Hacker News Reader

Comment by loufe | Hacker News Reader

The comments here are baffling. Is nobody seeing the easy opportunity for gathering amazing high-quality training data by inserting yourself as a MITM? If I were a competing lab, criminal, or opportunist I'd lie/cheat/steal/simply pay the difference to get the chance to listen into real life scenarios of usage of modern models in a high-impact business.

Seems a huge part of the story completely absent to me.

View on HN

There's a strong theory with some rumors to support it that some of the largest customers of these services are labs doing distillation. The resellers are basically crowdsourcing the farming of accounts for them to use.

Doing MITM on other people's sessions isn't as interesting as simply sending your own synthetic requests to the models at a 97% discount. They manufacturer the questions and responses they want to train on.

This is what every chinese "credit reseller" is doing, selling all the prompts and responses to labs in addition to selling quanted kimi/ds/etc as opus/fable

Arena.AI does this as well - they may appear as a leaderboard but what keeps business booming is their sale of traces to chinese labs

What's the source of information?

They are just offering API key swaps, not switching endpoints. No scope for mitm there. I'm sure users would only trust API key swaps, if only to be sure what models are actually being served

Isn’t that essentially OpenRouter and Cursor play.

Reply on news.ycombinator.com