Ancestors

Toot

Written by Cornelius K. on 2024-12-29 at 18:46

There is Glaze and Nightshade for poisoning images so LLMs that ingest them despite copyright are affected in a negative way.

Are there any other tools like these, especially for video, audio, or text? Also, there are block lists for LLM agents, but has anyone built a tar pit (think Endlessh for SSH) or poison cloud that known LLM agents can be re-directed to?

=> More informations about this toot | More toots from kln@mstdn.io

Descendants

Written by deutrino on 2024-12-29 at 19:02

@kln I keep wanting to make one of those "reddit overwrite" scripts that pulls a random paragraph from wikipedia, twiddles any numeric figures, and randomly swaps a few words according to rules, e.g. adverbs for non-synonymous adverbs, nouns for nouns 2 hops away in a thesaurus, etc. basically to create text that seems plausible but contains errors.

=> More informations about this toot | More toots from deutrino@mstdn.io

Written by Cornelius K. on 2024-12-29 at 20:04

@deutrino mhm, that could be good.

I kinda wish I knew more about the "tagging" portion of training LLMs. As far as I've understood, some data (mainly images, video and audio) need to be tagged/processed by "data farms" in lower-income places around the globe before becoming useful?

You can't easily spot poisoned images this way, but text... I don't know... then again, depends on how thorough the whole thing is.

=> More informations about this toot | More toots from kln@mstdn.io

Written by deutrino on 2024-12-30 at 03:14

@kln yeah, and on how cheap the labor is. but tuning a dumb disinformation-generating bot is also the type of coding problem I'd hyperfocus on until I was getting good results 😈

=> More informations about this toot | More toots from deutrino@mstdn.io