The dirty secret is we think we want the AI to play more realistically, but actually don't. This is true in many games but especially Civilization, where we rely on the AI behaving predictably to feel mastery over the environment, and for diplomacy to feel "fair". If the AI were playing to win, diplomacy would mostly evaporate as a game system. It'd break the fiction of the game.
That's what the developers say, but I think it's part true and part a cop out. The AI in Civ 5 & 6 is just so so bad, that it's only challenging when you play on the really hard modes that gives them all sorts of cheats. And then it's a bit less fun because you're walking the line between impossible because of the cheating and having to essentially exploit the AI because of the cheating.
I don't necessarily want an AI that tries to play like a human, though it would be a fun option, I want an AI that isn't just straight up terrible given the same starting resources and rules as the player.
Part of the problem, I think, is that each iteration of Civ, they make the game more complicated in a way that makes it even harder for the AI to do as well as a good player, but probably most players don't care and there isn't a lot of reason to become good at the game anyway.
Actually, if you play Stellaris, there is a mod called StarNet AI, which uses weights that are considered "meta". From my very limited personal experience (I got destroyed), it is not really fun to play an AI that uses very good strategies as it eliminates all previously viable, but not optimal playstyles.
Also, it feels _really_ good when you get to a new ship tier faster than your neighbors. The reverse is, however, very frustrating, as you are forced to play catch.
This comes up often in discussions about Axis & Allies Online. It has an AI, but it’s pretty basic and really only there as an aid to learn the rules before playing humans. The main way the game is played is ranked play on the built in ladder against humans, or custom games. There’s a discord that runs regular tournaments as well.
The problem is AI for complex games like that is absurdly hard to develop. The combinatorial complexity of a game like Axis & Allies is something like a hundred orders of magnitude greater than chess. It’s probably similar with Civ, probably a lot more so.
But aside from just competency, what makes playing humans so compelling is personality. Human players range the full spectrum from terrible to excellent, but even beyond that they vary massively in the ways they are terrible, and the ways in which they are excellent. With A&AOL there are top tiered players that employ radically different strategies, to great effect.
IMO it's the reverse. I could swear some of these games intentionally keep the gap between good-play and bad-play very small so as to hide the deficiencies of AI.
The lack of choice in buildings, reasonably inconsequential bonus tile yields, minimalist tech tree where you have to research everything, lack of synergies between buildings-resources-terrains, etc. all make a lot more sense when you see what horrendous decisions the AI makes even on hardest difficulties.
Sid had achieved competent but un-fun AI early on in the series and backed off. Players hate it because of how good it is at sacking poorly defended cities, especially early on when the AI knew too much about the global map.
A good AI shouldn't need to know any more about the global map than a human player would in the same situation. If they made it see the whole map, it's basically cheating.
In general with strategy game AIs it’s not feasible to make it play at decent human level, so the way they make it challenging to play is by cheating in various ways. Resource boosts, global vision, etc.
In particular, humans are very good at reasoning based on limited information. We can form hypotheses about where resources or objectives might be, or if an enemy unit goes out if the visible map, estimating where it could be on a later turn, or what it’s presence indicates about its home civilisation’s disposition out of sigh. That sort of thing is extremely hard to program, so the only way to compensate for the AIs inability to intuit information is by actually giving it the information.
Meta made an AI to play Diplomacy which is able to talk to other players and play the game. It doesn't use a LLM, but it was able to win a tournament against human players.
I suppose what I mean is, in the context of a computer game where you’re playing a computer, I don’t think it makes sense to talk about cheating. The way the game works is the rules. There may be different rules for the human and AI players, but the expectation that they are the same is an assumption ported over from board games. It’s not really a thing in native computer games. For example nobody expects the computer opponents in a FPS to obey the same rules as the human player. So I don’t think cheating is really an applicable term.
My point is that it's kinda weird to claim that we had AI that was so good at opportunistically playing "like a human" by e.g. picking on poorly defended cities that human players hated it, but then admit that its proficiency is at least in part because it knows the whole map - that, by definition, is not "like a human".
The means is not human like, but the behaviour might well be. A human might conduct a search with scout units and intuit city locations from observed units and such to identify and target cities. Doing that in software might be infeasible, so you give the model full information and maybe program in a delay based on the distance to enemy cities before targeting them. The implementation is different but the behaviour ends up being hard for human players to distinguish.
Yes, but the problem persists since a truly competent AI will have good guesses about where and when to strike into your hidden territory based on what’s revealed to it, using the same advances that give computational room to respect fog of war.
The main thing I'd want in a Civ AI is to be barely competant at combat. Right now, if you can get a couple archers up in time, you can beat an AI army that's 2-5x your size depending on the terrain. Yeah, it feels bad as a beginner when you haven't bothered to build any military and an AI comes and kills you, but a beginner should be playing a beginner mode.
