11 points | by myk-e 17 hours ago ago
4 comments
I'm curious how these 'distillation attacks' work and what the prompts look like to do them.
I’m pretty sure they just mean DeepSeek was paying to generate text, and then using this text as training data.
Which of course is completely wrong and totally hacking and illegal. Very different from using book torrents or people’s google docs to train your AI.
I’m not sure there’s even an adaptive pattern in prompts. The article sounds like a n>1E5 shotgun.
In my country we say: he who robs a thief has a hundred years of pardon.
I'm curious how these 'distillation attacks' work and what the prompts look like to do them.
I’m pretty sure they just mean DeepSeek was paying to generate text, and then using this text as training data.
Which of course is completely wrong and totally hacking and illegal. Very different from using book torrents or people’s google docs to train your AI.
I’m not sure there’s even an adaptive pattern in prompts. The article sounds like a n>1E5 shotgun.
In my country we say: he who robs a thief has a hundred years of pardon.