In-context Learning and Gradient Descent Revisited