:48

coming back to one thought. Where's

2:50

the money? Where is it? No,

2:52

really, where is the money? Where

2:54

is the money that this supposedly

2:56

revolutionary world-changing industry is making and

2:58

of course will make in the

3:00

future? And the answer is simple.

3:02

After hours of hours of grinding

3:04

through earnings of grinding through media

3:07

articles of grinding through all sorts

3:09

of things. I just don't believe

3:11

it really exists. It's real, but

3:13

it's small. Generative AI lacks the

3:15

basic unit economics, product market fear,

3:17

or market penetration associated with any

3:19

meaningful software boom, and outside of

3:21

open AI, the industry may be

3:23

pathetically hopelessly small, all while providing

3:25

few meaningful business returns and constantly

3:27

losing money. I'm going to be

3:29

pretty straightforward with everything I say

3:31

in this two-parta, because the numbers

3:33

and the facts in my hypothesis

3:35

are pretty fucking damning of both

3:37

the generative AI industry and its

3:39

associated boosters boosters. You're going to

3:41

get this episode, then there's going

3:43

to be a monologue about something

3:45

else or something related. I really

3:47

haven't got to it yet. And

3:49

then a second part, which I'm

3:51

recording immediately after this one. Little

3:53

behind the curtain there for you.

3:55

Anyway, in reporting this analysis, I've...

3:57

I've done everything I can to

3:59

try and push back against my

4:01

own work, and I've saw evidence

4:03

to counter the things that I've

4:05

seen, like the revenue and the

4:07

business models of these companies. Yet in

4:09

doing so, I've only become more convinced

4:11

of the flimsyness of generative AI in

4:14

the associated industry, and the likelihood of

4:16

this bubble bursting in a way that

4:18

kneecaps take valuations for a prolonged period,

4:20

or worse, hits the major stock market. Now

4:22

I really had originally written a far

4:24

more jocular and outraged and pissy script

4:26

but... While I was writing it, I

4:29

realized I really had to be blunt,

4:31

because what I'm describing is a systemic

4:33

failure. Venture Capital has propped up

4:35

open AI and anthropic, two companies

4:37

that have burned a combined $10.5

4:39

billion in 2024, and that number

4:41

is set to double or more

4:43

in 2025. The tech media has allowed

4:46

Sam Altman to twist them, to

4:48

validate completely fictional ideas, as a

4:50

means of propping up this unprofitable

4:52

environmentally destructive software company. And Big

4:54

Tech has become so disconnected from

4:56

reality that it is incapable of

4:59

seeing how little actual returns there

5:01

are in generative AI. And they're failing,

5:03

by the way. As I'll walk you through

5:05

in these episodes, the generative AI

5:07

industry is very small, with the

5:09

consumer market of the entire American

5:11

generative AI industry outside of chat-GPT

5:13

barely cracking 100 million monthly active

5:15

users, which puts them below a

5:17

lot of free-to-play games that you

5:19

get on your iPhone. Hyperscalers

5:22

have already spent hundreds of billions

5:24

of dollars in capital expenditures for

5:26

an AI industry that has the

5:28

combined monthly active users of a

5:30

free to play mobile game. I

5:32

really must repeat myself. It's insane.

5:34

But unlike most mobile games, generative

5:36

AI doesn't really make any money.

5:38

And for those of you wondering if selling

5:40

access to AI models is the solution, it's

5:43

important to know that Open AI, the market

5:45

leader in generative AI, made less than a

5:47

billion dollars on API calls in 2024. And

5:49

that's when people plug their models in for

5:51

those of you who don't understand, so it's

5:54

the difference between you to load up the

5:56

chat GBT app or someone has an AI,

5:58

a generative AI like chat. GPT Ryan

12:53

Reynolds here for Mint Mobile. I don't

12:55

know if you knew this, but anyone

12:57

can get the same premium wireless for

12:59

$15 a month planned that I've been

13:01

enjoying. It's not just for celebrities, so

13:03

do like I did, and have one

13:05

of your assistance assistance to switch you

13:07

to Mint Mobile today. I'm told it's

13:09

super easy to do at mintmobile.com slash,

13:12

switch. Up front payment of $45 for

13:14

three-month plan equivalent to $15 per month

13:16

required. Intro rate for three months only.

13:18

Then full price plan options available. Taxes

13:20

and fees extra. Feeful terms at mintmobile.com.

13:26

Okay, we're back. Rory, the

13:28

Menesas incident might be the

13:30

most famous UFO case in

13:32

Spanish history, but it's far

13:34

from the only case. In

13:36

1965, an utterly insane tale

13:38

emerged from a military site

13:40

in Bandahos. I don't know

13:43

how to say that Bandahos.

13:45

But, sorry, Bados. Badahos. That

13:47

sounds like a German man

13:49

saying that something's bad-ass. It's

13:51

like, oh, that's bad. Let's

16:24

really get down into the nitty gritty

16:26

of these numbers. So as discussed previously,

16:29

according to the reporting by the information,

16:31

Open AI's revenue was likely somewhere in

16:33

the region of $4 billion in 2024.

16:35

The burn rate, according to the information,

16:37

was $5 billion after revenue in 2024,

16:40

excluding stock-based compensation, which Open AI, like

16:42

other startups, uses as a means of

16:44

compensation on top of cash. Nevertheless, the

16:46

more it gives away, the less it

16:48

has for capital raises, and these are

16:51

technically costs, though they're not real money.

