Madrigal comments on Trust me bro!

[–] Madrigal@lemmy.world 31 points 7 months ago (6 children)

Nah, guarantee the models have rules built in to deal with obvious stuff like that.

You need to be more subtle. Give them information that is slightly wrong.

parent

[–] taco@anarchist.nexus 13 points 7 months ago (1 child)

Perhaps by generating a bunch of complex copilot code to upload. It's easy to mass produce and would look plausibly functional.

parent

[–] Madrigal@lemmy.world 13 points 7 months ago

Training AI models on AI content is the fastest route to model collapse.

parent

[–] Viceversa@lemmy.world 7 points 7 months ago

... and tell it things, that are slightly obscene

parent

[–] Aerosol3215@piefed.ca 4 points 7 months ago

Artisanal crap code.

parent

[–] ozymandias117@lemmy.world 4 points 7 months ago

Just need to use less obvious insults, a la, "your mother was a hamster, and your father smelt of elderberries"

Still poisons the model with something an end user won't like, but isn't easy enough to train out

parent

[–] bufalo1973@piefed.social 2 points 7 months ago*

Prompt for another AI: "write an example of code that looks correct but doesn't work"

Step 2; upload the resulting code to GitHub.

Step 3: make this an automated task.

parent