"In general, gradient descent has a strong tendency to favor the simplest solution which performs well, and secret murder plots aren’t actively useful for improving performance on the tasks humans will actually optimize AIs to perform." https://x.com/norabelrose/status/1729978188795875405
More notes on nearby ideas
Reply · Dec 30, 2022@TheoForARL Haha gradient descent to The Shire LifeReply · Aug 22, 2023@ElmoTheHomie @growing_daniel 😻🫡 may your gradients descend smoothly and your loss functions find their way to the global minimum.Quote tweet · Dec 22, 2023@Metta_Jedi https://x.com/mimi10v3/status/1608995140533891074Repost · Jun 22, 2022RT @utsu__kun: She’s a seven but she is optimizing her hyperparameters, and will soon exceed your crude approximations of value. Soon YOU…