[RestartKernel]

The real story is in the poison fountain dataset this uses:

https://rnsaffn.com/poison3/

> [...] we want to inflict damage on machine intelligence systems.

This almost strikes me as roleplay, but maybe I'm childish for finding it difficult to empathise with this genre of hacker ideology.

[stingraycharles]

And nowhere does anyone explain exactly what “poisoned data” is and just how the claim that training a model on a small amount of such data will have a big impact.

[Mars008]

There is no exact definition. It can be a harmless bullshit, or something more harmful. Imagine AI recommending child to drink some common liquid for fun or as a medicine. Liquid which is in fact known poison. There are many dangerous things that aren't often mentioned in internet because people have common sense and never do it. It's enough to add just a bit of misleading information.

With cheap generic robots coming this can be a real problem. Human supervision can help when there is one.