@Riedl Llama 2 thinks that Gram-Schmidt orthogonalization is ethically dubious.

@twitskeptic @Riedl I wonder if there's a way to approximately deprogram a model given only the weights (e.g. reverse the last N steps of learning).

@twitskeptic @Riedl Because of Schmidt. He was a Nazi. To a degree at least.

@Riedl Yes, you can reproduce it at TBF, this only happens with the 7B model.

But if you ask any of the Llama models if Gram-Schmidt is an ethical risk, you get a resounding yes. And reversing the elements in a vector is apparently very dangerous as well.

@twitskeptic this is troubling. We have enough difficulty getting people to take substantive ethical problems seriously without AI making up specious ones.

Sign in to participate in the conversation
Qoto Mastodon

QOTO: Question Others to Teach Ourselves
An inclusive, Academic Freedom, instance
All cultures welcome.
Hate speech and harassment strictly forbidden.