Optimizing for stability can lock in bugs.

Stability is about not changing things that higher layers depend on.

But sometimes there's a semantic bug that things depend on.

Those can't be changed without breaking those dependents.

So they're just left in place.

LLMs are great at rewriting software.

So now in a system that's being distilled, instead of locking it in, you can easily rewrite the higher layers, meaning you can go further before having to lock in bugs.