HomeVideos

The Secret Token Underworld

Now Playing

The Secret Token Underworld

Transcript

407 segments

0:00

Chinese resellers are offering Claude

0:02

tokens at 70 to 90% below official

0:06

Anthropic prices. Now, I don't even know

0:09

what to do with that statement. What are

0:11

we even talking about? There's a black

0:13

market for tokens? Yes. Yes, there is.

0:17

There is an entire economy built around

0:19

this this black market, this gray

0:21

market, this shadow API economy. So,

0:25

we're actually going to go over who's

0:26

using it, why they're using it, how this

0:28

actually even works. How do they have it

0:30

at 70 to 90% off because I mean that

0:33

sounds kind of nice. Did they solve They

0:35

solved something we don't understand.

0:36

And also, I'm going to toss my little

0:38

hat into the ring and I'm going to give

0:40

you a reason why I think things are as

0:41

cheap as they are. I got my own little

0:43

personal theory that I'll unleash at the

0:45

end. All right, so to understand this

0:48

entire gray market universe, you have to

0:51

understand the fundamental unit.

0:53

Everything is around this idea of a

0:55

transit station. What this is is it goes

0:59

from the great country of China and it

1:02

takes in requests going to these transit

1:04

stations and then out they go to say

1:07

Anthropic or OpenAI and then pipe the

1:10

results back in back to the people

1:12

behind these intelligence iron curtains.

1:14

These countries that are not allowed to

1:16

access OpenAI or Anthropic because of

1:20

government restrictions. Now, you're

1:21

probably thinking this this sounds like

1:22

a VPN. It is. This is exactly how a VPN

1:25

effectively works. You from your

1:27

computer make requests, it goes up to a

1:28

VPN, the VPN then goes off and makes the

1:30

request and funnel everything back to

1:32

you. This is exactly the same thing

1:34

except for they do it through these

1:35

transit stations. Now, you're probably

1:36

saying, "Well, what's the difference

1:37

between a transit station and say open

1:39

router?"

1:40

Nothing. Just just just legality, that's

1:42

all. Well, and and also

1:45

fraud and and many other things, but but

1:47

effectively nothing. So, it's just good

1:49

to keep in mind when I talk about

1:50

accessing the data and everything, it's

1:52

all done through these transit stations

1:54

cuz I think it makes a lot of the rest

1:55

of the story make sense. So, this is

1:57

kind of the fundamental unit of this

1:59

entire black market economy. But before

2:02

we continue, hey, I'd like to say thank

2:04

you to the sponsors. Hey, is that HTTP?

2:07

Get that out of here. That's not how we

2:09

order coffee. We order coffee via SSH

2:11

terminal.shop. Yeah, you want a real

2:13

experience? You want real coffee? You

2:15

want awesome subscription so you never

2:17

have to remember again? Oh, you want

2:19

exclusive blends with exclusive coffee

2:22

and exclusive content? Then check out

2:25

Cron. You don't know what SSH is?

2:27

Well, maybe the coffee's not for you.

2:30

>> Terminal [singing]

2:31

coffee

2:33

in hand

2:34

living the

2:36

dream.

2:38

>> When you hear that the Chinese are using

2:40

these transit stations, your first

2:42

thought is Dario was right. They are

2:44

doing distillation attacks. Well, no,

2:47

no, no, actually, calm down. That's not

2:49

actually what's happening. The people

2:51

who are using it is people who want to

2:53

use the models to do something. Now, I'm

2:55

not going to say what they're doing is

2:56

actually good, but they're doing things

2:58

with the models, whether it's just

3:00

programming, building applications, or

3:02

scamming people. But the point is it's

3:04

not a distillation attack in itself.

