Okay, lets start with the basics. This is a service that uses chat as a front-end to receive a task, then processes
words to extract the meaning understandable for a machine, then comes up with a multitude of answers, picks one with the highest rating evaluated by several
metrics, and then responds back.
So, myth #1: machine can take into account thousands of data points, just feed it stuff and let it think.
It doesn't if you can't describe the parameters of the task and teach it what is right and what is wrong. That's the semantics part when you define a task, and the metrics part, when you try to evaluate the correctness of the answer. You don't need a large language based model for that, you'd rather prefer to work with something closer to the procedural turn based evaluation of turn sequencing and logic behind decisions clearly explained in a concise output.
Myth #2: AI draws well, so you can feed it save files and teach it to "think" well too.
How do we know it draws well? We look at the results and bust the abominations. This comes from misunderstanding of human brain capability to perceive something alien in sound/visual information. We may miss a mistake at a glance, we may even accept the unreal, but once we notice it sorta haunts us, can't look past it, our counciousness alarms to notice. We know something is not right the second we pay a little bit of attention to the details, and trained professional's eyes/ears can get by with less information or less exposure on time scale. That means the verification process for AI produced results can be extremely efficient when it comes to images, sound or text written in native language.
With algorithms involving variables it's not that apparent, human brain will always diminish them to lockstep sequences, just to narrow the scope and verify the answer with surety. Because of such isolation, such calculation of different branches 1 by 1 and evaluation of right/wrong within a slice, the metric to teach an AI becomes efficiency-based. Not style-based or preference-based, as a human would do, and certainly not the best achiavable by heuristics, i.e. the resulting model wouldn't be able to fully trace a path, fail and adapt while abandoning the ballast to squeeze a couple more turns out of the chosen strategy. Lockstep analysis just eliminates such broad thinking, and large scale language models are no good for defining a task, AI tools would have to be super-tuned for Civ only to produce human level of thinking and commitment rather than simple tempo boosting.
PS
The best way to make computer opponents in Civ any good is adding agenda while scripting their gambits. Maybe even add a separated level of rewards as an incentive to pursue blitz to a power milestone. Not only it's extremely simple to produce and control, it also telegraphs other opponents what they are supposed to do to disrupt short term strategy. Ultimately it's a game, it's supposed to be open enough for a player to understand what game wants them to do, not some shogi tournament with one half of the playfield covered with a fog.
