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.