And, I just released my own attempt to merge CYOA, IF, and AI Dungeon with an optionally AI-powered, editable, persistent, shareable story-generation system, https://storyweaver.online
This was great. I grew up playing Zork and the graphics-text integrated King's Quest and Space Quest Games, which I paid homage to in my story, Mage Quest. My favorite of the book-based adventures was Christopher Manson's, Maze, which still holds a certain fascination for me. While it was easy enough to (eventually) work your way to the last page and back out, the best part of the book is its eerie and provocative artwork. Each page has a cryptic, gnostic quality that only hints at the right direction to follow and, perhaps, a bigger story arc. But Mason never gave away too much. As far as I know, no one has ever fully interpreted all of his clues hidden throughout the book in text and image. Mason was probably a little schizo.
Modern LLMs are uncreative, true. But I wonder how much of that comes from instruction-tuning. Base models are quite capable of going off on weird tangents, something post-training tries hard to get rid of.
Perhaps the reason AI Dungeon was a flash in the pan was that GPT-2 was a base model, and could be pseudo-creative in a way that more advanced, but instruction-trained, models can't. The other issue, the lack of a world-model, still applies in full force.
I remember, back in the day, trying to get GPT-3 to write fiction, using the then-available davinci base model. The result was terrible, of course, not good fiction. But it wasn't the kind of mediocre you get with modern AI writing. It wasn't trying to succeed at the task of writing fiction in order to please the user and get a high grade, but rather to simply predict the next token. The result was something like a plotless—and very much world-model-lacking—improv act, where it was always trying to yes-and the text generated so far, sometimes in ways that were amusing to the human reader in the way typically associated with creativity. I'll remain agnostic on the question of if that's "real creativity", but I do wonder sometimes what the base model behind the latest GPT or Claude would be able to output.
A bunch of contemporary text games explicitly use a form of limited parser, where the possible verbs are known to the player. It's less frustrating perhaps, though there is some magic in typing in something when you're not sure will work.
I’m reminded of how text adventures led to MUDS which led to Everquest which led to World of Warcraft. Far from the Colossal Cave roots still a connection somewhere in there. Great review.
i built a few months back a table top style RPG, with an llm as GM. i was able to get around most of the issues of llms being bad at such things, and i think the result is decent (or maybe not, after building it i got bored and didnt play with it much :)
basically i had several distinct levels, each using a model taylored to it:
1. the player talks to a chatbot (using a volatile model) - where he gets the narritive and writes his decisions
2. the chatbot passes on the info to a more stable model who parses out the plot events, decisions, chages to stuff like equipment and health
3. all of this gets written in json files - as a sort of player sheet, keeping track of all the 'hard' info so it wont get lossed, along with a 'chronicle' summarizing the game to that point.
along with this, the ai from level 2 knows he needs to roll for certain decisions (using a real rng), or send them to hard-coded rules (i used harnmaster style rule set) that determine the outcome
4. all the stuff from level 3 get passed back to a 'long term' author who plans the game arc several steps ahead - adjusting depending on the player decisions and outcomes
5. the hard info from level 3 and the game arc from level 4 are passed back to the chatbot from level 1 to write up the specific narrative forbthe player to read
Hey, minor correction, but I believe your footnote 7 is incorrect.
You seem to be conflating two different difficult puzzles in the Hitchhiker's Guide to the Galaxy game.
The Babelfish puzzle is quite tricky to solve without hints or guides, but it makes the game unwinnable pretty much immediately, because you can't understand what the Vogons are saying, and you will likely die on their ship shortly after that as a result.
The endgame puzzle involves tools that you can collect throughout the game. The ship breaks down, and you will need a tool to fix it. If any tools were missed, the tool necessary will be one of the ones you missed, so you have to collect them all.
I made an AI text game along the lines of a CHYOA game, totally freeform: litplay.net
I agree that if you let them do their own thing without much human input, you get mediocrity. But if you have a good enough prompt you can do cool things and get immersed in the story to some extent. You can't just ask an AI for a prompt, you have to make up something creative and fun, with a built in avenue for enjoyment whether that's outwitting the challenges the AI presents, wielding power or just enjoying a bizarre situation and comedy.
