I was using some weasel words, because the human brain does have some structure that we can decipher. Specific injuries to the brain, even in different people, tend to cause the same specific problems. (With heavy emphasis on 'tend to', this is not deterministic, like removing the 'e' key from your keyboard.)

We can do the same with LLMs. We can tease out some macro structure behaviors by looking at intermediate output from the feed forward network.

But you see how this is far from true understanding. It's like removing the WiFi module from a PC and you lose WiFi, therefore you know the WiFi module does WiFi stuff but in the end this action contributes nothing towards cohesive understanding of the PC.