16:53

unless there's a liquidity event, but you

16:55

don't need to worry about that. To

16:57

put this in blunt terms, based on

16:59

reporting by the information, and I'm repeating

17:02

myself here, but I really need you

17:04

to remember this, running open AI costs

17:06

$9 billion in 2024. The cost of

17:08

the computer To train models, to train

17:10

models alone, $3 billion, obliterates the entirety

17:13

of their subscription revenue, which is about

17:15

$3 billion, by the way. And the

17:17

compute from running models, $2 billion, takes

17:19

the rest and then some. They actually

17:21

end up losing an extra billion on

17:24

top of that. Sam Altman's net worth

17:26

is a billion dollars, by the way.

17:28

Casey Gagawa has now used this as

17:30

the Altman index, so it's like, you've

17:32

lost one Sam Altman, that's a billion

17:35

dollars. But just to be clear, it

17:37

doesn't just cost more to run Open

17:39

AI than they make. It costs them

17:41

a billion dollars more than the entirety

17:43

of their revenue to run the software

17:46

they sell before any other costs. Why

17:48

are we not more concerned about this

17:50

company? Now something else to note is

17:52

that Open AI also spends an alarming

17:54

amount of money on salaries. Over 700

17:57

million dollars in 2024 before you consider

17:59

that compensation. from stock. A number that

18:01

will also have to increase because Open

18:03

AI is growing, which means hiring as

18:06

many people as possible, and they're paying

18:08

through the nose for them. But let's

18:10

talk about how Open AI makes money.

18:12

Open AI sells access to its models

18:14

via its API and selling premium subscriptions

18:17

to ChatGPT. The majority of its revenue

18:19

over 70% comes from subscriptions to premium

18:21

versions of ChatGPT. The information also reported

18:23

that Open AI now has 15.5 million

18:25

paying customers, though it's unclear what level

18:28

of the service they're paying for, or

18:30

how sticky these customers are as in

18:32

how likely they are to stick around,

18:34

or the cost of acquiring these customers,

18:36

or really any other metric to tell

18:39

them how valuable these customers are to

18:41

the bottom line. Nevertheless, Open AI loses

18:43

money on every single paying customer, just

18:45

like its free users. Increasing paid subscribers

18:47

to a... Open AI services somehow increases

18:50

Open AI's burn rate. This is not

18:52

a real company. Now the New York

18:54

Times reports that Open AI projects it

18:56

will make $11.6 billion in 2025 and

18:58

assuming that Open AI burns at the

19:01

same rate it did in 2024, spending

19:03

$2.25 to make $1, Open AI is

19:05

on course the burn over $26 billion

19:07

in 2025 for a loss of $14.4

19:09

billion. Who knows what their actual cost

19:12

will be? Now you've probably heard about

19:14

soft bank coming in, soft bank's going

19:16

to feed the money and soft bank

19:18

said they're going to spend money on

19:20

this, that and the other. That round

19:23

has not closed yet. Masayoshi's son, a

19:25

complete fucking idiot who's lost thirty odd

19:27

billion dollars to soft bank, the... Japanese

19:29

mega conglomerate. He's dedicating billions of dollars

19:31

of revenue to buying open AI services.

19:34

Unless this is a straight up trade

19:36

where he's just sending money before the

19:38

services come in, I don't know if

19:40

it happens. And I'm going to get

19:42

into things like agents later, but the

19:45

information reported that open AI expects to

19:47

make $3 billion in revenue from agents.

19:49

By the end of this episode, you're

19:51

going to realize how fucking stupid that

19:53

sounds. We'll get there later. It's also

19:56

important to note that open AI's cost

19:58

a partially subsidizedized by its relation by

20:00

its relation. with Microsoft, which provides cloud

20:02

compute credits for its azure cloud service.

20:04

Not super technical, it's just when they

20:07

host people's software and files and such,

20:09

and the compute to run these models.

20:11

And they also offer this a steep,

20:13

steep discount to Open AI. Or put

20:15

another way, it's like Open AI got

20:18

paid with air miles, but the airline

20:20

lowered the redemption cost of booking a

20:22

flight with those air miles, allowing it

20:24

to take more flights than any other

20:26

person with the equivalent amount of points.

