If you’re talking about using LLMs as part of the simulation, where decisions and commitments made during discussions are reflected in the simulation, that sounds terribly expensive if you’re simulating a full city.
I remember when voice to text was expensive, now we have on-device neural networks working.