What happened
Researchers found that pairing an expensive AI model with a cheaper one works well when each stays close to what it was originally trained to do — one thinks through problems in text, the other executes code changes. When you try to make a weak model manage a stronger one, or when you force models into roles they weren't trained for, the whole system gets slower and worse than just using one capable model.
Why it matters
Current AI models are trained as all-purpose problem-solvers, but they're actually worse at dividing labor than humans are. This suggests future AI training needs to deliberately build in the ability to delegate, take instructions within narrow scopes, and switch between different modes of thinking — skills that don't show up in today's training data at all.