3:07

Instead, it's actually representing real

3:09

work. People actually using these

3:12

transit stations to do something. But

3:14

this is where things kind of get

3:15

interesting. Like, how would a transit

3:17

station actually work? From this amazing

3:19

article, how to buy cheap Claude tokens

3:21

in China, it actually kind of goes

3:23

through what's going on at these transit

3:25

stations. Upstream are the resource

3:27

providers, account merchants who bulk

3:29

register or acquire Anthropic accounts

3:31

at scale. Now, you're thinking, how

3:33

would you bulk register thousands, if

3:36

not tens of thousands of accounts with

3:37

Anthropic? Well, luckily, there's this

3:39

other article that actually explains how

3:41

you could do exactly that. See, there's

3:43

three problems. Each account needs a

3:45

unique browser fingerprint, a unique

3:47

residential IP, and a way through phone

3:49

and ID verification. Browsers have moved

3:52

past Playwright plus stealth plugins to

3:54

anti-detect browsers. Now, I did not

3:57

know this that there was these

3:58

anti-detection browsers. And there's

3:59

actually quite a few of them right here

4:01

listed right here. In fact, there's a

4:02

bunch of headless ones that are actually

4:04

identical to operating as Chrome would

4:06

so that the receiver of these messages

4:08

can't actually tell if you are or are

4:11

not from Chrome. Which means that

4:12

defenders that have been fighting this

4:14

network layer for a decade, they've

4:16

lost. You can no longer tell from the

4:18

wire if you're actually getting

4:20

legitimate traffic or not. At least

4:22

that's what's kind of being claimed with

4:23

these anti-detect browsers. But what

4:25

about SMS? That has to be hard, right?

4:27

Well, no, actually that's the cheapest

4:29

defeat in the pipeline is phone

4:30

verification, it turns out. It turns out

4:32

you can get a whole bunch of SIMs for

4:35

pretty dang cheap. And you can also

4:36

acquire SMS numbers pretty dang cheap.

4:40

See, the thing I guess a lot of us don't

4:41

think about is that at one point in

4:43

life, SMS was a pretty good proxy for

4:45

personhood. One phone number belongs to

4:47

one person. That's just no longer the

4:49

case anymore. It's really cheap to get

4:51

phone numbers these days. It costs about

4:53

.8 of an Abraham Lincoln, a one penny, a

4:55

cent. I think them British call them a

4:57

pence. I'm not really sure. I don't I

4:59

don't I don't speak British, okay? I

5:01

don't speak British. In other words,

5:03

these transit stations, they rely on

5:05

these SS SMS services to kind of do

5:07

phone verification. But then there's a

5:09

second problem. What about ID

5:11

verification? That has to be harder,

5:13

right? Well, it turns out jumping into

5:14

this 404 article by Joseph Cox,

5:17

no, actually it's not that hard. In

5:20

fact, a Telegram organized site operator

5:23

pseudonym John Wick selling AI-generated

5:25

driver's licenses, passports, and

5:27

national IDs for $15 each, auto

5:30

composited onto plausible carpet

5:32

bedsheet backgrounds, claimed the

5:33

throughput up to 20,000 documents per

5:36

day. Which by the way, even at half of

5:37

that, if they did that for a year, is

5:39

$50 million.

5:41

Now, these passports, they were actually

5:42

really sophisticated because the

5:44

technical move that makes them pass is

5:46

the MRZ, the the dashes, the little

5:48

carrots, lines at the bottom of the

5:50

passports encoding personal data plus

5:53

three check digits and a global checksum

5:55

per ICAO 9303.

5:58

Apparently earlier they were getting

5:59

caught by these validations, but they

6:01

got so sophisticated, these backends

6:03

that would validate if this was a real

6:05

license or not or a real passport were

6:07

like, "Yeah, bro, these are actually

6:09

really good. You can get by a whole set

6:11

of ID verification just via AI." There

6:13

was a whole like meta getting all these

6:15

accounts stolen because all you had to

6:16

do was go to the meta bot and be like,

6:18

"Yo, bro, I'm totally this person."

6:19

Generated deep fake of you holding an ID

6:22

of the person that doesn't actually

6:23

exist and then the meta bot was like,

6:24

"Sure, that's totally cool. You're now

6:25

this person. Would you like to reset the

6:27

password?" And a whole bunch of people

6:28

got their accounts stolen. But,

6:30

sometimes you need real IDs. And if you

6:32

need real IDs, this is where the story

6:34

gets a bit uneasy how these transit

6:36

stations operate. Anthropic has started

6:39

implementing identity verification. You

6:41

need a passport and a selfie to use

6:43

certain features. This is where the

6:44

advantages of AI transportations from

6:46

the crypto circle comes in. By the way,

6:48

crypto ruining everything, classic

6:50

crypto. We're back to the stage of being

6:52

our most familiar KYC manufacturers

6:55

braving the risks of Ebola in Africa to

6:57

find locals to help with face scan

6:59

verifications and selling IDs. Buying a

7:01

smiling teeth photo from a local for $5

7:04

and selling it for $100 to a ByteDance

7:06

programmer at Zaichen Road in Beijing.

