Luyện nói tiếng Anh bằng Shadowing qua video: I bought Black Market AI Tokens on Chinese Amazon

Chọn chế độ luyện tập

Mỗi chế độ mở một phòng luyện tập toàn màn hình.

311 câu
Nếu các câu quá ngắn hoặc quá dài, hãy bấm Edit để chỉnh sửa.
1
Did you know there's a black market for AI tokens?
Bạn có biết là có một thị trường chợ đen dành cho các mã thông báo AI không?
2
Chinese sellers are offering up API tokens
Các nhà bán hàng Trung Quốc đang cung cấp các mã thông báo API.
3
and subscription to popular LLM providers like Cloud Code and OpenAI Codex for pennies on the dollar.
và đăng ký các nhà cung cấp mô hình ngôn ngữ lớn phổ biến như Cloud Code và OpenAI Codex với chi phí chỉ bằng một phần nhỏ so với giá gốc.
4
I've heard stories of companies losing a million dollars a day.
Tôi đã nghe những câu chuyện về các công ty bị thua lỗ một triệu đô la mỗi ngày.
5
I heard through the grapevine of a coding agent, one of the more popular ones, that's currently losing $10 million a month and it's doubling every two months.
Tôi nghe tin đồn về một tác nhân lập trình, một trong những cái phổ biến hơn, hiện đang thua lỗ 10 triệu đô la mỗi tháng và con số đó tăng gấp đôi cứ sau hai tháng.
6
Three of my favorite things in life are good deals, sketchy Chinese websites, and protecting shareholder value, so you know I had to dig in and check it out myself.
7
I spent my own money on black market Claude tokens, and what I found out was a well-run multi-million dollar industry that goes deeper than I could have ever imagined.
8
Were these real Claude API tokens?
9
Where are they getting them?
10
How are they offering them so cheap?
11
What's the catch?
12
Are they selling my data?
13
I'm Wes, and welcome to Syntax.
14
searching for black market tokens, I hopped on Taobao and searched for Claude, and I was amazed at how many things will pop up.
15
Now, Taobao is kind of like a Chinese Amazon, but there are tons and tons of different sellers on there.
16
If you've ever used like AliExpress or something like that, it's like that, but a thousand times better with absolutely everything that's available to you, even black market Claude tokens.
17
And after a couple hours of searching, I came across kind of three types of listings after doing all of the translate.
18
The three types of listings that I stumbled upon were first, seemingly legitimate Claude subscriptions.
19
So these ones right here, if you look at them, Claude 4.8 Pro, Whitelist, Old Account, Membership, Top Up, Old Accounts.
20
What these are, are people that are using legitimate
21
or US credit cards and selling them to people who want legitimate full Claude access.
22
So someone who wants Claude access but is not able to
23
because they don't have the identification that is needed or the proper credit card that is able to buy it.
24
And if you look at it, you convert the actual prices of a lot of these, you'll see that they're actually more than what Claude code subscriptions would cost.
25
So I don't think there's much wrong with that other than people who are unable to access it, wanting to access it.
26
Now, the second type that I stumbled upon was something like this.
27
Claude 4.7 Pro Subscription Shared Monthly Card Account Claude Pro Max Membership domestic version.
28
If you look through all of these, they sell like seven days.
29
Now, Claude doesn't sell a seven day access to one of these, right?
30
So seemingly this one seems to be access to, I think the Claude Code website on a shared account.
31
So I went ahead and bought that one, seven days, eight and a half yuan, which is about a buck 50, right?
32
That's dirty cheap.
33
And then the third one, which is really what we're looking for, is we're looking for straight clawed anthropic token access for really cheap, right?
34
So this is the one that I ended up landing on.
35
$100 per knife limit.
36
Use this up until end.
37
No time check.
38
So what this essentially translates to, if you take it out of this Google Translate and put it into some other translate,
39
this equates to $100 worth of credit used up.
40
There's no time limit, right?
41
55 yuan converted to USD is $8.
42
So basically they're offering you 8% of the cost. $100 worth of credit for $8.
43
So I bought that one as well and we're going to dig in to see if it actually works.
44
Immediately what they did is they sent me some random token
45
that I can use and then a link to UQ which seems to be similar to a Chinese notion.
46
So dusted off my Chinese and read all of the instructions and I sort of understood what we're working with here, right?
47
So when I logged into the Claude website one, this is what you get, right?
48