I'd love God mode AI to put up any sort of half decent fight without needing to have so many more units that just go through an endless meat grinder because they can't coordinate or figure out how to back off and heal.
Realistically, if the Civ team were to want to do this, here's how I'd go about it:
1) Break the problem down into strategic goals and tactical actions, and then group the tactical actions into areas (e.g. diplomacy, resourcing, combat, research).
2) Process recorded online games of human players into streams of relevant events / choices.
3) Continually retrain your models on the output of 2.
You should thus have an AI that approximates the meta approaches of humans, which is really what we're talking about.
New players look exactly like bad AI: making poor choices, oblivious to things "the players community" already knows are always optimal, etc.
So really the problem is "How do we build a model that already knows what most of our players do, at this stage of the game's lifecycle?"
Which is essentially what one of the recent fighting games(?) did with their AI.
Is it that players hate good AI or that flagrant cheating breaks immersion? I think the latter would be an interesting AI target: can you make something which plays well but doesn’t do things which are obviously impossible for the player to do in-game - e.g. helping the AI with some extra resources isn’t glaring but having attacks perfectly target things the AI shouldn’t be able to see, or instantly repair damaged units, etc. is something you could never duplicate even if you were the best player in the world.
SMAC had a few restrictions that made dealing with its leaders harder, but more realistic. In Civ 4 and later Civs you can build complex deals and propose them to the AI leaders. If they are bad, they tell you they are not acceptable and you can try again and again. You can see the "diplomatic price" of each item.
In SMAC you could make a single proposal per turn. If it was good, they would accept it, if it was almost good, they would counter your offer, if it wasn't good enough, they would actually get madder at you and stop talking for the rest of the turn (or for a few turns if you relations were that bad). That was the only feedback you'd get.
This small change made parlaying with AI leaders feel more realistic. Instead of going "Hm, you won't give me Zaragoza and all your money, because 'we are losing' is only at +15, I'll come next turn when it should be at +17" you start building a mental model of each leader and proposing deals you think they should accept.
This is being repeated during every AI comptenecy discussion, giving an example of "you wouldnt want AI that never misses in FPS". But its completely missing point.
We want AI that is capable to never miss, and then is tuned down in a controllable matter. We want smart AI that are dumbed down in precise ways to create illusion of dumb AI, not dumb AI boosted by resources and damage modifiers to give illusion of a smart one.
You can make smart AI look dumb (thus weaker), but hardly ever you can make dumb AI look smart, thus now the strong AI is just dumb AI with advantage
My definition of a smart AI is an AI that is smart at mimicking a human - that means as a Game AI it would play an entirely different game of "mimicking a human" in essence.
The issue is that a lot of things a computer controlled player does is entirely too easy to implement with just a few random rolls. This starts to fall apart if you want to vary the difficulty. So you can get a credible if disdained "AI" in a lot less effort that say training a Neural net to predict what action a human would likely take next.
There simply is no "GPT-1" AI of playing a game yet much less a "GPT-3" level one and I do not think that it would be entirely easy at all to "tune" it in any way to get a consistent difficulty scale.
Totally disagree. Civ increases AI difficulty by gifting the AI free army units, resources and settlers. There are currently no 4x games where the AI can challenge an experienced player on even terms - the first game to do this (while also being a good game in and of itself) will be a massive success.
I remember (from HN) some genetic algorithm explorations of optimal research/build orders that discovered novel and better combinations then-unknown to the playerbase.
> The dirty secret is we think we want the AI to play more realistically, but actually don't.
The dirty secret is that we don't know if we want that. So far we don't have technology to make Civ AI play remotely like a reasonable human (at least not on a consumer grade PC). So we have absolute zero data point on how players will react to that.
I think evidence of this in the popularity of matchmaking systems. Players engaging in that aren't primarily looking for meaningful human interaction (they often mute the chat), but rather a human-skilled opponent.
One area where current advances in (chatbot-style) AI would make a lot of sense to apply in games is to make non-critical (for gameplay) interactions with NPCs more interesting, just as "background noise" to make the game world feel more alive.
E.g. you could start an "interesting" conversation with a random NPC in GTA or Hitman instead of getting the same pre-canned reactions over and over.
The downside then however is that the scripted characters which are important for driving the story forward will feel like pre-programmed robots compared to any 'unpredictable' random NPC ;)
If you can make an AI that can play like a smart human, you can make an AI that plays like a dumb human. And you can design a game that rewards cooperation between smart or dumb players.
You want it to behave predictably. That doesn’t mean it needs to be predictably incompetent, and it doesn’t mean predictably optimal. Predictability is orthogonal to competency.
