It’s one thing to think it’s obvious, but quite another to prove it. I think this is the true value of this kind of work, is that it’s helping to decipher what these models are actually doing. Far too often we hear “NNs / LLMs are black boxes” as if that’s the end of the conversation.