I'm no expert, but I think the future is more about extremely low latency and low power chips with LLMs etched directly onto them. You can create specialized chips that function as "neurons" in a larger system, generating the needed reactions with a very clearly defined set of constraints.