20:29

Until recently, OpenAI exclusively used Microsoft as

20:31

your services to train, host, and run

20:33

its models. But recent changes to its

20:35

deal means that OpenAI is now working

20:37

with Oracle to build up further data

20:40

centers to train, host, and run its

20:42

models. It's unclear whether this partnership will

20:44

work in the same way as the

20:46

Microsoft deal, with OpenAI provided credits and

20:48

discounts like before. If not, OpenAI's operating

20:51

costs will only go up. Per previous

20:53

reporting from the information, open AI pays

20:55

just over 25% of the cost of

20:57

a Zios GPU compute as part of

20:59

their deal with Microsoft. And that's about

21:02

a dollar 30 per GPU per hour

21:04

versus the regular a Zio cost of

21:06

$3.40 to $4 an hour. I know

21:08

that this sounds really technical, but in

21:10

very short, they're getting a sweet deal

21:13

from Microsoft and if anything happens then

21:15

they're completely fucked anyway. They're burning billions

21:17

of dollars. It's insane. Let's talk about

21:19

user numbers because Open AI has quite

21:21

a few. They recently announced that they

21:24

have 400 million weekly active users. Now

21:26

Weekly Active Users is a wanky number

21:28

and a very strange one for a

21:30

company like this. Open AI may pretend

21:32

to be a consumer company, but the

21:35

majority of their revenue comes from monthly

21:37

subscriptions, making them kind of a cloud

21:39

software company. Classically, cloud software companies report

21:41

monthly active users. That way you can,

21:43

I don't know. Compare one number, which

21:46

is the amount of active users you

21:48

have, with the paid users you have,

21:50

and then say, oh, that's a good

21:52

business. That's a good business right there,

21:54

man. Guess what? Open AI isn't giving

21:57

them monthly active users. Don't worry. I

21:59

might have estimated it. When I asked

22:01

Open AI to define what a weekly

22:03

active user was, it responded by... I'm

22:05

pointing me to a tweet by Chief

22:08

Operating Officer Brad Lightcap that said ChatGPT

22:10

recently crossed 400 million weekly active users.

22:12

We feel very fortunate to serve 5%

22:14

of the world every week. What a

22:17

fucking liar. It's extremely questionable that Open

22:19

AI refuses to define this core metric,

22:21

by the way. And without a definition

22:23

in my opinion, there is no way

22:25

to assume anything other than Open AI

22:28

is actively gaming its numbers. Now there's

22:30

likely two reasons they focus on weekly

22:32

active users. One, as described, these numbers

22:34

are easy to game. You can choose

22:36

any seven day period. And also the

22:39

majority of OpenAIs revenue comes from paid

22:41

subscriptions to chatGPT. And that latter point

22:43

is crucial because it suggests OpenAIs not

22:45

doing anywhere near as well as it

22:47

seems, based on the very basic metrics

22:50

used to measure the success of a

22:52

software product. The information reported on January

22:54

31st, OpenAi, like I mentioned, had 15.5

22:56

million monthly paying subscribers, and they added

22:58

in this piece that this was less

23:01

than a 5% conversion rate of OpenAIs'

23:03

weekly active users. A statement that's kind

23:05

of like dividing the number 52 by

23:07

the letter A. This is not an

23:09

honest or reasonable way to evaluate the

23:12

success of ChatGPT's still unprofitable software business,

23:14

because the actual metric, like I mentioned,

23:16

would have been to divide paying subscribers

23:18

by monthly active users, or the other

23:20

way around, I guess, a number that

23:23

would be considerably higher than 400 million.

23:25

And the reason they don't want you

23:27

to do that, by the way, is

23:29

because you would divide them five percent,

23:31

by the way. And... There's definitely lower.

23:34

But don't worry, I'm a sneaky little

23:36

shit, so I went and looked some

23:38

stuff up and I talked to some

23:40

people. Based on data from the market

23:42

intelligence firm Sensor Tower, Open AI's ChatGPT

23:45

app on Android and iOS is estimated

23:47

to have more than 339 million monthly

23:49

active users. And based on traffic data

23:51

for Market Intelligence Company, similar web, chatgpt.com

23:53

had 246 million unique monthly visitors, and

23:56

these were in January 2025. There's likely

23:58

some crossover with people using both the

24:00

mobile and web interfaces, though how big

24:02

that group beers is... kind of hard

24:04

to tell and remains uncertain. Though every

24:07

single person that visits chatgpt.com might not

24:09

become a user, it's safe to assume

24:11

that chatGPT's monthly active users are somewhere

24:13

in the region of 500 to 600

24:15

million. That's good, right? It's actual users

24:18

are higher than officially claimed. Right? That's

24:20

good, right? It's actual users are higher

24:22

than officially claimed. Right? But either free

24:24

ones definitely are. It would also suggest

24:26

that the real conversion rate is somewhere

24:29

in the neighborhood... but of 2.583% from

24:31

freed paid users on chatGPT, which is

24:33

astonishingly bad. And it's a fact that's

24:35

made worse by the fact that every

24:37

single user regardless of whether they pay

24:40

or not loses the money. Either way.