7:08

Other circles can't handle this kind of

7:10

hardship, but we KYC manufacturers in

7:13

the crypto circle can.

7:15

>> [laughter]

7:16

>> Yeah, bro, dude, they're so good, man.

7:18

Those crypto bros are so

7:19

They're just They're just braving the

7:21

conditions. Apparently there's even a

7:23

Worldcoin black market. I did not

7:24

realize that where they're actually

7:26

duping locals to scan their eyeballs and

7:28

then selling those black market IDs so

7:30

that you can be like, "Oh yeah, I'm

7:31

totally this person. I have my

7:33

Worldcoin. Yeah. I totally I I totally

7:37

am from Zimbabwe." So, one of the last

7:38

little spokes here is actual real

7:40

people. We're talking about real people

7:42

selling their IDs. Now, the crazy part

7:44

about this entire gigantic circle right

7:47

here is that there's tons of transit

7:49

stations. There's tons of SMS carriers.

7:52

There's tons of these AI-generated and

7:54

real people identification services. So,

7:56

to actually stop this stop this blob is

7:59

effectively impossible. These black

8:02

markets, these gray markets, they just

8:04

exist and you can take down one SMS,

8:06

it's not going to affect the the transit

8:09

station. It's not going to affect the

8:11

IDs. They all just keep on going this

8:13

this amorphous replaceable modular

8:16

system. I mean, it's a very

8:18

amazing feat of engineering as bad as it

8:21

is. Now, that's how they operate. That's

8:23

how these things work, but the real

8:26

question comes down to how are they so

8:28

cheap, right? Like, there has to be a

8:30

story behind it. Well, there's multiple

8:33

layers of fraud and let me explain. One

8:35

thing these transit stations do is they

8:36

actually use the Claude Max plan and

8:40

able to kind of create a pool of Claude

8:42

agents so that you can kind of route

8:44

traffic to to use each one of their

8:46

weekly limits. So, that way they get

8:48

cheaper tokens if needed. Now, that

8:51

obviously does not cut the cost to where

8:53

they really need it, but it gets really

8:54

close. It could possibly break them down

8:56

to even. But, that's not really the type

8:58

of fraud you were probably thinking

8:59

because that's not really where they do

9:01

a lot of the big stuff. But, the more

9:03

interesting way in which they save money

9:05

and allow for these cheap tokens is this

9:07

right here. Real money, fake models,

9:09

deceptive model claims, and shadow APIs.

9:12

If we go down here to page two, we see

9:14

the following. Alarmingly, our

9:16

experiments reveal significant

9:17

performance inconsistencies between

9:19

shadow APIs and the official APIs. On

9:22

high-risk medical benchmarks like

9:24

Med-QA, the accuracy of Gemini 2.5 flash

9:28

model drops precipitously from 83.82%

9:32

with the official API to approximately

9:34

37%

9:36

across all examined shadow APIs. In

9:38

other words,

9:40

they're saying, they're claiming,

9:41

they're showing you right here, "Hey,

9:43

you're going to get Claude Sonnet 4 6."

9:45

Now, you might actually just be getting

9:46

haiku. Maybe you're just getting Gemini.

9:48

Like, you don't actually know what

9:50

you're getting underneath the hood, and

9:52

this is really how they start cutting

9:54

those costs down. The other obvious one

9:56

is stolen credit cards and all that that

9:58

are used to pay for a lot of these

10:00

Claude Max plans that they're able to go

10:02

and run for some small amount of time

10:03

before they get caught. So, those are

10:05

effectively free money. But, the one

10:07

that I hate to say, the one that I can

10:09

already hear right now, Dario going, "I

10:12

knew it!" is that they take the data.

10:15

Cuz here's the thing, is these transit

10:17

stations, real people are trying to

10:20

solve real problems. Whether the real

10:22

problems are good or not is not up to

10:23

debate. It's that they're solving real

10:25

problems, and that means these question

10:28

and answer pairs that are coming back

10:30

out, these are real problems and real

10:33

results from high-end models. American

10:35

models, Frontier, Soda, all what is Soda

10:37

Deez Nuts? I'm not even really sure what

10:39

all that stuff means. But, nonetheless,

10:40

that's what's happening. So, they can

10:41

take these question answer pairs and

10:44

sell them to these Chinese labs to be

10:46

able to develop these open-source kind

10:48

of 6 months behind models. And this is

10:50

where the real money's at. This is where

10:52

the real gold is at. Is that these

10:54

transit stations actually get the best

10:56

thing of all, which is real data from

10:59

American companies. The good stuff.

