HomeVideos

The AI Economy is about to change

Now Playing

The AI Economy is about to change

Transcript

276 segments

0:00

All right, fellas. It's happening. The

0:03

cracks are starting to show. The

0:05

foundations looking a little bit shaky.

0:07

I'm talking talking about the AI

0:09

economy, the token economy. Uh the first

0:12

one comes courtesy of course of

0:13

Anthropic, everybody's favorite, the

0:15

good guys, am I right? I love their

0:17

lawyers. What ended up happening just a

0:20

couple days ago is Anthropic did

0:21

something called a painted door test. At

0:23

least that's the term I've always heard.

0:25

I think I've seen other people use fake

0:27

door test where you actually show

0:28

alternative pricing on your page to see,

0:31

hey, how much more money could we make

0:34

if we charged people more? How many more

0:36

people would leave the pricing page? So,

0:38

let me give you a a hypothetical. Let's

0:40

say that you raised the cost of your

0:41

product by 10%. And a,000 people visit

0:44

your pricing page. Typically, you'd get

0:47

a 100 customers out of your thousand

0:48

customers. But now, with your raised

0:50

prices, you only get 95 customers. But

0:53

of those 95 customers, you make more

0:55

money. So you say, "Hey, the area under

0:57

the curve is worth it. We're going to

0:58

lose a couple customers, but we're going

1:00

to gain more money." This is kind of the

1:02

idea of a painted door test. But

1:04

Anthropic did something a bit different.

1:06

Instead of doing a price change, they

1:09

just simply removed clawed code usage

1:11

from the $20 plan. See, in a typical

1:14

painted door test, after someone says,

1:15

"Okay, we'll pay this price." They're

1:17

like, "JK, bro. You're super special to

1:20

us, so you get actually a reduced price

1:21

because we love you." Well, this one,

1:24

there's an entire group of people that

1:25

when they visited this page saw they

1:27

couldn't use Claude code for $20 a

1:28

month. They paid $100 a month. It just

1:31

tricked some small percentage of people.

1:34

I don't think anyone's surprised by

1:35

this. And the people that are surprised

1:36

by it, you always hear kind of the same

1:38

thing. Well, they're making money off of

1:40

inference. like every time they do like

1:42

a request call, of course they make

1:44

money. It's like, yeah, I guess if you

1:46

measured how much a shoe cost by only

1:49

paying the employee to sell the shoe,

1:51

then you're like, "Yeah, you're making

1:52

money off of every shoe. They're making

1:53

money off of every request." Yeah, that

1:56

makes sense, but you're not actually

1:57

considering the real cost. Remember,

2:00

people that used Opus 45 are now using

2:02

Opus 46. Nobody's using Opus 47. A kind

2:05

of tarted model. No, no one really likes

2:08

that one. But that means all the cost

2:10

that went into 45 if they didn't recoup

2:13

that in inference cost 45 just cost them

2:17

money. And this is obviously what's

2:18

happening. Open AI got 122 or $120

2:23

billion investment and that's enough

2:26

money to run for 18 to 24 months. That

2:29

is like 5 to7 billion every single month

2:33

in the hole for the next 18 to 24

2:36

months. They had to do this test because

2:38

they have to know how much money they

2:39

can make because if they don't make some

2:41

sort of change, they're going to

2:42

continue to lose billions of dollars.

2:44

Now, if this just happened by itself,

2:46

I'd say, "Hey, Claw just needs to make

2:48

more money. They need to start becoming

2:50

more competitive because they're

2:51

competing against OpenAI." And then I'd

2:53

be like, "Okay, that's that's just

2:54

that." But this isn't the only case of

2:56

this happening because just a couple

2:58

days ago, guess who else decided to make

2:59

a bit of a price change? Microsoft. Oh,

3:02

beautiful Microsoft. I love Satia. I

3:05

love co-pilot. Isn't co-pilot is this is

3:08

this the greatest? You're probably

3:09

confused when I say co-pilot because

3:10

you're probably thinking of like one of

3:11

their 50 services. I I'm talking about

3:14

the GitHub the GitHub co-pilot in this

3:16

case. See the GitHub copilot you used to

3:18

pay some sort of amount of money and

3:19

then you could perform actions on

3:20

behalf. You had some sort of amount of

3:22

actions you could execute. Well, what's

3:24

the problem with that? Not every model

3:26

cost the same. If you're thinking about

3:27

composer too fast from cursor, that

3:29

thing costs like nothing. When you're

3:32

thinking about Opus 47, that costs a lot

3:36

more. We're talking like 20 times more

3:38

the cost. So obviously Co-Pilot, they

3:41

had to make some sort of reduction. And

3:42

that's what they did. You no longer just

3:44

get some amount of executes. You get

3:47

token usage because if you use a more

3:49

expensive model, well guess what? You

3:51

got to use more of your tokens.