24:42

And while it's quite common for Silicon

24:44

Valley companies to play fast and loose

24:46

with metrics, this particular one is... Well,

24:48

it's deeply concerning, and I hypothesize that

24:51

Open AI is choosing to go with

24:53

weekly versus monthly active users in an

24:55

intentional attempt to avoid people calculating the

24:57

conversion rate of its subscription products. As

24:59

I will continue to repeat, these subscription

25:02

products lose the company money every single

25:04

time. Now let's talk product strategy, shall

25:06

we? Because I don't think Open AI

25:08

really has one. Open AI makes most

25:10

of its money from subscriptions, approximately $3

25:13

billion in 2024, and the rest on

25:15

API access to its models, approximately a

25:17

billion. As a result, Open AI has

25:19

chosen to monetize ChatGBT and its associated

25:21

products in an all-you-can-eat software subscription model,

25:24

or otherwise make money by other people

25:26

productizing it. And just to be clear,

25:28

in both of these scenarios, Open AI

25:30

loses money on every transaction. OpenAIs products

25:32

are not fundamentally differentiated or interesting enough

25:35

to be sold separately. It has failed,

25:37

as with the rest of the generative

25:39

AI industry, to meaningfully productize its models

25:41

due to the massive training and operational

25:43

costs and a lack of any meaningful

25:46

killer app use cases for large language

25:48

models. The only product that OpenAI has

25:50

succeeded in scaling to the mass market

25:52

is the free version of chatGBT, which

25:54

loses the company money with every single

25:57

prompt and output. This scale isn't a

25:59

result of any kind of... product market

26:01

fit by the way. It's entirely media

26:03

driven with reporters making chat GPT synonymous

26:05

with artificial intelligence, a thing they regularly

26:08

write about without thinking. As a result

26:10

I do not believe that the generative AI industry

26:12

is real. It's not a real industry which

26:14

I will define as one with multiple

26:17

competitive companies with sustainable or otherwise growing

26:19

revenue streams and meaningful products with actual

26:21

market penetration. And I feel this way

26:24

because it... This market is entirely subsidized

26:26

by a combination of venture capital and

26:28

hyperscaler cloud credits, and, well, real money,

26:31

I guess. ChatGPT is popular because it's

26:33

the only well-known product, one that's mentioned

26:35

in basically every article in AI. If

26:38

this were a real industry, other competitors

26:40

would also be mentioned all the time.

26:42

They would have similar scale, especially

26:44

those run by hyperscalers, but as

26:46

I'll get to later, data suggests

26:48

that open AI is the only

26:51

company with any significant user base

26:53

in the entire generative AI industry,

26:55

and it's still wildly unprofitable and

26:57

unsustainable. Open AI's models have also

26:59

been entirely commoditized. Even its reasoning

27:01

model, O1, has been commoditized by

27:04

both Deep Sikhs R1 model and

27:06

perplexities agonizingly named R1 1776 model,

27:09

both of which have similar outcomes

27:11

at a much discounted price to

27:13

OpenAIs O1, though it's unclear and

27:16

unlikely in my opinion that these

27:18

models are profitable anyway. Open AI

27:20

as a company, well they just pissed

27:22

poor at product. It's been two years

27:24

and chat-GPT mostly does the same thing,

27:26

still costs more to run than it

27:29

makes, and ultimately does the same thing

27:31

as every other LLLM chatbot from every

27:33

other company. The fact that nobody has

27:35

managed to make a mass market product

27:37

by connecting open AI's models also suggests

27:39

that the use cases just aren't there.

27:41

Furthermore, the fact that API access is

27:43

such a small... part of its revenue

27:45

suggests that the market for actually implementing

27:47

large language models is relatively small. If

27:49

the biggest player in the space only

27:51

made a billion dollars in selling access

27:53

to its models unprofitably, and that amount

27:56

is the minority of its revenue, there might

27:58

not actually be a real industry here. And

28:00

I must be clear, if there was

28:03

user demand, this would be where it

28:05

was in the APIs. It would be

28:07

doing gangbusters, because people wouldn't be able

28:09

to help themselves, they'd just be all

28:12

over this generative AI share. But they're

28:14

not. Hey

28:25

kids, it's me, Kevin Smith! And it's

28:27

me, Harley Quinn Smith. That's my daughter,

28:29

man, who my wife has always said

28:31

is just a beardless, deakless version of

28:33

me. And that's the name of our

28:35

podcast. Beardless, deakless me. I'm the old

28:37

one. I'm the young one. And every

28:40

week we try to make each other,

28:42

laugh. And every week we try to

28:44

make each other, laugh, laugh with me.

28:46

I'm the old one. I'm the young

28:48

one. And every week, and every week,

28:50

and every week, we try to make,

28:53

we try to make, we try to

28:55

make, we try to make, we try,

28:57

we try, we try, we try, we

28:59

try, we try. We're, we try. We're,

29:01

we're, we're, we're, we're, we're, we're, we're,

29:03

we're, we're, we're, we're, we're, we're, we're,

29:05

we're, we're, I You get your podcast.

29:08

Hey y'all, it's your girl Cheekies and

29:10

I'm back with a brand new season

29:12

of your favorite podcast Cheekies and Chew.

29:14

I'll be sharing even more personal stories

29:16

with you guys. And I know a

29:18

lot of people are gonna attack me.

29:21

Why are you gonna go visit your

29:23

dad? Your mom wouldn't be okay with

29:25

it. I'm gonna tell you guys right

29:27

now. I know my mom had a

29:29

very forgiving heart. That is my story

29:31

on plastic surgery. This is my truth.

29:33

I think the last time I cried

29:36

like that was when I lost my

29:38

mom, like that, like yelling. I was

29:40

like, no. I was like, oh, and

29:42

I thought, what did I do wrong?

29:44

And as always, you'll get my exclusive

29:46

take on topics like love, personal growth,

29:49

health, family ties, and more. And don't

29:51

forget, I'll also be dishing out my

29:53

best advice to you on episodes of

29:55

dear cheekies. So my fiance and I

29:57

have been together for 10 years in

29:59

the first two years of being together

30:02

I find out he is cheating on

30:04

me not only with the women, but

30:06

also with men. What should I do?

