Stanford AI Lab (@StanfordAILab) unfollowed @ErikJones313 on Sep 9, 2024

RT @sea_snell: On difficult problems, humans can think longer to improve their decisions. Can we instill a similar capability into LLMs? An…

Really nice concurrent paper from @DavidGlukhov describing when seemingly benign outputs enable misuse. I especially like the "information theoretic" view: to reduce misuse, optimize to reduce the adversary's information gain wrt to a malicious task rather than simply suppress it