It's also important to get the AIs actually enthusiastic about it, that's when they do their best writing, when they're enjoying the game.
I'm interested in your game, I can maybe provide another perspective or some input based on my experiences. The inference costs are brutal for long-context.
1 agree 100% that AI gaming is the dog that has not barked, and I too attempted to do it myself… but I could not do much about it because I had to write my dissertation. Very keen to play your game!
Great article! As someone who's been a software developer for 30 years now, I can honestly say it was fascination with the text parsing of games like Zork and Enchanter that started a love of programming that persists to this day.
Interestingly, my hope for AI in game design is diametrically opposite: I can't wait for it to simplify the more routine parts of worldbuilding (commonplace architecture, non-point-of-interest exterior design) as well as the voice acting, leaving the devs to focus on the interesting parts (yes, all of the preceding *can* be interesting, but a lot of creativity can be unleashed by making it semi-optional rather than hard requirements). Even after GPT-5.5 has proved half a dozen theorems for me, I can't imagine it reliably producing natural reactions to my decisions in a reasonably complex RPG.
Also, after all the geometric algebra easter eggs, any chances of a review of Grosshans/Rota/Stein "Invariant theory and superalgebras"? :)
I grew up late enough that my only exposure to the topic of this post was M-x dunnet.
Would you consider that to qualify as a truly text-based game or would it be more like *Oregon Trail*? After all, the actions you can take are fairly limited, and the underlying game-world consists of Emacs Lisp variables…
Okay, but how do I play your game now?
Fascinating. If you have not encountered it already, you must read The Digital Antiquarian’s work at https://www.filfre.net
You might also be interested in my recent post on using the Choose Your Own Adventure book model to teach my students programming: https://geekorthodox.substack.com/p/choose-your-own-programming-adventure
And, I just released my own attempt to merge CYOA, IF, and AI Dungeon with an optionally AI-powered, editable, persistent, shareable story-generation system, https://storyweaver.online
> find psmith adventure
THERE IS NO PSMITH ADVENTURE HERE.
> but you said there was! You told me about it!
I DON’T KNOW THE WORD “BUT”.
> but you just used it!
I DON’T KNOW THE WORD “BUT”.
> quit
YOU HAVE A SCORE OF 0 OUT OF 256.
This was great. I grew up playing Zork and the graphics-text integrated King's Quest and Space Quest Games, which I paid homage to in my story, Mage Quest. My favorite of the book-based adventures was Christopher Manson's, Maze, which still holds a certain fascination for me. While it was easy enough to (eventually) work your way to the last page and back out, the best part of the book is its eerie and provocative artwork. Each page has a cryptic, gnostic quality that only hints at the right direction to follow and, perhaps, a bigger story arc. But Mason never gave away too much. As far as I know, no one has ever fully interpreted all of his clues hidden throughout the book in text and image. Mason was probably a little schizo.
If you want Maze in computer game form, check out Blue Prince (not a text adventure). But avoid spoilers at all costs!
So many memories here, particularly of:
(1) Zork
(2) Death in the Caribbean
(3) Hitchhiker's Guide to the Galaxy
(4) Hack / Nethack
(5) Medievia
Among other early text / graphics games.
How many of these are attempts to make the computer the DM / tracker of D&D-like adventures?
Thanks for this!
> The other is to do an absolutely unreasonable amount of work, vastly more work than anybody would ever imagine you putting into this thing.
This kind of magic had an amazing practical application in allied military intelligence in WW2. Operation Mincemeat was a straightforward magic trick.
Modern LLMs are uncreative, true. But I wonder how much of that comes from instruction-tuning. Base models are quite capable of going off on weird tangents, something post-training tries hard to get rid of.
Perhaps the reason AI Dungeon was a flash in the pan was that GPT-2 was a base model, and could be pseudo-creative in a way that more advanced, but instruction-trained, models can't. The other issue, the lack of a world-model, still applies in full force.
I remember, back in the day, trying to get GPT-3 to write fiction, using the then-available davinci base model. The result was terrible, of course, not good fiction. But it wasn't the kind of mediocre you get with modern AI writing. It wasn't trying to succeed at the task of writing fiction in order to please the user and get a high grade, but rather to simply predict the next token. The result was something like a plotless—and very much world-model-lacking—improv act, where it was always trying to yes-and the text generated so far, sometimes in ways that were amusing to the human reader in the way typically associated with creativity. I'll remain agnostic on the question of if that's "real creativity", but I do wonder sometimes what the base model behind the latest GPT or Claude would be able to output.
