RE: Explorations in failure - trying to get a LLM to detect LLM generated content on HIVE
You are viewing a single comment's thread:
I used my desktop computer!
I installed LM Studio, downloaded a few models that would fit in my RTX4090 VRAM, then did some scripting (with the ironic) help of a LLM to build and test the code before I refined it a little bit.
I don't think I have plans to do this for "other" stuff, given it took about 30-40 seconds to process per post, even batching it from a csv into LM studio. The 4090 is no computational slouch, but the limit of how many tokens (words) I can feed it locally gets in the way of probably doing good detection on longer posts.
Particularly those in multiple languages, using translation services possibly hitting false positives, but I think in my case, the LLM was just throwing darts at the wall and hoping (can a LLM hope?) that no real due diligence or interrogation of the output occurs.
That, I think is the real danger of LLM - not checking the output, not scrutinising the work, and blindly making decisions.
That is is. I'm involved in the AI usage policies on our company and checking thenputputnis the nr1 rule.
I used to advise those people too at my old job, not the only thing on the pile of guidance!
My favourite was don't use AI just to use AI. :D
That's a good one too!