The main issue I see is the spatial component is hard to describe in text. The new vision models make it easier, but still I imagine it's not trivial to integrate all the mechanics plus the spatial component on the limited prompt space.
I do think that 1) combining with the hand crafted AI and 2) having an "LLM advisors" system where for a given aspect (eg military) the "advisor" would present the options and tradeoffs to the "Main AI" and the role of the latter is to weigh the tradeoffs between the options presented by the advisors.
And what I do know is that it could be so much more immersive than the current hard coded AI!
Idefics, mingpt4, Next-GPT and LLava are open source multimodal LLMs that can read images.
Humans are decent at Civ exactly because we play in a naturally heuristic fashion and don’t try to do any tree search. To get an AI to cope with the this vast amount of information, available and hidden, and make sensible decisions without tree search, is a huge challenge. On top of that, it needs to deal with up to a dozen opponents, multiple of which may be human, and avoid getting ruthlessly exploited in trade deals and diplomacy and all the rest.
Here is one basic decision among dozens made every turn: do I accept an open borders treaty with Bob of the Carthaginians or not? I can’t see any armies he may or may not have (fog of war) and he’s never betrayed me before so I guess he’s trustworthy… oops! He used the open borders to move a vast army into my territory and right up to all my cities in a single turn and then declared war!
That’s one decision, among dozens made every turn, that led to total defeat.