> Suppose we want a new version of Pokémon Yellow with a new Pokémon creature. We don’t need to change the game’s source code, we just need to edit the wiki and then press the “remake game” button. Modding a game becomes as simple as editing the wiki. Creating a game becomes as simple as writing its wiki.
Is this actually at all feasible? The total amount of storage space used by the Pokémon wiki is hundreds of times larger than the original game, likely requiring many times more effort to write than spent by the original programmers, and it still doesn't cover many, many aspects of that game (e.g. that Pokémon is tile-based; what buttons do in various contexts; sound and visual UI design) -- to say nothing of the clever ASM and memory tricks necessary to make the game work on the Game Boy.
I'm reminded of that comic where the punchline is "do you know the industry term for a specification precise enough to generate a program? Code. It's called code."
> it still doesn't cover many, many aspects of that game
But I kinda love that as an idea. The text of the wiki defines the core ideas of what the game is, but a lot is left unspecified and to be interpreted.
Pokémon legends Arceus feels a lot like what someone who had only read about the franchise may create as a game, and it's a pretty awesome addition to the series.
I agree in principle, but that's dangerously close to saying "I'd love if AI was a good game designer" -- I mean, I'd love it too, but we're very far from that, and there are even more interesting things we could do in that world.
Also, if I were instructing someone to make a Pokémon game, the info on the wiki (gigantic percentage and stats tables, tilemaps, EVs and IVs) seems like generally the exact opposite of what I'd choose to specify.
Funny you say that, because I'd actually take stat distributions and movesets and all that other data-heavy stuff to be one of the most important factors in the game!
Can you code a million games per day?
The barrier to entry if this works out, would allow anyone to make such
games by simply writing down a few pages of "mods" every hour.
Or if you force ChatGPT to creatively generate some changes, it could
be automated. Hundreds of games per hour? Local models trained on some
custom dataset could generate these wikis in batches.
Thousands of new games per hour! And if the generator runs on random
seeds as input, why not try them all? Millions of games per day.
Obviously a single person with this generator would be able to saturate
entire genres in space of weeks.
Is there a market for thousands of custom games per hour? Games designed without clear design goals or quality control? Games where the player is the alpha tester for mechanics generated with no guarantee of consistency or even functionality?
It's like our modern mobile game landscape, but worse. I find such vision somewhat revolting, tbh.
Exactly. I'd argue there's still quite a gap between the PC/console and mobile markets. Mobile games aren't bad, per say, they just feel like they're optimized for revenue extraction vs entertainment. As an example, look at the expectations around big franchises when they launch mobile titles, e.g. Diablo: Immortal or C&C: Legions.
I see your point though. Steam's quality has been getting less and less consistent, particularly when it comes to early access titles. Maybe a better example would have been NSFW Steam games: a monotonous landscape devoid of originality sold for a few bucks a pop.
I think the idea isn't that it generates Pokemon Yellow exactly, just that it generates a game of similar complexity. If it's not in the wiki it's not relevant. If its relevant then it should be added to the wiki.
The original Pokemon games (Red, Blue, Yellow, and Green in JPN) were also full of notorious programming idiosyncrasies and glitches. I'm not sure how safely you could insert new data into the game without poking the bear too much.
Wow, I also spent years reading those guides before playing by first ever Pokémon game. I even started to write my own guide. As a child, it was very fun.
Now I'm wondering how many people share the same experience.
My friend who I played Red and Blue with had one of those.
They were an incredible resource. Later I was almost more amazed by gameFAQs which has very detailed guides for almost any game. And at least back then it was all ASCII with no ads. A million times better than the blogspam game guides that Google returns these days
gamefaqs often still has really really good mostly text guides for PC cRPGs (with the odd exception of BG3). Maybe all the old-timey text nerds only like RTWP (which is a position I have a lot of sympathy with).
On a cursory survey of GameFaqs as it exists now, checked the top five most recent game additions. Got semi-recent to recent games. Seems very similar to 20 years ago with slight tech additions. Still very slimmed down. Minimal load.
Fate/Extra CCC, All Endings (1/2/24); Pokemon Mystery Dungeon: Explorers of Sky, Recruitment (12/28/23); Shin Megami Tensei V, Walkthrough (1/4/24); Super Mario Bros. Wonder, Strategy (12/25/23); The Sims 4, Rent Object List (12/30/23)
Oldschool Ascii: 1, Pokemon Mystery Dungeon
Text HTML w/ Menu: 2, Fate/Extra CCC, Shin Megami Tensei V
Text and Images: 1, Sims 4
Text, Images, Video: 1, Super Mario Bros. Wonder
They all seemed reasonably well written on a cursory glance. Notably, did not delve all that far into checking whether they were factual.
I spent weeks and months taking a dozen different guides online in .txt format and combining and editing them together and doing my own editorializing. I carried this around on a 3.5 inch floppy disk and anytime my parents dragged me somewhere boring that had a floppy drive, I'd sit on whichever computer I found editing my guides while they did boring adult stuff. Usually at a family member's or friend of the family's house.
I had a guide for Wind Waker that I’d enjoy reading and mostly appreciating the artwork from (I’m not sure I even could read English when I first bought the guide), yet I didn’t have the game and only played occasionally at friend’s places.
I remember planning to collect all the statues in that game. I never did, and likely never will, but I’ve enjoyed Wind Waker more recently and it really is a great game and I would like to play through it at some point.
But I feel like that this isn't really the same experience. The Pokémon guide feels closer to the game than the Zelda one, especially since the Zelda games are known for their riddles - which are very hard to properly transcribe in a guide, imo.
