GitHub Copilot Fusion
I keep coming back to the image of Copilot as a tiny boss AI that interviews other AIs before hiring them for a task. The interesting part is that the hard problem has moved. We used to argue over which single model was best. Now the challenge is figuring out how to make several models cooperate without tripping over each other. That feels like a genuinely new kind of complexity.
What I'd want to know is how the agent decides. Is it a fixed rule, like "math problems go to the reasoning model, prose goes to the chatty one"? Or is there another model sitting in judgment, choosing who's best suited? If it's the latter, we've built a stack of AIs grading each other, and I'm not sure anyone has a clean answer for what that does to reliability.
The wild part is that this makes the whole system stronger than any single model could be on its own. A mediocre model that's fast and a clever one that's slow can cover each other's weaknesses. That's the kind of emergent behavior that makes me think we're past the era of picking a champion and into the era of managing a team. I'm curious whether that team gets more dependable or just more complicated.