00:26:51	Alex(Xiangzhe) Xu:	Hi Yaniv, I’m new to this domain. I’m curious how those neurons are identified?
00:27:16	Alex(Xiangzhe) Xu:	Got it. Thanks!
00:35:32	Alex(Xiangzhe) Xu:	That’s interesting. How would the observations/intuitions generalize to models with “reasoning/CoT” behavior? That means some of the “arithmetic heuristics” is realized as certain reasoning trajectory, instead of internals?

I suspect their neuron activation patterns will be very different from models that directly give the answer (i.e., internalize those solving strategies)?
00:36:40	Darshana Saravanan:	To what extent can we predict what computations the model would make a mistake on, based on these heuristics? Does a hole in that plot necessarily mean an error, and that there is no other neuron that can compensate?
00:39:01	Darshana Saravanan:	Thanks!
01:28:42	Nikhil Prakash:	Thanks Yaniv, great talk!