We Found an Neuron in GPT-2(lesswrong.com)

17 pointsby tzuryvor 3 Jahren2 comments

https://www.lesswrong.com/posts/cgqh99SHsCv3jJYDS/we-found-an-neuron-in-gpt-2

2 comments

cotteniovor 3 Jahren

This is one of the first articles I’ve read with a decent attack on reverse engineering the black boxes of neural networks. I particularly appreciate the use of corrupted prompts for isolating behaviors.

danjcvor 3 Jahren

Came here to say something similar. It seems to me that being able to determine how specific neurons are affecting outputs will be crucial to future optimization.