@GossiTheDog I can almost picture the anthropic tech employee: "I was just following orders and doing my job. The statistical model predicted with 99% accuracy the targets that would have been bombed by human decision makers in past wars according to data".
@elrohir training it on data from bombings on Germany in WW2 and bombing every town that is in reach
@elrohir "couldn't find our primary target, saw a church tower so we bombed around there."
@elexia That's the thing with this entire AI nonsense. It does reward-hacking by following LITERALLY its training reward function. It feels like everyone involved has never heard of the cautionary tale of the Monkey Paw Wish.