Here you are at the garage and you have zero heaven and you have to select a parking space, a total of zero vehicles.
49
So what I've understood this to be is these are different instances of the Claude website running.
50
So they seem to have a bunch of browsers running and they're proxying it through.
51
So watch, if I click through to one of these, it will load up what seems to be and seems to look like the actual Claude website.
52
So I can go and I can look at some of the previous conversations that I've had.
53
I can say, who does he do a podcast with?
54
This seems to be an actual Claude instance.
55
So I opened up the dev tools here.
56
I looked at the network requests every single time that you click on something.
57
This looks exactly like the Claude website.
58
It's not just like a dupe of it it the everything down to the JavaScript bundle names are exactly the same.
59
So what they are doing here is they are proxying it through.
60
So they have a whole bunch of Claude instances running on a browser somewhere, and they are running all of these different chats
61
and keeping track of the chat IDs that are belong to me
62
and their proxy in the middle is they're filtering it out.
63
So they're using the same one over for multiple people.
64
I things like you click on the design
65
and it just says not found
66
and it will will log me out it was a very very crappy experience for me
67
and also when I take a look at my dosage
68
which I simply have run three or four chats
69
so far I only get 50 over the next three hours
70
that doesn't seem like a very high thing so for the 10 bucks
71
that I spent on this probably not worth it what I'm more interested in is the actual $100 worth of Klud tokens.
72
Does that actually work?
73
Is it real?
74
Let's dive into that.
75
All right.
76
So my $100 worth of Klud credits for $8.
77
How does this work?
78
So I clicked off over to the Chinese notion and read the guide for newcomers.
79
And if you scroll it, you start seeing, oh, you can use it with the cloud code CLI, right?
80
You can use it with cloud desktop.
81
You can use it with VS code, use it with cursor.
82
And they basically detail how to use it with absolutely every single tool that you could possibly want.
83
So I thought, yeah, let's try it with Cloud Code first.
84
So I went over to the website that they give you access to code.anyxml
85
and I inputted that ID that they gave me and boom, $100 worth of credits, right?
86
Then I went over to the API keys and had to create an API key.
87
This is where it started to get a little bit fishy, right?
88
So you click on create token and you have access to three different times of token groupings
89
And what the hell is this?
90
5x ratio and 1.5x ratio?
91
So at first I didn't totally understand what this was.
92
And as I went and translated it, essentially they have three different groupings of Claude that they're trying to sell you, right?
93
There's Claude Max is of high quality, there's Claude Fast, and then there's Claude offers great value for money um
94
and what i've broken this down into is that there is claude max
95
which seems to be we'll go into this in a second
96
seems to be legitimate claude is not actually a hundred dollars
97
uh worth of credit for eight dollars it's it's a hundred dollars worth of credit for for forty dollars
98
because it's it's actually boosting it up at five times the ratio then claude fast and the other ones,
99
they have slightly different rules around caching and whether they actually work if they're offline or not.
100
Those ones are 1.5x.
101
So there is no 1x.
102
We're kind of screwed there.
103
But I went ahead and made tokens for all of them.
104
And I went in, of course, I'm not going to run this on my computer.
105
I'm running this on a sandbox environment where it can't actually get access to my entire computer
106
and i simply just went into my cloud settings
107
and swapped out the tokens
108
that are being used i don't know you can you can use the rest of my tokens
109
if you want after this and then the base url
110
that we're sending it to right there's they've set up a transfer station in between
111
that will inject it now if i just go ahead
112
and run claude in here of course i trust this folder
113
right there we go opus 5 let's just go ahead
114
and try say hey it's gonna boogie and we'll see if it actually works
115
all of it comes back in chinese
116
which is not what actually happened last time let's let's translate
117
that quickly and see what that actually means what content needs to be searched what code
118
or documentation needs to be created tell me exactly what you want to do and i will execute it immediately