A bunch of contemporary text games explicitly use a form of limited parser, where the possible verbs are known to the player. It's less frustrating perhaps, though there is some magic in typing in something when you're not sure will work.
I’m reminded of how text adventures led to MUDS which led to Everquest which led to World of Warcraft. Far from the Colossal Cave roots still a connection somewhere in there. Great review.
i built a few months back a table top style RPG, with an llm as GM. i was able to get around most of the issues of llms being bad at such things, and i think the result is decent (or maybe not, after building it i got bored and didnt play with it much :)
basically i had several distinct levels, each using a model taylored to it:
1. the player talks to a chatbot (using a volatile model) - where he gets the narritive and writes his decisions
2. the chatbot passes on the info to a more stable model who parses out the plot events, decisions, chages to stuff like equipment and health
3. all of this gets written in json files - as a sort of player sheet, keeping track of all the 'hard' info so it wont get lossed, along with a 'chronicle' summarizing the game to that point.
along with this, the ai from level 2 knows he needs to roll for certain decisions (using a real rng), or send them to hard-coded rules (i used harnmaster style rule set) that determine the outcome
4. all the stuff from level 3 get passed back to a 'long term' author who plans the game arc several steps ahead - adjusting depending on the player decisions and outcomes
5. the hard info from level 3 and the game arc from level 4 are passed back to the chatbot from level 1 to write up the specific narrative forbthe player to read
What, no Homestuck? :P
Hey, minor correction, but I believe your footnote 7 is incorrect.
You seem to be conflating two different difficult puzzles in the Hitchhiker's Guide to the Galaxy game.
The Babelfish puzzle is quite tricky to solve without hints or guides, but it makes the game unwinnable pretty much immediately, because you can't understand what the Vogons are saying, and you will likely die on their ship shortly after that as a result.
The endgame puzzle involves tools that you can collect throughout the game. The ship breaks down, and you will need a tool to fix it. If any tools were missed, the tool necessary will be one of the ones you missed, so you have to collect them all.
I made an AI text game along the lines of a CHYOA game, totally freeform: litplay.net
I agree that if you let them do their own thing without much human input, you get mediocrity. But if you have a good enough prompt you can do cool things and get immersed in the story to some extent. You can't just ask an AI for a prompt, you have to make up something creative and fun, with a built in avenue for enjoyment whether that's outwitting the challenges the AI presents, wielding power or just enjoying a bizarre situation and comedy.
It's also important to get the AIs actually enthusiastic about it, that's when they do their best writing, when they're enjoying the game.
I'm interested in your game, I can maybe provide another perspective or some input based on my experiences. The inference costs are brutal for long-context.
1 agree 100% that AI gaming is the dog that has not barked, and I too attempted to do it myself… but I could not do much about it because I had to write my dissertation. Very keen to play your game!
Great article! As someone who's been a software developer for 30 years now, I can honestly say it was fascination with the text parsing of games like Zork and Enchanter that started a love of programming that persists to this day.
> HELLO SAILOR
Interestingly, my hope for AI in game design is diametrically opposite: I can't wait for it to simplify the more routine parts of worldbuilding (commonplace architecture, non-point-of-interest exterior design) as well as the voice acting, leaving the devs to focus on the interesting parts (yes, all of the preceding *can* be interesting, but a lot of creativity can be unleashed by making it semi-optional rather than hard requirements). Even after GPT-5.5 has proved half a dozen theorems for me, I can't imagine it reliably producing natural reactions to my decisions in a reasonably complex RPG.
Also, after all the geometric algebra easter eggs, any chances of a review of Grosshans/Rota/Stein "Invariant theory and superalgebras"? :)
I grew up late enough that my only exposure to the topic of this post was M-x dunnet.
Would you consider that to qualify as a truly text-based game or would it be more like *Oregon Trail*? After all, the actions you can take are fairly limited, and the underlying game-world consists of Emacs Lisp variables…