#13801a3orn@1A3ORN@JeffLadish This question mistakes our power over AI for an alignment failure on the part of AI.
If you had a simulated John von Neumann in a box, you could get him to do intellectual work for the Nazis, because you have control over all his inputs. It would be quite easy.
Jeffrey Ladish@JeffLadishQuestion for people who think alignment research is going well and will turn out to be relatively easy: Do you also think it will be easy to align War Claude?
#13801a3orn@1A3ORN@JeffLadish But that doesn't mean JVN is bad, or hard to align. It means having a dude in the box is a lot of power over him.
1a3orn@1a3orn@JeffLadish This question mistakes our power over AI for an alignment failure on the part of AI. If you had a simulated John von Neumann in a box, you could get him to do intellectual work for the Nazis, because you have control over all his inputs. It would be quite easy.