3 ms·Gradient Descent on Token Input Embeddings3 points by kp1197 1y agokp1197 1y agoDoes performing gradient descent on token input embeddings lead to interpretable results? And if not, why?