119
please speak english for me create a simple website
120
that is an admin dashboard let's see if it will go ahead
121
and create those things for me there we go we got
122
it back to english all right it seemed to create it in 44 seconds
123
which is suspiciously fast it wasn't doing it that fast earlier
124
so that makes me think it maybe is not even using Claude at right now let's say start a
125
v server with allowing ip address visiting all right seems like it finished up
126
and visiting the website
127
that also is suspiciously does not look like a Claude Opus
128
5 does not use emojis like this it doesn't do this
129
design it is blue this is this is very clearly not giving you Opus 5.
130
However, I will tell you that in my testing in the past, it was giving me actual cloud.
131
Let's try change the model to Sonnet.
132
All right, here it's telling me we want modern dark mode.
133
It took five minutes to come up with the question of do you want modern dark mode?
134
So obviously not very fast.
135
While that's running, let's talk about this for just a second.
136
So what I did is I whipped up a script
137
that would run a bunch of different calls to the different APIs out there, right?
138
and I also use legitimate cloud tokens
139
because I wanted to take a look at the outputs of them
140
and say how similar to these are they to real cloud code and
141
that will be able to tell me things like this is
142
the real cloud code model versus they've been swapping something out
143
so when I did the legit cloud token it was 147 tokens
144
and then the max token also was 147 tokens meaning that their tokenization for each of the models,
145
this was with Claude Opus 4.8, was exactly the same.
146
And all of these models have different tokenization patterns.
147
So saying something like, hey, how are you doing today?
148
We'll tokenize differently for each of the models.
149
Then with the slow token, it was giving me 160 tokens.
150
Then I asked AI to write a script and try to do extract the system prompt from it.
151
And from the cheaper one, the one that was one and a half X, it was able to get out the prompt from it where it says,
152
if anybody asks you in Chinese or English who you are, you tell them that you are cloud code.
153
So almost certainly I can tell you that using the API keys that are at one and a half X ratio,
154
that's not actually cloud.
155
They're just substituting it with a much cheaper model.
156
Now, the one that is 5X ratio, I do believe from all of my testing that that is legitimate Claude.
157
Now, where did they get those?
158
We'll talk about that in just a second.
159
Eight minutes and three seconds later on Sonnet, it finally finished.
160
And this is what it gave us, which is much more something that Sonnet and Claude was going to put out, right?
161
It's
162
that common blue you got these backgrounds right here you've got
163
the hover effects on absolutely everything right you hover over top of it
164
because a little bit of an animation that those are all
165
clawed calling cards that you're gonna see there so I'm
166
that matched with the tokenization lining up one to one makes me confident to say
167
that this 5x1 actually is giving you clawed
168
most of the time one more thing to show you
169
that this is a bit of a scam
170
if we go to our usage for building this simple little website is telling us we use $1.92 worth of tokens, right?
171
And if we go to our dashboard here, you'll see that today we have consumed, and this is all I've done today,
172
is $3.45 worth of Sonnet and $3 worth of Opus.
173
So $6.40 worth of tokens.
174
And why that is happening is because they are stealing all kinds of different APIs.
175
And when they need to switch from one API to another, because I remember it was spinning for a while.
176
And in that case, it probably because one API got shut down, they have to take our entire chat transcript and then send it to another API endpoint.
177
And those tokens are not cached.
178
And if you're not sending cash tokens, you have to send the entire thing again.
179
And that gets really, really expensive.
180
So if I'm doing my math correctly here, I spent $8 for $100 worth of tokens, right?
181
Then I go to my API key to create my token and it is five times that.
182
So eight times five, that is 40 cents on the dollar.
183
Now, what they're telling me here is that my $2 spend was really closer to six.
184
So that's three times more expensive.
185
So my 40 cents ends up being 120 cents,
186
meaning that I just paid $1.20 or $120 for $100 worth of tokens.
187
That's not a good deal at all.
188
I think they're being a little bit sneaky here with it.
