Eric Michaud on scaling, grokking and quantum interpretability
The Inside View
00:00
The Quantization Model of Neural Scaling
Eric Michaud: What if there were a bunch of things that you need to learn to do prediction well in something like language? And so these things could be pieces of knowledge or different abilities to perform certain types of specific computations. So we call these the quanta of the prediction problem. Eric, I still don't really understand your quantum paper. It's possible for smooth loss curves on average to average over lots of small discrete phase changes in the network performance. We'll go more into details of your papers in the next two hours.
Transcript
Play full episode