3:53

Therefore, you won't get as many calls

3:55

out of it. This is a change. It needs to

3:57

become more economically viable. But

3:59

before that, to pay for all my tokens,

4:01

sponsor time.

4:02

>> Look at all these engineers sitting at

4:05

their neat little desks. It takes dirty

4:07

work to keep a code base clean. Every

4:10

day, sickos are out there committing

4:12

unreed code. And when that happens,

4:15

llinters won't save you. You need

4:18

someone like me.

4:19

>> Let's go.

4:21

>> Feature free scrumbag. Who you calling

4:23

scrumbag? What's this slop you're trying

4:25

to push? Unnecessary comments. Global

4:28

state nested turnaries.

4:30

>> H my bad. I didn't even read the code

4:32

yet.

4:33

>> You disgust me. Step away from the

4:34

keyboard.

4:35

>> Just let me explain.

4:36

>> Is that a mouse? HE'S MERGING A PROD.

4:38

YOU HAVE THE RIGHT to remain silent.

4:40

Anything you push to GitHub Canon will

4:42

be used against you. You have the right

4:43

to a debugger. But if you cannot afford

4:45

one, a public stack trace will be made

4:47

available to you.

4:50

And one more code criminal off the

4:52

streets where they belong. HR.

4:57

Look, I didn't even I know I didn't

5:00

review any of the code, but I was going

5:02

to have Code Rabbit review it from the

5:03

start with oneclick fixes install

5:05

enforcement. I don't need Merge Cop. I

5:08

would never merge unreviewed code, but a

5:10

first pass with Code Rabbit always makes

5:12

things go faster. Actually, you can try

5:14

it too at code rabbit.ai.

5:17

>> Next week on Merge Cop.

5:19

>> The Diffler's out there, and I'm going

5:21

to be the one to deprecate them.

5:26

See, the difference between Microsoft

5:28

and Anthropic is Microsoft makes money.

5:30

I know, crazy concept for Anthropic, but

5:33

Microsoft makes a lot of money. Whether

5:35

you like it or not, they're one of the

5:36

biggest companies in the world. And

5:38

therefore,

5:40

they can actually kind of take like a

5:42

bit of a nose dive for a while

5:44

accumulating users. But even at this

5:46

point, they're going, "Hey, we can't we

5:49

can't do this. We got to make more. This

5:51

this is silly. What's going on here?

5:53

This doesn't even make sense according

5:55

to Microsoft. Now the real winner

5:57

honestly from all of this is Google.

5:59

Classic Google. They are pouring like a

6:02

hundred plus billion dollars a year into

6:04

AI and they can just do that. And guess

6:06

what? After they pour hundred billion,

6:08

$200 billion into AI, they still make

6:12

money. That's wild. Like they can do so

6:15

much money for the next year after year

6:18

and they don't have to worry about will

6:20

investors still find me attractive?

6:22

That's probably why you're not getting

6:23

the same level of hype coming from

6:26

Google that you get from these

6:27

companies. Like, just think about that

6:29

for a second. Google's also competing on

6:31

the frontier. Google's also attempting

6:33

to win the market. Google's also trying

6:35

to convince everybody that their AI is

6:36

the best AI, that they're going to be

6:38

able to shepherd it and take it into the

6:39

future despite the fact of inventing the

6:41

T and GPT and somehow fumbling the bag

6:44

and not being the first one to market.

6:46

Nobody knows the answer to that one. And

6:48

AI is really expensive. Uber just got

6:50

done claiming that within four months

6:52

they spent their entire year's budget on

6:55

AI. Gee, I can't believe this is

6:57

happening. How could you tell every

6:59

employee to maximally use AI? By the

7:02

way, we're judging you on AI usage. Oh

7:04

my gosh, you're using too much AI. How'd

7:06

that happen? What the hell? You're not

7:08

supposed to be using a year budget in 4

7:10

months. What are you even doing? What

7:12

are people doing with all those tokens?

