kp1197 17 hours ago

Does performing gradient descent on token input embeddings lead to interpretable results? And if not, why?