One thing I’m curious about is where the real bottleneck ended up: generating optimization candidates or validating that they were actually improvements.