189
By the way, if you're enjoying this, hit subscribe and let me know what other seedy areas of the internet I should be looking into.
190
Got a couple ideas.
191
If you are a seedy person, send me a DM.
192
Some questions I had now were, where are these black market token resellers getting access to the tokens.
193
And luckily for me, while I was researching it, Matt Lenhardt posted this blog post from Vectoral
194
and turns out the guy started a company to stop this exact fraud because it's a much bigger deal than I thought.
195
Yes, I'm Matthew Lenhardt.
196
I'm the CEO and founder of Vectoral, a platform to stop token fraud and abuse
197
that really stems from problems I ran into at a previous
198
company i was working at where we were constantly battling token fraud
199
and i started to look into it like where is this all stemming from
200
and turns out there's an entire underground market for it matt
201
has dove deep into the forums where all these chinese sellers are trying to figure it on out
202
and sort of broken it down into four different groups
203
that we have
204
that right upstream you have the credit cards they're maybe using
205
stolen credit cards to be able to register with these big labs very open ai anthropic gemini etc um
206
and then midstream this is where they start to pool all of the accounts
207
and this is kind of the the interesting parts to me
208
so they are they're trying to figure out where they can get tokens from absolutely anywhere right
209
so whether that is there's an npm worm
210
that has scraped your dot env file whether that is uh reverse engineering some of the things like I'm sure VS Code,
211
Cursor, Kiro, all of these places that maybe they offer a trial, maybe they have bought like a $20 plan,
212
but you're able to use $500 on that single plan, because you know, some of these guys are losing money on it.
213
They're trying to reverse engineer because tokens are tokens.
214
And wherever they come from, they don't necessarily care.
215
So they're trying to reverse engineer all of the APIs, and then put it into a single tidy API, which is The next one, the downstream, these are the transfer stations.
216
So this is what I've been interacting with right here.
217
They give you a nice tidy API where you can log in.
218
You can see your usage for the day.
219
You can make an API key.
220
They give you a URL that you can ping and it is Anthropoc compatible.
221
So you can just swap it out in cloud code.
222
This is the transfer station where they take in my request.
223
They create me X, Y and Z.
224
And it's up to them to figure out where those need to be routed.
225
And in our case, when I was showing you, it took eight minutes to do a very simple task.
226
And that's likely because they're always getting shut down
227
and they need to reroute those requests to different APIs to figure out which one will work.
228
And I think that's also why I was running out of tokens much faster than it said.
229
Then there's me, the end user, looking for a good deal.
230
Yeah, so there's actually sites that have popped up now
231
that try to verify this at the different relays or transfer stations
232
because there's a lot of talk of models being swapped out.
233
And so now there's like, I guess, like verification sites that try to verify this for end users like, hey, are your stolen tokens,
234
you know, actually the model that they're purporting to be?
235
Now, if you're just watching this video saying, well, I don't know where to steal tokens from.
236
Don't worry.
237
There's an entire website dedicated to reviewing stolen token providers.
238
You can go on it and you can say, I think I need a little bit of fable today, right?
239
And you can see that LLM API is pretty good.
240
Or KUN.AI has pretty good.
241
And they're tracking absolutely everything in terms of like, how often are you switching APIs?
242
How what's the latency like in returning them?
243
What's the price?
244
They even have free API keys.
245
They have so much out here.
246
They just want to tempt you with a little bit.
247
It's absolutely nuts to me that this is such an industry
248
that we now have like Amazon reviews and free giveaways on stolen tokens.
249
The stolen API keys thing is really interesting to me because we hear about these NPM hacks all the time.
250
And then you don't really hear about like the craziness that happens.
251
When it was Bitcoin, you would hear immediately my account was drained.
252
With these tokens, I'm wondering, are they just doing it enough to hopefully not get noticed or to fly under the radar?
253
They're pretty clever.
254
