one of the most ominous papers i've read, where ppl training AI to discover ways to improve human health instead go "hm, i wonder what happens if we just flip the direction of training from 'help' to 'hurt'"
nature.com/articles/s4225
Conversation
Replying to
I'm more and more inclined to think we need a Butlerian Jihad against AI research (including my own former work on neural networks, genetic algorithms, & evolutionary robotics)
6
2
66
Show replies
It doesn't take much effort to un-improve human health. We don't really need an AI to teach us that. ;)
Replying to
Wow how selfish do you need to be to discover this and then decide to publish it for the entire world to know?
9
4
If you are setting your training space around VX and you get a bunch of stuff that seems like VX well what did you expect. What I really want to see is someone build something from the molecule space they claim is separate from the known poisons.
1
16