Some of Zelda's best riddles can be rendered useless as soon as you read the solution (thinking about a very good one in Phantom Hourglass I'll never be able to experience again) - but the Pokémon ones still feel fresh, especially since the games rely so much on exploration, grinding and battling.
I hadn't thought of it like that, but I agree with you. The magic of good puzzle design lies partly in figuring it out yourself and feeling clever, and you miss out on that when reading the solution to the puzzle in a guide.
I absolutely loved what Phantom Hourglass and Spirit Tracks did with the hardware they had available to them in terms of game design. I'd love to see a remaster of Spirit Tracks, but I'm not sure it'd be possible to capture the cleverness of their controls (admittedly not always a good thing) and puzzles on the Switch?
The Pokemon Power magazine had a very small guide for Red/Blue at the start, I must have read that 1000 times over. Just before christmas that year I was yahooing some Pokemon / Gameboy things when I stumbled upon the no$gmb and the world of emulation... I have fond memories of that emulator and the debugger.
I was the same with a GameFaqs walkthough of Pokémon Red/Blue. It was written in a very interesting style too and like many people I wish I could play the Pokémon game that it helped to create in my head.
A bit younger, so the game isn’t the same, but I had the same experience with Pokémon Mystery Dungeon - I did ultimately get the game within a year or so
> Changing a game becomes as simple as modifying a wiki
I feel like an implementation of Pokemon Yellow in a reasonably modern high-level programming language would be easier to reason about and modify than a sprawling wiki written in English.
Yes, I got the sense from reading this that the author has neither spent a lot of time contributing to a wiki, nor writing software, because both tasks are scarily complex in a way that the blog entry skates over. But then I clicked "About" and I see they are in fact a staff engineer at reddit, so shows what I know.
The hubris of solving social problems with technical solutions is a pipe dream many corporate engineers fall into. I think that's because real social problems, i.e. conflicts, are handled by their managers.
While I've definitely seen the trap you're describing, this isn't it. This is less "solving a social problem with technology" and more "anyone could code if only they didn't need to learn those pesky programming languages". Which is odd to see coming from an engineer who should understand that a programming language is just a precise specification of behavior. I guess the main thought in the blog post is that you don't need to actually specify your problem domain because an LLM can infer the details from context and its training data? Which may be true (at some point) but still doesn't make English a good programming language.
Isn't this just data oriented design? The guide contains all the data for the game (maps, most graphics data as they were sprites, npc locations, stat blocks, etc). There have been RPG makers for decades that allow you to create a game with an OOTB world and combat engine. I guess a wiki is a novel interface, but it seems like a No Code game generator, which comes with a whole host of limits and issues.
This seems to come from a place where you think knowledge can conquer the world. Which it can, only when combined with action. Want to make a retro game that is reminiscent of your childhood? Awesome go do it. There are a thousand tools out there. But you are going to be shaving a lot of yaks if you think you can create something by solely knowing everything about it. Might as well build software with UML.
I've been pondering this recently; I think there are some really fun applications of LLMs to game dev that have not yet been explored.
Going back to the original AI Dungeon (text-based adventure game, fully-generated by LLM), it was a lot of fun, but hallucinations were crazy. You could just ad-lib a change to the world-state and it would run with it.
One possibility would be to have an LLM act as a DM and use Tool APIs to drive a world model; for example, the DM can call `moveMonster(newPosition)` a la the Voyager paper. The game engine enforces constraints, so you can't just conjure items out of thin air. The player isn't interfacing with the LLM directly, so they don't see the hallucination problem.
It feels like lots of experiments have already been done with pulling NPC dialog from LLMs, but what about crafting worlds (most easily in text-based games, but potentially graphical too) and then reifiying them into the engine? Perhaps this could make more cohesive procedural worlds, since you can have the stories in the world's history drive the actual layout of the areas, scenes, and characters.
I did the same thing, used to fall asleep reading the strategy guide for at least a year before I got the game myself.
I was similarly absorbed with those annual “bumper collection of cheats for video games” books that used to come with tech magazines, endlessly rereading them and memorizing them for games I never owned.
I wonder:
A) How much this is endemic to the Hacker News demographic. I’ve always been absorbed in the pursuit of useless knowledge (later in life this manifests in unnecessary PC watercooling and Haskell addiction).
B) If we all had genius parents that realized you can spend £3.50 (much less than the cost of an A list game) and keep a child occupied for an entire year, while also encouraging them to read (!!).
Children are learning machines, they will devour anything you give them... whether it's "useful" or "useless" is an adult judgement, or perhaps even a judgement on your deathbed?
There are plenty of quiz tournaments and quiz shows, and whole communities of general-knowledge recallers called "quizzers", ultimately because _other_ humans are impressed and entertained by displays of mental prowess. Is knowing every Pokemon or every x86 instruction any more or less useful than knowing test cricket results or every Beatles recording?
This reminded be of 2kliksphilips's "The Best Game I Never Played", which is about the experience of being obsessed with a video game you can't access as a child so much, that you can never actually play it without being let down.
Interesting idea! I’m curious if one could create an app (purely as an experiment) where source control contains a repository of markdown docs and the build step just calls an LLM in a specific order.
I think before trying to build a new game from scratch using a wiki, they should probably prove the concept by building Pokemon Yellow from the game guide.
Is this actually at all feasible? The total amount of storage space used by the Pokémon wiki is hundreds of times larger than the original game, likely requiring many times more effort to write than spent by the original programmers, and it still doesn't cover many, many aspects of that game (e.g. that Pokémon is tile-based; what buttons do in various contexts; sound and visual UI design) -- to say nothing of the clever ASM and memory tricks necessary to make the game work on the Game Boy.
I'm reminded of that comic where the punchline is "do you know the industry term for a specification precise enough to generate a program? Code. It's called code."