30:08

Okay, where do I start? That's not

30:10

love. He doesn't love you enough because

30:12

if he loved to you, he'd be

30:14

faithful. It's going to be an exciting

30:17

year and I hope that you can

30:19

join me. Listen to Cheekies and Chill,

30:21

season four as part of the Michael

30:23

Dura podcast network available on the iHeart

30:25

Radio app, Apple Podcasts, or wherever you

30:27

get your podcasts. Welcome to Paud of

30:30

Rebellion, our new Star Wars Rebels rewatch

30:32

podcast. I'm Vanessa Marshall. Hi, I'm Tia

30:34

Sercar. I'm Taylor Gray. And I'm John

30:36

Lee Brody. But you may also know

30:38

us as Harrison Dula's Specter too. To

30:40

being ran, Specter 5. And Ezra Bridger,

30:42

Specter, Specter 6 from Star Wars Rebels.

30:45

Wait, I wasn't on Star Wars Rebels,

30:47

am I in the right place? Absolutely.

30:49

guests like Steve Bloom voices Zaborellio, Specter

30:51

4 or Dante Bosco voice of Jai

30:53

Kel and many others. Sometimes we'll even

30:55

have a lively debate. And we'll have

30:58

plenty of other fun surprises in tribute

31:00

to. Oh, and me? Well, I'm the

31:02

lucky ghost crew, Stowaway, who gets to

31:04

help moderate and guide the discussion each

31:06

week. Kind of like how Canaan guided

31:08

Ezra in the ways of the force.

31:10

You see what I did there? Listen

31:18

to Potter Rebellion on the I-Hart radio

31:20

app, Apple podcast, or wherever you get

31:23

your podcast. My name is Brendan Patrick

31:25

Hughes, host of divine intervention. This is

31:27

a story about radical nuns in combat

31:29

boots and wild-haired priests, trading blows with

31:31

Jay Edgar Hoover in a hell-bent effort.

31:34

to sabotage a war. Jayagahouhova was furious.

31:36

Somebody violated the FBI and he wanted

31:38

to bring the Catholic left to its

31:40

knees. The FBI went around to all

31:42

their neighbors and said to them, do

31:45

you think these people are good Americans?

31:47

It's got heist tragedy, a trial of

31:49

the century, and the god-damnest love story

31:51

you've ever heard. I picked up the

31:53

phone and my thought was this is

31:56

the most important phone call I'll ever

31:58

make in my life. I couldn't believe

32:00

it. I mean, Brendan, it was divine

32:02

intervention. Listen to divine intervention on the

32:05

i-heart radio app, Apple Podcasts, or wherever

32:07

you get your podcasts. Some might argue

32:09

that Open AI has a new series

32:11

of products that could open up new

32:13

revenue streams, such as operator, it's agent

32:16

product, and deep research their research product.

32:18

And I'm so fucking tired of hearing

32:20

about agents. Whenever you hear someone say

32:22

agent, really look at what they're saying,

32:24

because they want you to think autonomous

32:27

bit of software. What they're actually talking

32:29

about is either a chat bar, or,

32:31

well, the dog shit, the Open AI

32:33

and Anthropic and Anthropic have warmed up.

32:35

which will get to shortly. But first,

32:38

let's talk costs. Both of these products

32:40

are very compute-intensive. operator uses open AI's

32:42

computer-using agent, the CUA, which combines open

32:44

AI's models with virtual machines that take

32:47

distinct actions on web pages in this

32:49

extremely unreliable and costly way, where they

32:51

take screenshots as they scroll down. And

32:53

it just doesn't fucking work. I had

32:55

a whole thing about Casey Newton writing

32:58

about this. It's just so bad. Like,

33:00

Casey Newton, please go outside challenge. Just...

33:02

Just go outside Casey, stop with the

33:04

compute, you don't know what you talked

33:06

about. But failures with these, and remember

33:09

these models, pretty much all of them,

33:11

are inconsistent, and the more in-depth the

33:13

thing you ask them to do, the

33:15

more likely there's going to be a

33:17

problem with it. So think about it

33:20

like this, failures from something you've asked

33:22

them to do will either increase the

33:24

amount of attempts you make to get

33:26

the thing you want, or make users

33:29

not use it at all. Not a

33:31

really great idea. Now let's talk deep

33:33

research. They use a version of Open

33:35

AI's O3 reasoning model, which is a

33:37

model so expensive because it spends more

33:40

time to generate a response based on

33:42

the model reconsidering and evaluating steps as

33:44

it goes. The Open AI will no

33:46

longer launch O3 as a standalone model.

33:48

And that's really a good thing. when

33:51

you see a company be like, yeah,

33:53

you can't touch it, it's too expensive.

33:55

In short, these products are extremely expensive

33:57

to run, and this means that any

33:59

time their outputs aren't perfect, which is

34:02

to say a lot of the time,

34:04

there's a high likelihood that they'll be

34:06

triggered again, which will in turn spend

34:08

more compute. But let's talk about the

34:11

product market fit, because this is really

34:13

important. To use operator or deep research,

34:15

currently requires you to pay $200 a

34:17

month subscription. Sam Altman recently revealed still

34:19

loses the money because people are using

34:22

it more than expected and that is

34:24

a quote. Furthermore, even on chatGPT Pro,

34:26

deep research is currently limited to 100

34:28

queries per month, adding that it is

34:30

very compute intensive and slow. Though Altman

34:33

has promised the chatGPT plus and free

34:35

users will eventually get access to a

34:37

few deep research queries a month, Well,

34:39

that's not good for their cashburn. That's

34:42

actually bad for the cashburn. I'm not

34:44

sure it's going to make them... I'm

34:46

not really sure how that turns into

34:48

money anywhere. But let's talk about operator.

