logoalt Hacker News

harshreality • yesterday at 10:49 PM • 1 reply • view on HN

Imagine you have a small forum that can run cheaply on one server/k3s cluster. You want to classify new content to detect most spam and ease administrative burdens. You rate limit classification[1] so you won't be hit with surprise bills. Sounds fine? Nope, because now any idiot can submit CSAM and have your AI provider drop you.

So basically, table stakes for running a user-submitted-content site will be to pay for your own GPU node/cluster and manage your own inference stack (running a worse model than you could use in the cloud). Otherwise you're at the mercy of opaque and kafkaesque (by necessity) service termination without warning for CSAM (real or hallucinated), and subsequent refusal to communicate about anything related to it.

[1] either by queuing all posts for classification before showing them publically, or just best-effort after-the-fact detection of spam or abuse, and upon detection they're moved to a spam queue.


Replies

fc417fc802 • yesterday at 11:17 PM

> So basically, table stakes for running a user-submitted-content site will be to pay for your own GPU node/cluster and manage your own inference stack (running a worse model than you could use in the cloud).

Or just use one of the many third party inference providers that markets an image classification endpoint intended for user generated content.