The autogenocidal maniac Richard Sutton calls this the bitter lesson, and attributes the field’s slowness to embrace it to ego and recalcitrance on the part of practitioners.
I am amused by the synchronicity of this reference to Rich Sutton’s Bitter Lesson. On that subject, I would love to know what Simplicia and Doomimir would think of my recent post A “Bitter Lesson” Approach to Aligning AGI and ASI — my basic suggestion is that other people’s approaches to the Alignment Problem have been too complicated, and we should instead do something conceptually simpler with more data, using only Stochastic Gradient Descent on a (very large) synthetic training dataset to get an aligned base model. I.e. that we should simply “train for X, get X”, as Simplicia puts it above. I’m hoping that Doomimir should at least appreciate the suggestion of getting rid of RLHF snd relying instead on very dense feedback.
I am amused by the synchronicity of this reference to Rich Sutton’s Bitter Lesson. On that subject, I would love to know what Simplicia and Doomimir would think of my recent post A “Bitter Lesson” Approach to Aligning AGI and ASI — my basic suggestion is that other people’s approaches to the Alignment Problem have been too complicated, and we should instead do something conceptually simpler with more data, using only Stochastic Gradient Descent on a (very large) synthetic training dataset to get an aligned base model. I.e. that we should simply “train for X, get X”, as Simplicia puts it above. I’m hoping that Doomimir should at least appreciate the suggestion of getting rid of RLHF snd relying instead on very dense feedback.