34:50

operator is this agent product where you're

34:53

meant to be able to be like,

34:55

hey, look, go and look, something up

34:57

for me, and it only works like

34:59

30% at the time, and it's just

35:01

very bad. And as I covered in

35:04

my newsletter a few weeks ago, this

35:06

product, and it claims to control your

35:08

computer, prime time, and I don't think

35:10

it has a market. The way they're

35:12

selling this is that you'll be able

35:15

to make it do distinct tasks on

35:17

the computer, but even Casey Newton and

35:19

his article was like, yeah, it only

35:21

works sometimes, and the things it works

35:24

on are like searching trip advisor. Imagine

35:26

this, if you will. What if for

35:28

the cost of boiling a lake and

35:30

throwing an entire zoo into the lake

35:32

and boiling the animals inside it, you

35:35

could sometimes be able to search trip

35:37

advisor in two minutes versus ten, like

35:39

five seconds. The future is so cool

35:41

I love living in it. But let's

35:43

talk about deep research for a second.

35:46

It's already been commoditized. Proplexed AI and

35:48

XAI have launched their own versions immediately,

35:50

and deep research itself is... not a

35:52

good product. As I covered in my

35:54

newsletter last week, the quality of the

35:57

writing that you receive from Deep Research

35:59

is really piss poor. And it's rivaled

36:01

only by the appalling quality of its

36:03

citations, which include forum posts and search

36:06

engine optimized content instead of actual news

36:08

sources. These reports are neither deep nor

36:10

well researched, and cost-open AI a great

36:12

deal of money to deliver. And just

36:14

to give you a primal deep research

36:17

is meant to be, you're meant to

36:19

be able to type something in, and

36:21

it does like a 3,000 word word

36:23

report. I really, if you should go

36:25

and look up, go to my newsletter,

36:28

Where's Your Ed.at, it's the, it's the,

36:30

it's the, it's the piece before the

36:32

ones that's going to come out, when

36:34

these episodes come out. I forget the

36:36

name exactly. You need to go and

36:39

look at how shit deep researchers, it's

36:41

come out. You need to go and

36:43

look at how shit deep researches. It's

36:45

incredible that this money losing juggernaut piece

36:48

of shit, thinks that this is a

36:50

real product, and don't work very well.

36:52

Let's talk about how they make money.

36:54

Let's talk about how they make money.

36:56

How they make money. How they make

36:59

money. Or don't. Both operate in deep

37:01

research, like I told you, currently require

37:03

you to pay $200 a month to

37:05

a company that loses money all the

37:07

time, that also loses money on the

37:10

$200 a month. Neither product is sold

37:12

in its own, and while they may

37:14

drive revenue to the ChatGPT Pro product,

37:16

as said before, said product loses open

37:18

AI money. These products are also compute

37:21

intensive and have questionable outputs, making each

37:23

prompt very likely to create another follow-up

37:25

prompt. And the problem is you're asking

37:27

something that doesn't know anything that probabilistically

37:30

generates answers to research something. So as

37:32

a result, the research isn't going to

37:34

be any good. It's not like it's

37:36

going to research it and go, hey,

37:38

what would be a good source? It's

37:41

going to say, what matches the patterns?

37:43

What matches all the patterns that I've

37:45

been trained on? Ah, that's fine. Who

37:47

gives a ship? It's like having the

37:49

world's worst intern except the intern gets

37:52

a concussion every 10 minutes. But in

37:54

summary, both operator and deep research are

37:56

expensive products to maintain, assault through an

37:58

expensive $200 a month subscription that, like

38:00

every other service provided by Open AI,

38:03

loses the company money and due to

38:05

the low quality... their outputs and actions

38:07

are likely to increase user engagement to

38:09

try and get the desired output, incurring

38:12

further costs for Open AI. Well, you

38:14

know, like Ed, Ed, you say, Ed,

38:16

you're just being, you're just being a

38:18

hater, right? Just being a hater, things

38:20

don't look great today, but this early

38:23

days, it isn't early days, but still,

38:25

Ed, it's early days, things don't look

38:27

great today. What about the future prospects

38:29

for Open AI? Prospects, prospects for Open

38:31

A. Things can't be that bad prospects

38:34

for Open AI. A week or two

38:36

ago, Sam Altman announced the updated roadmap

38:38

for GPT 4.5 and GPT5. Now these

38:40

are their next generation models that have

38:42

been hyping up for the best part

38:45

of a year. Except GPT 4.5 didn't

38:47

exist before. It was always GPT 5.5.

