Gradient Descent on Token Input Embeddings

lesswrong.com

3 pointskp11971 comment

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