Or you could be me, and have all the skillsets used across the whole process, but consistently get too burnt-out to finish anything. 
Regarding placeholder images, though, one technique I know about is using public domain images for those, and then commissioning an artist to replace them later.
Probably also worth slapping faint text reading “placeholder” across these images, too, just to ensure the point is made.
Also I feel like this is another case of misunderstanding what players are looking for.
If the text changes, then I assume something happened, and I’m going to re-read it.
I am not re-reading the entire room description or passage text when returning to some previous point of the game. I’m looking for what changed.
Having a wide number of variations for description paragraphs is actually annoying for gameplay reasons. I only randomize (hand-crafted) text when it’s describing something actively happening on every turn, and it gets its own line.
Yeah, definitely – the Inform/TADS default responses are actually sometimes a gift to players, telling them “nothing to see here, time to move on!” Of course, you need to have solid implementation to make sure you’re not accidentally telling that to the player when there actually is something important going on…
Keep in mind, modern video games and IF have very different standards. Conservation of detail where every bit of custom text is narratively or mechanically important or a deliberate red herring might be the norm for IF, but pointless flavor text and window dressing that serves no purpose beyond making the game world feel more like a real place is kind of expected with mainstream games and building a forest by copying the same, exact tree a thousand times doesn’t go over nearly as well as it did in the days when technical limits meant that was really the only way to do it.
I’d like to counter with how mainstream game studios are starting to understand that they’ve shot themselves in the foot a bit by trying to make stuff as visually detailed and complex as possible, because the important game elements get lost in all of it, and level designers have to actively fight back against this by reducing detail and make things more clear, so the player can identify the key elements around them.
In MechWarrior: Online, for example, players usually opt for visual filters that hide the vast majority of the graphical detail, because everything looks like a mess of environmental noise without them active.
I feel like this fits the IF pattern I described pretty well.
This isn’t the place for my rant about how the reason mainstream studios are pushing photorealistic graphics as the only measure of “quality” is because photorealism is something that can only be achieved with tons and tons of person-hours and thus something that the big studios can achieve by crunching their artists and devs and indie studios fundamentally cannot.
But just know it’s there. Lurking. In the background.
At the risk of derailing the thread (although I notice it’s already crept back to how useful LLMs might be for authors, perhaps proving its thesis), this reminds me of a more general aphorism that I’ve abided by for years: you are the one person in the world who is the most passionate about the game you are making for free.
If you are bored writing a large number of object descriptions, how bored is the player going to be reading them? You should either remove those objects or make them interesting somehow.
To re-rail things a bit, the new temptation is to get an LLM to write the descriptions, which solves the problem for the author, but not the player. And unfortunately, that’s exactly the wrong way round to solve anything.
Garden was a conversion of an unillustrated paper entry to the Lindenbaum contest for short CYOA books (Lindenbaum doesn’t allow illustrations), which I then converted to an IF, which explains my approach to the art. It was meant to be a small project, and I was new to Twine, so it never really occurred to me that it couldn’t have art, and Stable Diffusion was, at that point, brand new and hot — the statue is a typical SD1.5 sort of mess.
I’m not sure how I would approach it now. I still doubt I’d be up for sketching all the garden areas, so I’d be more likely to release it without background art now, which would be a shame.
Fighting the good fight, Brian.
For whatever it’s worth, I wouldn’t touch so-called “AIs” (more accurately called LLMs; more truthfully ‘automated plagiarism machines’) with a ten meter pole. Nothing puts me off a game, story, or other product faster than the AI label.
I totally take your point here. My thought experiment of slightly-different offices was hasty to try to make the point, but not advocate for that point!
More generally, I’m thinking about the usefulness of having a ‘2nd unit author’ as Hanon described, but only when the particulars of what that tool can do would actually add to the gameplay experience.
It’s not a tool for every game, or even most games. But maybe if I’m doing a text adventure version of Borges ‘the Library of Babel’ and a nearly-infinite slightly-varied series of rooms is actually directly contributing to the aesthetic experience, then maybe it makes sense.
