Eric Michaud on scaling, grokking and quantum interpretability
The Inside View
00:00
The Power Law Governing How Useful Knowledge Is for Prediction
There's a power law governing how useful these things are for prediction. So you could imagine like there are some facts like I don't know like Max Tegmark is a physicist at blank where like there's some frequency that like knowing this like piece of knowledge is like useful prediction. The most frequently useful pieces of knowledge maybe will be learned by the small networks and then the big networks will have circuits or they'll know more niche information.
Transcript
Play full episode
