Standard, Huge, Deity, and usually Shuffle map
Yea, sorry if I tilted ye a bit. I read your post and realised that it looked like I was replying to you. That really was not my intention.
I won Immortal last night in 6 hours on small/standard/Pangaea with Rome so I have to completely concede your points as well
If you are interested in why I place so much weight in map size and game speed in terms of difficulty, here's a very brief explanation. There will be heaps of specifics that I'm wrong about and I haven't actually looked at Civ 6's game files - this is based off maybe Civ 3/4 & possibly 5.
Most people think strategy game AI is like a chess AI where it is analysing/reducing permutations in a way that replicates deduction (I doubt chess AIs still function this way). This isn't really achievable (naively at least) in Civ because every piece moves like a queen that can turn corners, there can be hundreds of pieces later in the game, the gameboard isn't uniform etc etc.
In reality, whenever I've gone to mini-mod/tweak strategy games, the AI is always defined in terms of
subsumptive behaviours which are a very approximate approach to modelling behaviour in very complex contexts.
These a basically a bunch of pre-canned behaviours like "attack the lowest health unit" or "run away". Each performs some specialised analysis and produce a number that is given a weight in the moddable files. So someone who is modding a game can only adjust which behaviours a unit has and their weight, not how the behaviours actually perform their analysis.
E.g. an archer might have a behaviour list like this:
1) move into range of enemy ranged units and attack: 40%
2) move into range of melee and attack: 60%
3) move to opposite side of river from infantry: 60%
4) move to opposite side of river from cavalry: 0% (because you'll get rekt, son)
And in a given game, there might be a situation where the behaviour's individual analysis comes up like this:
1) 50 2) 70 3) 90 4) 100
multiplied by the weights:
1) 50 * 40% = 20 2) 70 * 60% = 42 3) 90 * 60% = 56 4) 100 * 0% = 0
The highest is 3) 56 so the archer moves to the opposite side of the river from an infantry unit. (Please forgive me if I've forgotten my 10 times tables - it's been a while since I've actually calculated anything myself.)
etc etc - and you can already see that, in actuality, we need those behaviours to be broken apart and fed into each-other in a deductive way. The list doesn't actually make sense but does *fully* explain why the AI is so appalling at combat. This is also why Victoria's combat puzzles are such a powerful insight into beating the AI.
If you follow some of the links on the wiki page, you'll probably encounter various academic solutions to these problems along with some analysis of their computational complexity (how long it takes them to run). Real solutions are a lot more open ended than that but the academic stuff is very important for understanding how concepts are related to each-other and the feasibility of the solutions.
In comparison to neural nets (called "deep learning" or "machine learning" these days), this stuff can/has been referred to as "classical AI". It's a crap ton faster than neural nets. It might take Firaxis 5 years on a server farm to create a machine learned AI for Civ (it might take months or days too). Every time they changed the rules (i.e. patches), they'd have to retrain it as well.
When applied to things like city placement, the AI will probably analyse the tiles in a city's radious as it's highest priority behaviours. For example, if you have some massive eruptions early in the game, the AI will often settle cities on those sites whereas a human might realise that most of the tiles are mountains and settle off to the side, with an understanding that, in practice, we only need 3 of the 6 high yield tiles.
I haven't actually looked at Civ 6's files but in the past there hasn't been a separate list of behaviours for different map settings as, somewhat obviously, it'd take a million years to playtest them all (or a generative automated testing framework and game designers who stayed within the constraints that enable that - which is a general rant about programming that I shouldn't get into).
Going back to the city placement example - as humans, we can calculate the time sensitivity of our decisions and understand that the first 3 unimproved tile yields are much more important than the rest. The AI gets free builders to combat this and free settlers to make up for their terrible city placement. It also appears that the AI basically can't play the adjacency game and certainly does not respond to patch changes or differences between Civs (although I think it does change the priorities of the districts it builds based on Civ - again, haven't looked at the files to know this).
When you make the map larger or expand the time scale, you're increasing the number of bad cumulative decisions the AI makes VS the number of good cumulative decisions a human makes. That's why, even with all the AI's advantages, you'll see the player with an average city pop of 7 say whereas the best AI is around 4 or 5. On Deity where, as you know, the AI's advantages are utterly insane.
Adjacency and 1UPT aren't well suited to the AI's decision making method so this problem is disproportionately worse in Civ 6. This is why small/standard & either continents, pangaea or island plates are the only settings where the AI actually works in Civ 6.
It is also why the people saying the game is harder in June 2019 might be right. The OP was clicking "Play Now" which is applying the settings for which the AI has been play-tested. Additionally, June 2019 has changed the player's decision making and punished a small army early with barbs - these are both things that change (i.e. temporarily decrease) the ratio of bad AI decisions to good player decisions.
That took a while to explain sorry. Hopefully you see what I'm getting at. Again, play how you want but understand that if you want to play huge maps against a decent AI, you have to play Civ 4 because the AI can't cope with Civ 5/6.