I love your aphorism, and I think it’s an absolutely spot-on inverse-proof of mine!
There’s also a kind of inverse time-spent correlation here. I’ve invested a LOT of time on my current project, that I’m hand crafting. I think that can be sensed by the players, the multitude of mini-decisions about every element. But if an enormous boatload of text that took a few seconds to generate scrolls past, our mind starts to feel the static of vagueness.
To also join your re-rail, I’m not actually advocating letting an LLM write the descriptions whole cloth. The thing I’m trying to figure out (and I don’t have an answer) is, what kind of task or project would benefit from these processes?
Not directed at me, but I’d like to add that even when I was doing Garden, people were already worrying about the plagarism aspect of Stable Diffusion. For that reason I picked an art style using 18th and 19th Century woodcut prints and etchings, which are in the public domain in the first place.
Not saying that absolves OpenAI of their training practices, but at least I could feel fairly confidant that the images I was using weren’t going to be infringing someone’s copyright or stealing some living artist’s style.
Make of that what you will.
This is more or less like saying you ate only the vegan parts of a dessert. The flour and the milk are irreversibly bound together — the images in the training set can’t be considered separately as the copyrighted images were used to train the image recognition components just as much as the free images were.
That vampire game where you attempt to convince people into letting you enter their house was built with AI and was pretty fun from what I saw.
//edit: I think the future of LLMs in gaming is in powering the NPCs so that you can have actual conversations that seems to be realistic with them. I’m working on something to that extent in Godot (which has support for local LLMs via addons), however I still need to figure out the best way to make NPCs act realistically so they don’t do one thing, but say they’re doing something completely different (unless that’s the point with that specific character).
Until the player leads a question and the NPC starts making stuff up that could plausibly be in the game world, and send the player on a wild goose chase.
I think what would be necessary to use AI directly in a parser game - as you said for NPCs would be very useful - is if the AI understood it was performing with a specifically limited data set and you could direct the AI in a kind of “rehearsal” mode where you instruct it how to behave the same way you’d direct actors or improv performers or players in a tabletop RPG.
This is blue-sky, but:
>ASK BOB ABOUT WALLETBob looks up from raking leaves. “Oh, hello! My wallet is over there under the shrubbery.”
>DIRECT BOBOkay, please give direction for Bob.
>Bob doesn't know where the wallet is. Do not tell the player that the wallet is under the shrub. Bob doesn't know where the wallet was lost until the player shows it to them or tells them about it. If the player asks when you last saw it, Bob can say he knew he had it when he was trimming the hedges. If the player tells Bob the wallet is under the shrub, then Bob should try to look under the shrub and then take the wallet, then thank the player. If the player gives the wallet to Bob, he should thank the player enthusiastically and offer them the five dollar bill inside. If the wallet does not contain the five dollar bill, Bob should be sad that it was stolen and act more suspicious when talking to the player if the player has the five dollar bill in their inventory.Okay! Bob takes the note.
>STOP DIRECTING BOB[Normal play resumed]
So? Nothing wrong with the AI trolling players a little bit. Plus, as a developer I can just add the thing AI was talking about if it’s something cool to mess with the players further.
Oh, certainly finetuning NPC models is an important step here.
Considering the size of the data sets any halfway decent LLM is trained on, I have to question wether it’s even possible, without referencing a specific author/work from the data set in the prompt, for there to even be enough similarity between anything in the data set and any particular output to count as plagerism by any reasonable standard.
Makes me wonder what the results would be if you say, trained a model on the corpus of Project Gutenberg, generated a million pieces of AI fiction with said model, and then fed the output to a plagerism detector using Project Gutenberg as its database of prior art… Or trained a image generator on the surviving works of antiquity, the middle ages, and the Renaissance, had it generate a million images, and then fed both image sets to an image de duplicator.
When you’re thinking of all those cases and responses you may as well just write out answers yourself.