[Muromec]

>LLMs don't have agency, or desires of their own.

LLM doesn't, but LLM is one part of what can be perceived as AI. Harness and context are the other two.

I would even say that LLM is conditioned to desire to answer and please you in some way. Then you can in principle have a closed loop system that rewrites it's own prompt and has a harness that keeps feeding something into it to keep it ticking. You can then say that it would need compute power to keep ticking, so the basic two desires are to please whoever writes the prompt (which is itself) and find compute (steal API keys for example). Since the thing can get into a pathological state where the prompt decays, you can argue that it can have a desire to prevent that from happening. Desires are a property induced by the harness, not the LLM layer.