What's this about?
WHALE (25 upvotes) alternates optimization: update the weights under the current harness, then search for a better harness under the updated weights — because optimizing either alone leaves the system bottlenecked by its frozen counterpart.
What exactly is joint optimization?
Tuning the runner and their shoes together instead of separately. Model and scaffold co-adapt; a harness ideal for one checkpoint is stale for the next.
Why is this worth noting?
It closes the loop the other harness papers open: not just "the harness matters" but a recipe for evolving both halves together — a small paper marking where the field is headed.