New visualization of unified neural scaling laws shows how parameters like model width, depth, and dataset size dynamically predict test error

It maps error rates against one million training steps.