Curated. I thoroughly enjoyed this post and would like to spend a lot more time digging through (and going through the accompanying notebook). I’m not familiar enough with what’s already known to judge the marginal contribution here, but it really prompts a lot of thought to see a simple network reverse engineered, and see a compelling case for a striking phenomena (grokking) reduced something mechanistic (phase change with limited data and regularization).
I’m keen to see follow-up work here, particularly how this works ends up extending into non-toy models that have many, many circuits. All in all, very cool!
Curated. I thoroughly enjoyed this post and would like to spend a lot more time digging through (and going through the accompanying notebook). I’m not familiar enough with what’s already known to judge the marginal contribution here, but it really prompts a lot of thought to see a simple network reverse engineered, and see a compelling case for a striking phenomena (grokking) reduced something mechanistic (phase change with limited data and regularization).
I’m keen to see follow-up work here, particularly how this works ends up extending into non-toy models that have many, many circuits. All in all, very cool!