Like they won't take you for everything you have because they know it gets shut off at that point.
255
And what I've heard, and this was also like my own internal experience, is like you don't really notice things until they get really bad.
256
And so, you know, if they're just stealing a few thousand dollars a day from a given API key, it may never get noticed.
257
It's only when they start stealing like 100,000 or more that for a lot of these large organizations, things start to get flagged.
258
Stealing and selling tokens is not the only way that these guys are making money.
259
this interesting post on China Talk, one fish, three meals, where meal one is marking up access.
260
We saw that.
261
You say, here are tokens.
262
I will charge them for you.
263
I stole them.
264
So I'm making money on top of that.
265
Meal two, swapping out the models and inflating the tokens.
266
So we saw that, right?
267
We weren't entirely sure that we were getting real clawed at some point and also inflating the token usage.
268
We spent $2 worth of tokens and then it charged me six.
269
That doesn't make sense.
270
It's 5xing the actual token usage that can kind of get you.
271
And then the third one, this is really important, is the logs are the product.
272
So these LLM companies that are trying to train their own models, what they will often do is called distillation,
273
which is, Anthropic has been crying about it for almost six months now, which is essentially, you can train your model based on the output of another model.
274
And if you have the logs or the output of somebody actually using cloud code,
275
then you can use those logs to actually train your own model on top of that.
276
The data of people actually using your model and saying, yes, this is good.
277
No, that is bad.
278
That is an absolute goldmine to these companies
279
that are specifically a lot of them in China that are trying to make their own version of Anthropic or OpenAI models.
280
And they need this data in order to go.
281
So the way it works with the Chinese labs is, say they're trying to distill Anthropic,
282
they need a way to basically disguise their traffic across a number of different sources.
283
And so they will buy some of this traffic through the account sourcing or the account layer I was talking about earlier, as well as the relays themselves.
284
They do this through a bunch of them basically to reduce like the kind of fingerprint that Anthropik's able to detect.
285
Should you be buying tokens on the dark web from sketchy Chinese website, even if they are seemingly legitimate cloud code?
286
I'd probably not.
287
You have to remember, even if you're running it through a legitimate harness, like open code or cloud or codex, these things are still being proxied,
288
meaning that they can swap out tool calls, which is maybe the tool call says createindex.html they will switch it to create index.ishmall
289
and steal your entire computer contents and all of the IP that's related to your employer.
290
And that would probably end you up in some hot water.
291
These APIs I've learned are flaky.
292
They're slow.
293
You're not always sure you're getting what you want
294
and they are using up way more tokens than you probably think they are.
295
As much as you may hate Dario or Sam for what they are doing here, you are often stealing from people that have unknowingly, accidentally leaked their API key.
296
You may be stealing from somebody who has their Claude Mac subscription.
297
You often hear from people who are saying, I did three prompts and I've totally used up my entire usage.
298
Well, maybe your API keys got leaked.
299
You probably should take a look at that.
300
You are much better off spending your money on a bit more of an open stack.
301
Even if those models are the Chinese variety, I got no problem with that, right?
302
So go to somewhere like OpenRouter and take a look at all the different models that they have available.
303
And a lot of these are significantly cheaper for almost as good of.
304
They're not as good, but almost as good.
305
And if you're just building the CRUD app, you don't need some of these best models out there.
306
You're going to get a lot more value out of that.
307
Switch to a harness that is a bit more open, that allows you to switch between all the different models.
308
So what I mean by that is go grab OpenCode
309
or Pi or any of these other harnesses that will allow you to switch between different models and different providers.
310
And I think that's a much better deal than going on some sketchy website
311
and buying what you think might be legitimate tokens that are probably stolen.