If the diplomacy mechanic cannot survive with players playing to win, then it’s clearly the diplomacy mechanic that’s wrong — diplomacy in reality is an inevitable result of multiple players playing to win; you don’t do diplomacy for fun, but to shift the world-state slightly in your favor by unifying against common threats.
Something must be fundamentally wrong with the model if it cannot encourage the most basic action within that model, without a player acting completely against his own interest
I think maybe the thing to do is not so much use LLMs (which I would still use for generating variety) for the AI's decision making, but to bring to bear the many lessons from TikTok and various other attention engines that we find on the internet.
Let the AI know everything, in the way that a father knows everything when teaching chess to his kid. But instead of just nurturing an independent chess player, feed the player with dopamine experiences that keeps him playing. Early victories, the joy of discovery, interesting diplomatic situations. Once they can play better, stretch it out a bit, more intricate setups.
I'm not sure why you would think they, when multiplayer games obviously show that intelligent opponents are nice to play against and that diplomacy is not only possible but much more interesting between humans.
Even here, comments are pointing out how bad Civ's AI is. The solution isn't more realistic AI. It's better deterministic "AI" that allows you to learn how it plays and devise different strategies for how to deal with it. Unfortunately, AI is largely an afterthought for developers and players don't really understand what they want either.
That just leads to everybody complaining about how bad the AI is in Civilization and Stellaris. Sure, these companies don't care because they have financial incentives not to care, but we're seeing the effects of studios not caring about AI right now. Not sure why you're defending them.
By «everybody», you probably mean the most experienced players, since it takes a good 100 hours for a human player to start being competent at the game, while your median player only has 40ish hours under their belt.
AI doesn't come free, GalCiv2 has been built around the AI being competent, but is IMHO a boring game as a result.
I guess I would agree with you for Civ 5 (and 6 ?) where the decision to switch to the 1 unit par tile system has crippled the AI... (while being pushed by Firaxis wanting to make the game more friendly to those sub-median-experience players !) But then I would argue that this system is bad for Civ itself, and this is an outlier.
(There's also the example of both AI Wars being built around embracing that the AI plays a different game, which from I hear works very well. It's questionable how applicable that is to your average 4X.)
> By «everybody», you probably mean the most experienced players, since it takes a good 100 hours for a human player to start being competent at the game, while your median player only has 40ish hours under their belt.
Nah, I mean everybody as in pretty much anybody I see talking about AI in these games, it's almost always negatively. And honestly, I don't care how experienced players are. If your average player is complaining about your AI, then you've failed.
Experienced players are just better at quantifying and identifying how AI fails and how it affects the gameplay.
Honestly, I struggle to see what your point is. That devs just shouldn't bother with AI?
And how likely is «pretty much anybody you see talking about AI in these games» to be overwhelmingly someone with way more experience than ~40 or even 100 hours ?
My point, trying to restate it in a different way, is that these language models might be low-hanging fruit to make an AI interacting with whom feels more like a head of state (and a specific historic one too !)
Though I'll note that Alpha Centauri has already achieved something like this using a very simple dialog AI (in a specific context).
Current civ6. Harder modes just give massive starter advantages and statistical bonuses.
3 starter cities compared to your 1. That can be brutal to overcome. Top players like potato mcwhisky basically have to know every little thing possible to win.
a lot about the nonsensical lust for 'ai' will be solved, once people realize it's not literal human 'intelligence', it's just a collage of data based on statistical analysis of past data and a prompt.
right, thats why some moves by alphago were called beautiful and deeply creative by human go masters?
idk know why this keeps getting repeated, it is NOT a statistical analysis of past data and a prompt. It is not just pattern matching. It is building models to predict patterns, which is a very different statement. And the fact of the matter is, is that is mostly still an unknown blackbox. It is an area of active research to understand the mechanisms of this blackbox.
And therefore it's "intelligent"? eye roll AI as we know it is nothing more than marketing. At least in the gaming sense everyone knows there are no actual smarts.
On closer examination it appears that you are correct. Sorry, I think I am mainly weary of all the "AI" hype so it's easy to have a knee jerk reaction.
> once people realize it's not literal human 'intelligence'
I suspect that our brains are so profoundly hardwired to recognize certain signal patterns as proof of conscious intent that soon, even experts who should know better are doomed to be fooled. The layman has no chance.
Hard disagree, I find civ to be crappy in single player because it’s AI is so dumb and just gets unfair advantages to remain competitive, it turns the game into just cheesing the stupid AI until you win at harder difficulties. So I ended up only playing multiplayer, but since the devs don’t care about making a reasonable networking stack it just crashes all the time. Hopefully they fix the AI, but if not I hope they actually make the networking stack not suck.
