Every Model Learned by Gradient Descent Is Approximately a Kernel Machine | Dark Hacker News