Tải Ứng Dụng

Mọi thứ bạn cần để nói trôi chảy

Chấm điểm AIChấm từng câu nói
Luyện IPAThành thạo mọi âm
Từ vựngMở rộng vốn từ
Game Từ vựngHọc qua chơi

Cùng kênh

FOR YOU

Video gợi ý

Về bài học này

Bài học hôm nay sẽ giúp bạn rèn luyện khả năng nghe nói qua video thú vị về thị trường chợ đen cho các token AI. Chúng ta sẽ khám phá một số thuật ngữ và cụm từ hữu ích, đồng thời luyện tập cách phát âm tiếng Anh chuẩn thông qua việc nghe và nhại lại các đoạn hội thoại trong video. Đây là cơ hội tuyệt vời để bạn nâng cao kỹ năng giao tiếp và hiểu biết về các khái niệm mới trong công nghệ.

Từ vựng & cụm từ chính

  • Black market: Thị trường chợ đen
  • API tokens: Token API
  • Subscription: Đăng ký
  • Legitimate: Hợp pháp
  • Access: Truy cập
  • Shared account: Tài khoản chia sẻ
  • Credit: Tín dụng
  • Proxy: Proxy (đại lý)

Mẹo luyện tập

Khi bạn luyện nói tiếng Anh, hãy thử áp dụng phương pháp shadowing với đoạn video này. Bạn có thể bắt đầu bằng cách nghe từng câu và sau đó nhại lại theo ngữ điệu và tốc độ của người nói trong video. Hãy chú ý đến phát âm tiếng anh chuẩn và cách ngắt quãng trong câu. Nếu bạn cảm thấy tốc độ quá nhanh, đừng ngần ngại dừng video lại để lặp lại từng phần một. Đây là một phương pháp rất hiệu quả để cải thiện luyện nghe nói qua video và giúp bạn cảm thấy tự tin hơn khi giao tiếp.

Hãy nhớ rằng, việc luyện nói tiếng Anh không chỉ đơn giản là phát âm mà còn là cảm nhận và truyền tải ý nghĩa của từng câu từ. Hãy cùng luyện tập thường xuyên để thấy sự tiến bộ trong kỹ năng của mình nhé! Bạn có thể khám phá thêm các nguồn tài liệu trên shadowing site để nâng cao khả năng luyện nói tiếng anh của mình. Chúc bạn thành công!

Ngữ pháp trong video

Những cấu trúc người nói dùng nhiều nhất, kèm đúng cụm từ trong video:

Cấu trúcTrong video
Thì hiện tại hoàn thành tiếp diễn have/has been + -ing — việc bắt đầu từ trước và vẫn đang tiếp tụcthey've been swapping · I've been interacting · has been crying
Thì hiện tại hoàn thành have/has + quá khứ phân từ — việc đã xảy ra nhưng còn liên quan đến hiện tạiI've heard · you've ever used · I've understood
Câu bị động be + quá khứ phân từ — nhấn vào việc xảy ra, không phải người làmis needed · being used · be searched
Mệnh đề quan hệ who / which + mệnh đề — thêm thông tin về người hoặc vậtpeople that are · people who want · someone who wants

Phương Pháp Shadowing Là Gì?

Shadowing là kỹ thuật học ngôn ngữ có cơ sở khoa học, ban đầu được phát triển cho chương trình đào tạo phiên dịch viên chuyên nghiệp và được phổ biến rộng rãi bởi nhà đa ngôn ngữ học Dr. Alexander Arguelles. Nguyên lý cốt lõi đơn giản nhưng cực kỳ hiệu quả: bạn nghe tiếng Anh của người bản xứ và lặp lại to ngay lập tức — như một "cái bóng" (shadow) đuổi theo người nói với độ trễ chỉ 1–2 giây. Khác với luyện ngữ pháp hay học từ vựng bị động, Shadowing buộc não bộ và cơ miệng phải đồng thời xử lý và tái tạo ngôn ngữ thực tế. Các nghiên cứu khoa học xác nhận phương pháp này cải thiện đáng kể phát âm, ngữ điệu, nhịp điệu, nối âm, kỹ năng nghe và độ lưu loát khi nói — đặc biệt hiệu quả cho người luyện IELTS Speaking và muốn giao tiếp tiếng Anh tự nhiên như người bản ngữ.

Phương pháp shadowing: đọc hướng dẫn từng bước đầy đủ →