7:14

I don't even know. Like, I don't think

7:16

we're going to be like, "Well, back it

7:17

up everybody. Uh, we're not going to use

7:19

AI anymore. You're going to have to go

7:20

back to hand coding everything because

7:22

AI is just not economically viable. No,

7:24

no, no, no, no, no, no, no, no, no, no.

7:26

They will find a way. Sure, it may be in

7:28

a couple years, but you know what?

7:30

They're going to find a way to make this

7:32

thing viable. But for now, we're

7:34

starting to see the cracks. Things just

7:36

can't be as free as they once were, and

7:38

the amount of usage you're going to be

7:39

getting is clearly and obviously going

7:41

down. I don't want to be the person

7:42

that's like, "Oh, Mr. Anti-AI for all

7:44

reasons." Like that's why I make this

7:46

video to show you how ridiculous their

7:48

marketing is. This is why they do such

7:50

hyped out marketing. This is why Daario

7:53

is constantly telling you, "Hey, you're

7:54

out of a job here very, very soon. You

7:56

you you Yeah, we're going to take your

7:57

job." I feel super bad about it. Oh, I

8:00

just feel so bad about me taking all of

8:03

your money. I'm so sorry. It's so

8:06

dangerous. But this is the reason why

8:08

they're doing it. They need to raise

8:10

money. This is why Google doesn't do it.

8:12

They don't need to raise the same kind

8:14

of capital that uh Daario and Sam need

8:16

to raise. I think there's plenty of

8:18

great uses for AI. I I do use it on the

8:20

regular. Even though I participate in

8:23

Daario's least favorite activity, hand

8:25

coding every single I'm a tried coder.

8:26

Okay, I like triad coding, but I also

8:28

like AI. There's plenty of times that

8:30

it's actually super convenient. And so

8:32

ultimately, I just hope to kind of bring

8:33

a a bit of a more middle ground to

8:35

things because tech is exciting. Like

8:37

think about this for a second. We get to

8:40

build like anything that your mind can

8:42

come up with. Like that is such an

8:44

incredible privilege to have. Most of

8:46

the world was spent dying because you

8:49

stepped on some piece of metal that you

8:50

didn't see. And oh, sucks. You got

8:52

tetanus now. Looks like you're dead.

8:54

Dysentery on the Oregon Trail. Yep. Too

8:58

bad. Shouldn't have been doing whatever

9:00

you were doing, which was called living

9:01

cuz now you're dead. Instead, we

9:03

actually get the opportunity to build

9:05

all this stuff. And for me, this is the

9:06

most exciting time period to ever live

9:08

in because now I get to just have any

9:11

kind of personalized experience. Like a

9:13

quick pro tip, you're learning

9:15

something, dude. If it's open source,

9:18

clone down the repo, open up AI and say,

9:20

"Hey yo, based on the code, give me an

9:22

example of how to do this. Explain that.

9:24

Explain this." It's like personalized

9:26

documentation. Super cool. Anyways, I

9:28

wanted to end on a high note, you know,

9:30

uh, just because I just feel like the

9:32

yapping that's going on is just so kind

9:35

of jaded. The name is the tokenogen.

Interactive Summary

The AI economy is currently experiencing significant financial challenges, with major players like Anthropic and Microsoft adjusting their pricing models. Anthropic conducted a "painted door test," effectively charging some users more to gauge revenue potential, while Microsoft shifted GitHub Co-Pilot to a token-based usage model to account for the varying costs of different AI models. OpenAI, despite a massive investment, is reportedly losing billions monthly due to the high costs of developing and running AI. In contrast, Google is positioned as a strong contender, capable of investing hundreds of billions annually into AI while remaining profitable, thus not requiring the same aggressive marketing for capital. The escalating expense of AI, exemplified by Uber exhausting its annual AI budget in just four months, indicates that current "free" or low-cost usage models are unsustainable, and a reduction in usage and price adjustments are anticipated. Despite these economic hurdles, the speaker believes AI has immense potential and that solutions for its long-term economic viability will eventually be found.

Suggested questions

5 ready-made prompts