38:49

Now GPT 4.5 didn't exist before. It

38:51

was always GPT 5.5. Now GPT 4.5

38:54

will be open AI's last chain of

38:56

thought model of its reasoning. Yes. GPT5

38:58

will be, and I quote Sam Altman,

39:00

a system that integrates a lot of

39:02

open AI technology, including O3. What the

39:05

fuck are you talking about? Orman also

39:07

vaguely suggests that paid subscribers will be

39:09

able to run GPT5 at a higher

39:11

level of intelligence, which likely refers to

39:13

being able to ask the models to

39:16

spend more time computing an answer. He

39:18

also suggests that the GPT5, and I

39:20

quote, will incorporate voice, canvas, search, deeper

39:22

research, deeper search, and more. Fucking bedbath

39:24

beyond, motherfucker. Come My man, your company

39:27

spent $9 billion to lose $5 billion.

39:29

Why is anyone taking this seriously? This

39:31

is ridiculous. But both of these statements,

39:33

all of these statements honestly, vary from

39:36

vague to meaningless. But I hypothesize the

39:38

following. GPT-4.5 will be an upgraded version

39:40

of GPT-40. Open AIs Foundation model you're

39:42

probably using right now. And it's code

39:44

named Orion. Eryan. could literally be anything.

39:47

But one thing that Altman mentioned in

39:49

the tweet is that open AI's model

39:51

offerings have got too complicated. They'd be

39:53

doing a... with the ability to pick

39:55

what model you used. Gossying this up

39:58

and he's claiming it's unified intelligence. This

40:00

fucking guy. If I said this shit

40:02

to a doctor, they'd institutionalize me. They'd

40:04

say you sound like a lunatic. But

40:06

anyway, as a result of doing away

40:09

with the model picker, which is literally

40:11

the thing you click and you choose

40:13

GPT-400 or GPT-400-many or like the I-1

40:15

reasoning things, I think they're going to

40:18

attempt to attempt to moderate to moderate

40:20

cost by picking. If there's one thing

40:22

I've noticed with Open AI, they're not

40:24

very good at automating anything. So I

40:26

expect this to be bad. And I

40:29

believe that Altman announcing these things is

40:31

a very bad omen for Open AI.

40:33

Because Orion has been in the works

40:35

for more than 20 months and was

40:37

meant to be released at the end

40:40

of 2024, but it was delayed due

40:42

to multiple training runs that resulted in,

40:44

to quote the Wall Street Journal, software

40:46

that fell short of the results researchers

40:48

were hoping for. As on the side,

40:51

the Wall Street Journal refers to Orion

40:53

as GPT-5, this was from several months

40:55

back, but based on the copy in

40:57

Altman's comments, I believe Orion refers to

41:00

a foundation model. Open AI, which is

41:02

one to replace the core GPT, one

41:04

that powers ChatGPT. Open AI now appears

41:06

to be calling a hodgepodge of different

41:08

mediocre models, something called GPT-5. It's almost

41:11

as if Altman's making this up as

41:13

he goes along. Now the journal further

41:15

adds that as of December Orion performed

41:17

better than Open AI's current offerings but

41:19

hadn't advanced enough to justify the enormous

41:22

costs of keeping the new model running.

41:24

With each six-month long training run, no

41:26

matter how well it works, costing over

41:28

$500 million. Open AI also, like every

41:31

generative AI companies, running out of high-quality

41:33

training data, the data necessary to make

41:35

its models smarter based on the benchmarks

41:37

specifically made up to make LLLM seem

41:39

smart, And I should note that being

41:42

smarter means completing tests not new functionality

41:44

or new things that it can do.

