Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

I'm not convinced that current generative AI is a good fit for this kind of game. IMO, the heart of the text adventure game is the world model, and LLMs are notably lacking here. It's hard to believe the game is simulating a real place when it doesn't even have object permanence.

That said, my favorite human-authored text adventure game (I prefer that name to "interactive fiction" because I'm primarily looking for entertainment, not literary value) is Lost Pig:

http://grunk.org/lostpig/

Playable online with a Javascript-based interpreter at:

https://iplayif.com/?story=http%3A//mirror.ifarchive.org/if-...

It's a comedy, and just as with graphical adventure games, I think the whole adventure game concept works best with comedy. Even human-authored world models are inevitably flawed, and the resulting absurdity best matches the tone of comedy. I also recommend another comedy, Brain Guzzlers From Beyond!:

https://ifdb.org/viewgame?id=f55km4uutt2cqwwz

Both these are relatively modern games, written after the commercial collapse of the genre. They were the winners of the 2007 and 2015 Annual Interactive Fiction Competitions respectively. More information about this:

https://www.ifwiki.org/The_Annual_IF_Competition



I'm applying generative AI to interactive fiction with Spellbound

https://www.tryspellbound.com/app/scenario/65821/create

(You can click on the dice at the bottom to turn on D&D mode)

I've taken the approach of starting with the #1 problem with Gen AI for this application: that it writes bland prose with not much going on by default.

From there you can layer on systems that address things like object permanence, but even with a basic engine capable of generating fun to read pages of text I think you already get a pretty fun experience


> I'm not convinced that current generative AI is a good fit for this kind of game. IMO, the heart of the text adventure game is the world model, and LLMs are notably lacking here. It's hard to believe the game is simulating a real place when it doesn't even have object permanence.

The mechanics still need to be coded; however, like the OP, I believe there's an enormous opportunity to enrich the game content with LLMs. Like procedurally generated maps in roguelikes, LLMs can be used to create an order of magnitude more unique interactions in the game than what you could provide by crafting each dialogue tree by hand. While not as good or memorable (on average) as a hand-written character, it should be more than enough for Villager A, who normally would be completely mute.


Perhaps the game could be developed collaboratively online, so interested contributors could add dialog for the villagers and other characters, instead of trying to just outsource to hallucinating LLMs.


yep, that's what I was thinking. A lot of filler can be written using llm.


There’s a lot of IF like Eat Me or Midnight Swordfight which really lean on excellent writing, black comedy and wordplay. LLMs are still largely terrible at humour. Anything by Chandler Groover is recommended:

https://ifdb.org/search?searchfor=author%3AChandler+Groover

Someone else mentioned Emily Short and there are also fantastically written games like Sunless Seas/Skies.


Agreed about the importance of a world model. People get enamored with increasing text volume with LLMs, but unless it ties back into distinct game state interactions, players quickly recognize it as fluff and it ends up not really adding to the game.

It's not exactly a text-based adventure game (more text-based trading), but recently I spent some time messing around with integrating an LLM with a text-based world model. It's not 100% reliable, but I've had some pretty satisfying interactions with it: https://github.com/heyitsguay/trader

Basically, you're a trader traveling around a world buying and selling goods, but the economy is tuned such that you can only really get ahead by conning the NPCs >:D




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: