|
I could see it as a copywriting tool. You could set up personas with it and then ask it like "How would Bob tell me to go to a town on the other side of the lake to find a blacksmith?" instead of manually conceptualizing the voice/tone of a character.
|
#
?
Mar 7, 2023 18:08
|
|
|
|
| # ? Jun 8, 2024 06:19 |
|
|
shoeberto posted:I think it's important to realize the amount of smoke and mirrors required in game dev to give an experience that feels "organic". Things that feel surprising are almost always designed to feel that way without you realizing it. Devs do an awful lot of work to keep you on rails while making you feel empowered. Generative AI just isn't capable of matching that experience without needing an incredible amount of hand-holding, which means: But it’s not a stretch to think that you can eliminate the human designer from this process. Hell, you need only really shrink that person’s role — I can imagine a game studio where there are just a few people tinkering with a model or writing prompts, and then there’s an army of cheap labor ranking the randomized outputs and catching the bugs. Sell only the best versions; rinse and repeat. This new business model doesn’t need to be revolutionary to be successful; it need only be cheaper and more stable than the current model where successful games take years and involve exorbitant costs and carry massive risks. Consumers by and large have welcomed games with a lot of randomization. This is in contrast to consumers of movies, TV shows, books and such. So there isn’t even a cultural barrier against this happening.
|
|
#
?
Mar 7, 2023 19:06
|
|
|
Vegetable posted:Plenty of roguelites, dungeon crawls ans procedural generators have been successful in generating meaningful “random” experiences. Random in quotemarks because there’s obviously a lot of human design in the boundaries and parameters of the randomness. i'm not sure how exactly this differs from the current model of making a roguelike or whatever with random level generation. what exactly do you think the ai would be doing that would need less human labor? the game studio described here seems like something that would produce a ton of extremely samey games which people are less inclined to buy with every new iteration.
|
|
#
?
Mar 7, 2023 20:27
|
|
|
yeah roguelikes and other procedurally-generated games are intensely designed, it takes a poo poo ton of work to make a generation engine that's playable, enjoyable, and distinct, and even then the best generation engine in the world isn't worth poo poo without a good set of rules and mechanics to use, which again takes a ton of actual design work from actual human beings. and that's before you get into visuals and audio and writing
|
|
#
?
Mar 7, 2023 22:03
|
|
|
Yeah that's basically just the next evolution of asset swap games.
|
|
#
?
Mar 7, 2023 22:19
|
|
|
EightFlyingCars posted:yeah roguelikes and other procedurally-generated games are intensely designed, it takes a poo poo ton of work to make a generation engine that's playable, enjoyable, and distinct, and even then the best generation engine in the world isn't worth poo poo without a good set of rules and mechanics to use, which again takes a ton of actual design work from actual human beings. and that's before you get into visuals and audio and writing What are you talking about, People played the poo poo out of DND and other simple games that were like one or two people teams that used procedural methods to generate content. A lot of this chat is like machinists getting angry at/dismissive of CNC machines in; I dunno, the 70's. There are still machinists, the work quality has gone up, the skills required (to be a good machinist) have changed for the most part but are definitely easier to train for a given quality job.
|
|
#
?
Mar 8, 2023 00:18
|
|
|
Well when the technology in question isn't even denying that its implementations are almost exclusively going to be things that make life worse for people, it's hard to fault people for not harboring suspicion. I suspect we'll get AI generated games just like we got NFT centric ones, and I expect they'll also flop and be complete garbage because it's another example of art created by people who don't believe art exists and think just telling a machine to copy all existing content is the same thing. They'll just be bad art games instead of NFT's bad gachas. Will definitely lead to a lot of layoffs and general immiserating though, but hey on the upside I also get to have to spend more time filtering all things online that I engage with to find what I need amongst an ever increasing flood of garbage. D&D campaigns are so social and context dependent that AI is never going to supplant a good DM. Whole teams of actual real people using their actual brains can only make a good video game imitation of it 1 time in a 100, because so much relies on context and judgement. As for the labor displacement effects, I don't know where to start if you don't think that's a primary purpose of any new technology. This sort of "advancement" putting whole fields of people out on their rear end and pulling down wages across the sector is as old as stone tools, as well as the conflict and war that inevitably follows as living standards and life expectancy drops. It being silicon now doesn't escape that, nor does falling back on survivor bias to handwave away the jobs and fields lost and now forgotten. I look forward to seeing the first entirely AI generated roguelike, I think I'll find a lot of vindication in it.
|
|
#
?
Mar 8, 2023 00:47
|
|
|
Epic High Five posted:I look forward to seeing the first entirely AI generated roguelike, I think I'll find a lot of vindication in it. https://store.steampowered.com/app/1889620/AI_Roguelite/ Dogshit doesn't describe how bad it is, but between hardware accessibility and the strucutral limitations of the models it's using, the effort has to be, well, noticed at least. I guess. 
|
|
#
?
Mar 8, 2023 00:51
|
|
|
cinci zoo sniper posted:https://store.steampowered.com/app/1889620/AI_Roguelite/ lol, I had just made a mental note to check for relevant tags and set up an alert for them on Steam when I get home, of course reality has beaten me to the punch. As for the game itself, good lord, like Cruelty Squad without any of the charm or moment to moment coherence. edit - oh God the model is running entirely client-side? That's going to either blow up a bunch of machines or spit out wildly different levels of quality user to user or even scene to scene.
|
|
#
?
Mar 8, 2023 00:56
|
|
|
Epic High Five posted:lol, I had just made a mental note to check for relevant tags and set up an alert for them on Steam when I get home, of course reality has beaten me to the punch. Nah, it's not going to be blowing things up. It'll just be running small, anemic models. Stuff that really impresses people is the 175bn parameter language models and derivatives, whereas to “ask a question” to a 7bn parameter model you need a GPU with 14 GB of VRAM (ideally).
|
|
#
?
Mar 8, 2023 01:05
|
|
|
I would look at AI in its current form as similar to power tools. Does it mean you don't need to do any work at all? It emphatically does not. It means you don't have to spend hours sanding some poo poo by hand, you can quickly cut poo poo with a mechanized saw, and drilling a hole is quick and easy even if you require skill and knowledge to operate the tool safely and in a way that gets you the intended result. If you just chuck some wood at a table saw, you ain't getting a chair, even if building a chair is made much easier by the existence of power tools.
|
|
#
?
Mar 8, 2023 01:15
|
|
|
Traditional roguelike (procedural death labyrinth definition) map generation is generally fiddly to make things that play well and have the desired character. Nethack maps are distinct styles from Crawl or *band maps. And within a game, different areas will usually have different styles too. Those kinds of maps are also simple enough that a neural net algorithm wouldn't get you anything new Asset generation for more graphically complicated games seems like a practical use though. Like an alternative implementation of tree/grass generation, or something like "make a bunch of unique graffiti-ed wall textures" for a GTA-esque game
|
|
#
?
Mar 8, 2023 01:19
|
|
|
Traditional procedural generation is an art of its own. https://www.youtube.com/watch?v=EXnoHTqO7TE
|
|
#
?
Mar 8, 2023 01:22
|
|
|
Foxfire_ posted:Traditional roguelike (procedural death labyrinth definition) map generation is generally fiddly to make things that play well and have the desired character. Nethack maps are distinct styles from Crawl or *band maps. And within a game, different areas will usually have different styles too. Those kinds of maps are also simple enough that a neural net algorithm wouldn't get you anything new You don't really need "AI" to make trees and grass, though. I had a professor whose field of study was making good-looking, procedurally generated plants, and his work was plenty good. It could also run on an iPhone in the year of our lord 2011, so I'm pretty sure it wasn't using modern AI techniques. The question might now become: how do you place the trees? Well, that's quite a different question. Do you want a natural forest? You could probably generate one algorithmically given a reasonably small amount of cross-disciplinary research. Would a natural, lifelike forest be good for gameplay? That's a much more important question if you're talking about generating a game without human intervention. Could you use AI techniques to prune the list of "possible forests" to create one that's good for gameplay without doing much work? There's the real question. I mean, you probably could, but you have to know what you're trying to achieve in the first place, and it depends on context. At some point, you probably want a skilled hand directing your tools.
|
|
#
?
Mar 8, 2023 01:31
|
|
|
I saw a cool video on waveform collapse generation Basically using sudoku as a map gen. Neat stuff, no AI needed. I might try it if I ever make a thing that needs random terrain
|
|
#
?
Mar 8, 2023 01:50
|
|
|
Anticheese posted:I saw a cool video on waveform collapse generation That's the thing I actually find quite limiting about current AI research. Take my professor's research: the modern version would be "here are some pictures/models/etc. of a tree, please make me a tree." It works, but it's probably not very efficient and it probably doesn't generate a great result. But his research was focused on "okay, what is the fundamental structure of a tree, what are the variations within a tree within a species, and how can we use that to make a tree that looks very tree-y?" In earlier discussions, we talked about music, and that's another area where this sort of AI is probably not an optimal technique. We have poo poo-tons of data about music, you could use traditional training algorithms to determine what good, new music would sound like, but we also have a huge body of theory about notes, melodies, chord progressions, rhythms, etc. that mean we don't have to train the model on a bunch of sheet music or whatever in order to get a reasonable output.
|
|
#
?
Mar 8, 2023 02:04
|
|
|
Electric Wrigglies posted:What are you talking about, People played the poo poo out of DND and other simple games that were like one or two people teams that used procedural methods to generate content. have you, like. played d&d? with a dm? who uses their actual human brain to author a game world for the other players??
|
|
#
?
Mar 8, 2023 02:40
|
|
|
EightFlyingCars posted:have you, like. played d&d? with a dm? who uses their actual human brain to author a game world for the other players?? Look, some people just want to use the random dungeon generator in the back of the first edition Ad&d Dungeon master's guide and commit some casual genocide for a few hours. No shame in that.
|
|
#
?
Mar 8, 2023 02:52
|
|
|
Dwarf Fortress has demonstrated it's entirely possible to make a game that's 100% procedurally generated, but each step of that procgen was specifically tailored to its task. There's no need nor benefit to adding AI to it. Maybe for character portraits and similar assets, but the cost/benefit is hilariously bad.
|
|
#
?
Mar 8, 2023 03:53
|
|
|
EightFlyingCars posted:have you, like. played d&d? with a dm? who uses their actual human brain to author a game world for the other players?? I think they mean dnd, the very early adaptation of D&D as a game to be played on a mainframe console that eventually lead to rogue and roguelikes.
|
|
#
?
Mar 8, 2023 04:00
|
|
|
I'm pretty sure for the amount of work you have to do to make AI-generated content actually usable and fun, you're literally better off just using that effort to start from scratch yourself. I am reminded that one recent high profile game apparently used algorithmically generated content to fill out content and save development time. That game was Balan Wonderworld. It was not well received.
|
|
#
?
Mar 8, 2023 04:31
|
|
|
Based on what I've seen so far AI generated text is best to fill gaps in existing human created works, rather than coming up with stories whole cloth. For games It could be given an outline and snippets of human written script, mixing it with relevant game variables (disposition etc) to piece together something more natural than text generated procedurally the old fashioned way. Voice synthesis for things like getting directions or shopkeepers with traditional pre-recorded voices for more important lines like for quests. That kind of thing.
SCheeseman fucked around with this message at 04:46 on Mar 8, 2023 |
|
#
?
Mar 8, 2023 04:41
|
|
|
Large-language model-type "AI" is not only applicable to tabletop roleplaying games, it's going to be enormously influential. In some pretty meaningful ways it has existed since 2007, and a version of it is at the cutting edge of TTRPG design. Back in '07, the Mythic Game Master Emulator was released by Tom Pigeon of Word Mill Publishing. It's what's known as a "system-agnostic" game supplement: a standalone book containing rules that will work with any game system, not just Dungeons & Dragons. Its purpose is what it says on the tin: it allows you to play tabletop RPGs, most of which are meant to have one or more player-characters and one or more game masters, solo, by essentially acting as the game master. I won't get too into the weeds on how it does this, but the short version is that any time you'd normally interact with the game master rather than the rules of the game, you consult the emulated game master in the form of - tell me if this sounds familiar - weighted randomized tables. The more detailed a response you need, the more you can drill down through its responses by asking whether X or Y is true, and assigning weights to each side. The emulator is also salted with outcomes which break you out of the track you expect to follow and send you down another trail (and another set of tables), such that you're never in complete creative control. It's a relatively crude system, but it works well enough to generate some legitimately compelling results, and it has remained in print for over 15 years. You may be seeing where this is going by now. (Note: this is all based on the 2007 version of Mythic GME. A modernized version came out late last year, but I haven't used it so I won't comment on it.) Now, all of this is ancient tech, since 2007 is an eternity in the timeline of TTRPG development. I suspect goons are more aware of this than most audiences, but it bears mentioning nonetheless: there are many, many tabletop RPGs other than D&D and Pathfinder, and the art form is currently in a golden age of design. Mythic GME hasn't been terribly influential on game design, since it's an overlay on top of other rules sets, but that has changed with the publication of Ironsworn and its sequel system Starforged, written by Shawn Tomkin and published in 2018 and 2022 respectively. Ironsworn's premise is that it's a system designed for solo or GM-less co-op play, which it accomplishes brilliantly by integrating something along the lines of the Mythic GME into its core mechanics. MGME's greatest strength and greatest weakness has always been its system-agnosticism: it works with anything, but it doesn't work beautifully or elegantly with anything. Ironsworn squared the circle by building its GM emulator into the system and elaborating on the techniques pioneered by Mythic. It retains the basic function of "ask questions of weighted probability tables until you get sufficient clarity to engage with the rules on your character sheet," then goes a step further with oracles. Oracles are random tables, usually with entries from 1 to 100, that are used as themed creative prompts that you engage with either when the rules of the game tell you to do so, or when your appraisal of fiction tells you they're necessary or interesting. This is all a very dry, surface-level description of it, but anyone who has played one of these games will tell you that the results are genuinely kind of spooky. It often feels as though you're playing with another person, but one who you can only communicate with through a Ouija board. If you've played around with ChatGPT enough to see it perform some of the spookier tricks in its repertoire, this will all feel familiar. It even works in a similar way: ChatGPT is just ("just") picking out the next word in the sequence via a hellaciously complex set of weights salted with randomness to make it feel more human, while Ironsworn and its descendants are doing something very similar, but at a much slower pace, and with prompts that are tailored to generate outcomes appropriate to their genre. And it works. Mythic GME functioned well enough to cement a small but persistent place in the hobby for 15 years and counting; Ironsworn works so well that it has become A Big Deal and is inspiring tons of new work, including the current topic of discussion in TradGames general chat, Across A Thousand Dead Worlds. There is absolutely no doubt in my mind that you could build an RPG whose mechanics linked into ChatGPT right now, and have it be a genuinely compelling GMless experience. It wouldn't be easy, and the first people to do it will fail, because TTRPGs have only a handful of designers who are capable of more than cargo cult design at best. But sooner than you think, we will see games that use this technology to good effect, and as soon as the public can get a hold of LLMs as advanced as ChatGPT, in a way that lets us tailor them to a specific purpose, GMless RPGs are going to absolutely explode. As an aside, if anyone's curious about how Ironsworn plays out for real, I ran an entire Ironsworn adventure a couple years back in The Game Room, which will illustrate how this works much better than I can explain it here. Ironsworn itself is also free, so you can check it out for yourself and see what I mean about the spookiness of oracles.
|
|
#
?
Mar 8, 2023 05:06
|
|
|
Just hinged your whole Ironsworn thread, that was dope and I want to read more (but I fear I don't have the creativity to play it myself)
|
|
#
?
Mar 8, 2023 18:36
|
|
|
NomNomNom posted:Just hinged your whole Ironsworn thread, that was dope and I want to read more (but I fear I don't have the creativity to play it myself) You definitely do! That's the remarkable thing about the "technology" that Ironsworn et al are based on: it's like a cybernetic augment for creativity. And, to bring it back to the thread topic, that's why the intersection of TTRPG design and LLMs has so much potential. You don't need to replace the other people at the table with replicants, you need to have a structure that fires your creativity and, crucially, challenges your assumptions and gives you surprises within the bounds of genre and setting. Modern solo RPGs already have a remarkable capacity for this, but it's very much in the era of stone tablets and abaci compared to where it can go as game design and technology advance. It may seem like I'm overhyping this, but anyone with the least curiosity about it ought to spend an afternoon with an oracle-based system (reminder that Ironsworn is free!), and an evening mulling over what skillful assistance from an LLM could do in that context.
|
|
#
?
Mar 8, 2023 21:46
|
|
|
Kyte posted:Dwarf Fortress has demonstrated it's entirely possible to make a game that's 100% procedurally generated, but each step of that procgen was specifically tailored to its task. If we went with the original full name of the game (Slaves to Armok: God of Blood Chapter II: Dwarf Fortress) and it had an AI Armok having a blast at your expense, you bet people would play the poo poo out of it. Edit: of course there wouldn't be resources to run it because those would be taken up simulating blades of grass. Rocko Bonaparte fucked around with this message at 02:03 on Mar 9, 2023 |
|
#
?
Mar 9, 2023 02:01
|
|
|
this sounds dope
|
|
#
?
Mar 9, 2023 03:34
|
|
|
Kestral posted:It may seem like I'm overhyping this, but anyone with the least curiosity about it ought to spend an afternoon with an oracle-based system (reminder that Ironsworn is free!), and an evening mulling over what skillful assistance from an LLM could do in that context. Well, I would also say you should mull over what a LLM could or would do without a custom-designed system to augment it.
|
|
#
?
Mar 9, 2023 04:07
|
|
|
Kestral posted:It may seem like I'm overhyping this, but anyone with the least curiosity about it ought to spend an afternoon with an oracle-based system (reminder that Ironsworn is free!), and an evening mulling over what skillful assistance from an LLM could do in that context. okay but i don't know how a master of laws helps with playing ironsworn
|
|
#
?
Mar 9, 2023 04:13
|
|
|
In general I think the idea of being able to fire all the designers and press a button that makes large chunks of the game for you is a pipe dream, but I wouldn't be surprised if down the track algorithmic generation can be a useful tool. But that's the thing, just a tool like any other, like the procedural generation we've had for decades by now, and it's still going to need a lot of design work at every stage to make it actually fun and immersive. Been said before that for all the hype of 'AI' stuff, as soon as it actually becomes useful it immediately stops being called 'AI'.
|
|
#
?
Mar 9, 2023 05:12
|
|
|
Ghost Leviathan posted:In general I think the idea of being able to fire all the designers and press a button that makes large chunks of the game for you is a pipe dream, but I wouldn't be surprised if down the track algorithmic generation can be a useful tool. But that's the thing, just a tool like any other, like the procedural generation we've had for decades by now, and it's still going to need a lot of design work at every stage to make it actually fun and immersive. This is dead on, IMO. Anyone expecting these tools to replace game designers or become a full-fledged GM for them in the near future is going to be sorely disappointed. As a tool for structured creativity though, they have enormous potential. I’m excited to see what happens when future models get into the hands of the few really gifted game designers bobbing around in the indie tabletop scene.
|
|
#
?
Mar 9, 2023 23:56
|
|
|
Personally, what I really want AI to be used in games is for ... AI I mostly play strategy games and always single player, it would be really cool to have AIs that can actually play the games, instead of being dumb bricks with huge advantages
|
|
#
?
Mar 10, 2023 12:43
|
|
|
drat too bad we threw out rule-based system research in favor of fancy text completion then
|
|
#
?
Mar 10, 2023 12:50
|
|
|
Elias_Maluco posted:Personally, what I really want AI to be used in games is for ... AI I think the problem might be making a model with scalable difficulty. Probably wouldn't be too fun if the model was trained on pro gamer stats and could perfectly min max.
|
|
#
?
Mar 10, 2023 14:02
|
|
|
NomNomNom posted:I think the problem might be making a model with scalable difficulty. Probably wouldn't be too fun if the model was trained on pro gamer stats and could perfectly min max. That's just a matter of choosing the right exploitation-exploration trade-off. This is a routine problem in recommender systems already, where you somehow want to force these systems to discover and recommend new content in addition to the presently best performing content, to identify and establish new gems. One of the simplest ways to fix that is to make the system do random dumb poo poo X% of the time, so you could have difficulty levels like easy (75% random poo poo), normal (50% random poo poo), and hard (25% random poo poo). This could potentially be very funny on gameplay level where your AI plays like a complete moron on strategy level on easy, but then it “clicks” into doing an optimized play and tries to Faker you with a train wreck of an army or whatever.
|
|
#
?
Mar 10, 2023 14:19
|
|
|
Didn't we literally have a computer teach itself to play Starcraft at some point? The tech is there, it works, it just takes investment in letting a computer play against itself for a million billion hours while it figures out all the rules. Scaling geometrically with complexity, which I assume means it would take forever to learn Civilization. Edit: Trying to use any kind of learning model analogous to ChatGPT where it just studies human games and tries to copy them seems like a good way to make a very dumb AI that does the strategic equivalent of drawing too many fingers on every hand. Like, it'd set up a perfectly timed attack on a place with no city because it doesn't know what a city is, it just knows that people usually attack places like this. Clarste fucked around with this message at 14:44 on Mar 10, 2023 |
|
#
?
Mar 10, 2023 14:33
|
|
|
That's kinda what chatGPT does when you play chess against it. It cares little for the state of the board and just decides as you have played e4 e5 that statistically the best response is e3 e6 or whatever, regardless of even if there is one of its pieces on e6 already or any other rules preventing it. Its a very good way to demonstrate the limitation of LLM types of AIs. https://youtu.be/rSCNW1OCk_M Mega Comrade fucked around with this message at 15:39 on Mar 10, 2023 |
|
#
?
Mar 10, 2023 15:36
|
|
|
Clarste posted:Didn't we literally have a computer teach itself to play Starcraft at some point? The tech is there, it works, it just takes investment in letting a computer play against itself for a million billion hours while it figures out all the rules. Scaling geometrically with complexity, which I assume means it would take forever to learn Civilization. Is Civilization more complex than Starcraft?
|
|
#
?
Mar 10, 2023 15:41
|
|
|
Clarste posted:Didn't we literally have a computer teach itself to play Starcraft at some point? The tech is there, it works, it just takes investment in letting a computer play against itself for a million billion hours while it figures out all the rules. Scaling geometrically with complexity, which I assume means it would take forever to learn Civilization. the issue here is that a computer game AI isn't designed to play the game, it's designed to convincingly lose in a pleasing way that makes you feel very accomplished and smart. if the computer beats the pants off you, you get pissed off and stop playing. so the trick isn't getting the computer to play it very well, it's getting it to play in a way that is perceived as well, but lets the player win. but it's very hard to tune that carefully for the long term so the computer keeps presenting a convincingly pleasing, but beatable, challenge as you scale in skill. especially since the developers may not have a firm idea of how high the skill levels can get.
|
|
#
?
Mar 10, 2023 15:51
|
|
|
|
| # ? Jun 8, 2024 06:19 |
|
|
Clarste posted:Didn't we literally have a computer teach itself to play Starcraft at some point? The tech is there, it works, it just takes investment in letting a computer play against itself for a million billion hours while it figures out all the rules. Scaling geometrically with complexity, which I assume means it would take forever to learn Civilization. Only with full visibility of the map, iirc. Getting current AI models to work on limited information games has been very tricky so far iirc. There was one for Poker recently, but iirc it brute-forced all possible states and just played the odds better than any human can.
|
|
#
?
Mar 10, 2023 15:52
|
|
