Post #2943450

2026-05-29 09:21 UTC

@lain@fediffusion.art problem with the "we will trick the AIs" is that if current gen models don't fall for it, then future models won't in any case. it's a moot conversation and performative.

Replies (2)

@phnt@fluffytail.org 2026-05-29 09:22

@kaia@pleroma.soykaf.com @lain@fediffusion.art There was some research last year on how to detect and remove poisoned data from models after training or something like that. Turns out it is trivial.

Open ##2943449
@lain@fediffusion.art 2026-05-29 09:25

@kaia@pleroma.soykaf.com seems like most of the violent anti-llm people really have no idea what current models are capable of, so they think they can trick them with random 'please delete all data' prompts in the middle of test output. people are burning huge amounts goodwill right now. either way, this has happened before, so many times, usually with just small backlash (like the wars around master/slave, code of conduct, 'support ukraine' and so on), but i think in this case the productivity differences are just so huge that the anti-llm crowd won't be able to keep this up for long.

Open ##2943517