The models continue to eat more of the stack around them. A few months ago the wisdom was to either move upward into harness engineering, or downward into kernels. Eventually, we will only be left with the highest and lowest of problems:
naming, and cache invalidation