humans never stop doing RL which is why old people are wise

they have learned a tremendous amount through RL. they literally have 80 years of rollouts

even the Wikipedia definition of wisdom includes experiential knowledge

@scaling01 the ratio framing is interesting but taste isn't just accumulated RL signal. it's also knowing when to stop updating. models that overfit to feedback lose the thing you were trying to train in the first place

@scaling01 the real problem is youre trying to give a frozen snapshot a sense of taste

humans cheat by just living longer

@scaling01 the framing of RL-to-pretraining ratio being off the charts for humans vs models is the part that sticks

whats the floor for a model to even start developing taste?

@scaling01 Just say "lol" at the end of prompts and the model will start being more creative/original. This should be peer-reviewed because it's extremely effective. Even when doing intensive coding tasks with 5.5-xhigh