41:46

Sam Orpman deputizing Orion from GPT5 to

41:48

GPT4.5 suggests that Open AI has hit

41:50

a wall with making its new model,

41:53

requiring him to lower expectations. for a

41:55

model open AI Japan president, Tagao Nagasaki

41:57

had suggested would, and I quote, aim

41:59

for 100 times more computational volume than

42:01

GPT-4, which some took to mean 100

42:04

times more powerful when it actually means

42:06

it will take way more computation to

42:08

train or run inference on it. I

42:10

guess he was right. Now, if Sam

42:13

Altman, who is a man who loves

42:15

to lie, is trying to reduce expectations

42:17

for a product, I think we should

42:19

all be really, really, really worried. Now

42:21

large language models which are trained by

42:24

feeding the massive amounts of training data

42:26

and then reinforcing their understanding through further

42:28

training runs are hitting the point of

42:30

diminishing returns. In simple terms, the quote,

42:32

friend of the show Max Zef of

42:35

Tech Crunch, everyone now seems to be

42:37

admitting you can't just use more compute

42:39

and more training data with pre-training large

42:41

language models and expect them to turn

42:43

into some all-knowing digital god. Max is

42:46

a fucking legend. capture the entire tech

42:48

media, has been its relationship with Microsoft,

42:50

because access to large amounts of compute

42:52

and capital allowed it to corner the

42:55

market for making the biggest, most hugest

42:57

large language model. Now that it's pretty

42:59

obvious this isn't going to keep working,

43:01

Open AI is scrambling, especially now deep

43:03

seekers commoditized reasoning models, and prove that

43:06

you can build LLLMs without the latest

43:08

GPUs. It's unclear what the functionality of

43:10

GPT4 or GPT5 or GPT5 or GPT5

43:12

will be. Does the market care about

43:14

an even more powerful large language model

43:17

if said power doesn't do anything new

43:19

or lead to a new product? Does

43:21

the market care if unified intelligence just

43:23

means stapling together various models to produce

43:25

more outputs that kind of look and

43:28

sound the same? As it stands, Open

43:30

AI has effectively no moat beyond its

43:32

industrial capacity to train large language models

43:34

and its presence in the media. Open

43:37

AI can have as many users as

43:39

it wants, but it doesn't matter because

43:41

it loses billions of dollars and appears

43:43

to be continuing to follow the money

43:45

losing large language model paradigm, guaranteeing or

43:48

lose billions of dollars more if they're

43:50

allowed to. This is the biggest player

43:52

in the generative AI industry, both the

43:54

market leader and the recipient of almost

43:56

every... every single dollar of revenue that

43:59

this industry generates. They have received more

44:01

funding and more attention than any startup

44:03

in the last few years, and as

44:05

a result, their abject failure to become

44:07

a sustainable company with products that truly

44:10

matter is a terrible sign for Silicon

44:12

Valley and an embarrassment to the tech

44:14

media. In the next episode, I'm going

44:16

to be honest, I have far darker

44:19

news. Based on my reporting, I believe

44:21

that the generative AI industry outside of

44:23

open AI is incredibly small, with little

44:25

to no consumer adoption, and pathetic amounts

44:27

of revenue compared to the hundreds of

44:30

billions of dollars sunk into supporting it.

44:32

This is an entire hype cycle fueled

44:34

by venture capital and big tech hubris,

44:36

with little real adoption and little hope

44:38

for a turnaround. Enjoy tomorrow's monologue, and

44:41

then the final part on Friday. Thank

44:50

you for listening to Better Offline.

44:52

The editor and composer of the

44:54

Better Offline theme song is Matosowski.

44:57

You can check out more of

44:59

his music and audio projects at

45:01

matosowski.com. m-a-t-t-o-s-o-s-w-s-k-i.com. You can email me

45:03

at easy at betteroffline.com or visit

45:05

betteroffline.com to find more podcast links

45:08

and of course my newsletter. I

45:10

also really recommend you go to

45:12

chat.where's your ed.at to visit the

45:14

discord and go to r slash

45:17

betteroffline to check out our credit.

45:19

Thank you so much for listening.

45:21

Better offline is a production of

45:23

Coolzone media. For more from Coolzone

45:25

media, visit our website Coolzone media.com

45:28

or check us out on the

45:30

iHart radio app, Apple podcast or

45:32

wherever you get your get your

45:34

podcast. In

45:56

a world of economic uncertainty

45:58

and workplace transformation, to lead

46:00

by example from visionary C-sweet

46:02

executives like Shannon Skyler of

46:04

PWC and Will Pearson of

46:06

I-Hart Media. The good teacher

46:08

explains the great teacher inspires.

46:10

Don't always leave your team

46:12

to do the work. That's

46:14

been the most important part

46:16

of how to lead by

46:18

example. Listen to Leading by

46:21

example. Executives making an impact

46:23

on the I-Hart Radio app.

46:25

Apple podcasts or wherever you

46:27

get your podcasts. I

46:29

am Bob Pittman, Chairman and CEO of

46:31

I Heart Media. I'm excited to introduce

46:33

a brand new season of my podcast,

46:36

Math and Magic, Stories from the Frontiers

46:38

of Marketing. I'm having conversations with some

46:40

folks across a wide range of industries

46:43

to hear how they reach the top

46:45

of their fields and the lessons they

46:47

learned along the way that everyone can

46:49

use. I'll be joined by innovative leaders

46:52

like Chairman and CEO of Elf Duty

46:54

Tarang Amin. Legendary singer-songwriter and philanthropist, Jewell.

46:56

Being a rock star is very fun,

46:58

but helping people is way more fun.

47:01

And Damien Maldonado, CEO of American financing.

47:03

I figured out the formalize has to

47:05

work hard, then that's magic. Join me

47:07

as we uncover innovations and data and

47:10

data and analytics analytics and the ever-important

47:12

creative spark, the magic. Listen to Math

47:14

and Magic on the iHeart Radio app,

47:16

Apple Podcast, or wherever you get your

47:19

podcast. I'm ready to fight. Oh, this

47:21

is fighting words. Okay, I'll put the

47:23

hammer back. Hi, I'm George M. Johnson,

47:25

a best-selling author with the second most

47:28

banned book in America. Now more than

47:30

ever, we need to use our voices

47:32

to fight back. Part of the power

47:34

of black queer creativity is the fact

47:37

that we got us, you know? We

47:39

are the greatest culture makers in world

47:41

history. Listen to Fighting Words on the

47:44

I-Hart Radio app, Apple podcast, or wherever

47:46

you get your podcast. Hey y'all, it's

47:48

your girl Cheekies and I'm back with

47:50

a brand new season of your favorite

47:53

podcast Cheekies and Chew. I'll be sharing

47:55

even more personal stories with you guys.

47:57

and as always, you'll get my exclusive

47:59

take on topics like love, personal growth,

48:02

health, family ties, and more. And don't

48:04

forget, I'll also be dishing out my

48:06

best advice to you on episodes of

48:08

Dear Cheekies. It's going to be an

48:11

exciting year and I hope that you

48:13

can join me. Listen to Cheekies and

48:15

Chill, season 4 on the I Heart

48:17

Radio app, Apple Podcast, or wherever you

48:20

get your podcasts.