Making deep learning go brrrr from first principles (2022)