ATRIUMsearch → argument graph
ClaimVideo · 18:49 — 20:01

Black market token sellers make money through three methods: marking up stolen access, swapping models while inflating token counts, and selling usage logs to competitors for model distillation.

The host outlines three ways sellers profit: first by marking up stolen tokens, second by substituting cheaper models while inflating token usage, and third—most importantly—by selling user interaction logs to AI companies for model training through distillation. ✦ AI generated

Wes · Syntax · 2026-08-07 · original ↗

starts at this moment · 18:49

Meal one is marking up access. We saw that. You you you say here are tokens. I will charge them for you. I stole them so I'm making money on top of that. Meal two, swapping out the models and inflating the tokens. So, we saw that, right? We weren't entirely sure that we were getting real cloud at some point and also inflating the token usage. And then the third one, this is really important, is the logs are the product. So, these LLM companies that are trying to train their own models, what they will often do is called distillation, 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. And if you have the logs of the output of somebody actually using Claude code, then you can use those logs to actually train your own model on top of that.

verbatim transcript · starts at 18:49

Transcript · around this moment

18:30never get noticed. It's only when they start stealing like 100,000 or more that for a lot of these larger organizations things start to get flagged. >> Stealing and selling tokens is not the only way that these guys are making money. This interesting post on China Talk One Fish Three Meals where meal one is marking up access. We saw that. You you you say here are tokens. I will

18:49charge them for you. I stole them so I'm making money on top of that. Meal two, swapping out the models and inflating the tokens. So, we saw that, right? We weren't entirely sure that we were getting real cloud at some point and also inflating the token usage. I we spent uh $2 with the tokens and then it charged me six. That doesn't make sense. It's five x'ing the actual token usage.

19:13That could kind of get you. And then the third one, this is really important, is the logs are the product. So, these LLM companies that are trying to train their own models, what they will often do is called distillation, 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. And if you have

19:37the logs of the output of somebody actually using Claude code, then you can use those logs to actually train your own model on top of that. The data of people actually using your model and saying, "Yes, this is good. No, that is bad." That is an absolute gold mine to these companies that are specifically, a lot of them in China, that are trying to make their own

20:01version of Anthropic or OpenAI models, and they need this data in order to go. >> So, the the way it works with the Chinese labs is uh say they're trying to distill Anthropic, right? >> They need a way to basically disguise their traffic across a number of different sources. And so, they will buy the some of this traffic through the account sourcing or the account layer I was talking about

20:28earlier, as well as the relays themselves. They do this through a bunch of them basically to reduce like the kind of fingerprint that Anthropic's able to detect. >> Should you be buying tokens on the dark web from sketchy Chinese website, even if they are seemingly legitimate Claude code? Uh probably not. You have to remember, even if you're running it through a legitimate harness, like Open Code or Claude or Codex,

Around this claim