I still can't imagine hooking up an LLM to real hardware because of exactly this.