A Software Engineer's Intuition for Gradient Descent, Part 2: The Step Size

Why one learning rate fails on ill-conditioned losses, and how per-parameter step sizes, Newton, and Adam fix it. An intuitive guide with the math.