Because LLM are Markov Chains on steroids. They're useful for sure. But they won't suddenly start to create a better (for whatever better is) version of themselves or start pushing the boundaries of the machines they're running on.

Or maybe I'm wrong and the current "Vibe coding" push is in fact LLMs getting "coders" to compile a distributed AI. Or multiple small agents which goal is to get lot of hardware delivered somewhere it can be assembled for a new better monolithic AI.