11:01

Okay, this is pure, unadulterated Opus

11:04

4. They even probably got a little bit

11:06

of Fable. How much do you want to bet?

11:08

They probably even got a little taste of

11:09

that Fable.

11:11

But, now it's time for my theory. My

11:14

theory for how they make even more money

11:16

with all this. Now, you've probably seen

11:18

these tweets throughout the time talking

11:21

about how all these users of Twitter,

11:23

there's literally like thousands of

11:24

these tweets where they show like, "Oh

11:26

my gosh, oh you know, after like one

11:28

prompt, my weekly limit's completely

11:30

gone." And you're sitting here thinking

11:32

like, how how is Claude that bad? Like,

11:34

why would you pay money for that? Like,

11:36

all these people on the internet are

11:37

just so angry about usage limits. Like,

11:40

I wonder if there's there's something

11:41

wrong. Well, then you think about, well,

11:43

wait a second. What about Shai Halud

11:46

that happened, you know, about a year

11:48

ago? Shai Halud was a worm in NPM that

11:51

would self-propagate and steal all these

11:53

tokens.

11:54

And then you had Shai Halud 2, you had

11:56

Shai Halud 3. Really, what you're seeing

11:58

is that there's all these vibe victims,

12:00

as I like to call them, that have no

12:02

idea how computers

12:04

work and have been told over and over

12:06

again, "Hey, bro, you're a software

12:07

engineer now, okay? Your programming

12:09

language? English. Go get 'em, tiger,

12:12

okay? Yeah, you probably need like five

12:14

Claude Code Max subscriptions. Yeah, no,

12:16

no, you totally got this. It's totally

12:18

normal to run into your weekly limits

12:19

all the time. You got this, bro. Go get

12:21

'em. Go get 'em." And all of this just

12:24

like massive worm nonstop getting your

12:27

system hacked by these completely insane

12:29

dependency graphs that exist out there

12:31

that if you want to even just say hello

12:33

world on a website, you need to go and

12:35

download tens of megabytes of

12:37

JavaScript. Crazy, right? Like, I don't

12:40

know what times we live in, but that's

12:42

what's happening. Also, you got like

12:43

Arch Arch is constantly getting hacked

12:45

these days. PyPI? Oh, PyPI is hacked

12:47

these days. Like, everywhere you go,

12:49

every single one of these

12:51

ecosystems of package managers are just

12:53

fraught with worms and deception. And

12:55

so, how much do you want to bet there's

12:57

an entire class of people that are

13:00

effectively have their computers acting

13:02

like little mini VPNs for these transit

13:05

stations. Oh, you have Claude Code up?

13:08

We're going to spawn another little

13:09

Claude Code in the background and do a

13:10

little piece of work. Okay, hey, thank

13:12

you, not a big deal. Have fun with your

13:14

weekly limits, not a big deal. Right?

13:16

Like, they don't they don't want to

13:17

abuse it. They don't want to get it

13:19

taken from them. All they need to do is

13:21

just do it sometimes, just a little bit.

13:23

Hey, bro, just a little bit, you know,

13:24

like when the person's already active.

13:26

So, in case their session runs out, it

13:28

feels kind of crazy. I am completely

13:30

convinced that this is happening right

13:32

now, that there's millions of requests a

13:34

day that are through people's laptops

13:36

that they don't even know about because

13:38

they're software engineers now, okay?

13:41

And NPM's totally reasonable ecosystem.

13:44

So, there you go. That is the clawed

13:46

black market token 70 to 90% off. It's

13:50

absolutely wild out there. I'm going to

13:52

include a bunch of sources inside the

13:54

description. Go check them out.

13:56

It is an in-depth enterprise that's

13:58

going on. The name is the primogen.

Interactive Summary

This video examines the existence of a 'shadow API economy' where Chinese resellers provide access to premium AI models like Claude at significant discounts. The host explains the infrastructure behind these 'transit stations,' which bypass government restrictions, and details how they overcome hurdles like account registration, phone verification, and identity checks. Furthermore, the video uncovers how these operators cut costs through fraudulent means—such as swapping powerful models for cheaper, inferior alternatives—and highlights the lucrative practice of harvesting high-quality training data from unwitting users to improve competing models. Finally, the host presents a theory that everyday users, through compromised software dependencies, might inadvertently be powering this network with their own computing resources.

Suggested questions

4 ready-made prompts