kp1197 19 hours ago

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