ChatGPT created a text adventure for me | Hacker News Reader

ChatGPT created a text adventure for me

andrewhuth.substack.com

33 pointsahuth16 comments

If like Andrew you find yourself in a situation where a GPT-generated text adventure sounds exciting, let me urge you NOT to do it and browse ifdb.org instead. There are thousands of super creative text adventures there made by a super-active community that are just begging to be played. You could live 10 lives and still find something new to play every day.

Why not do both?

Because the human-created ones are better bya wide margin.

I'd be interested in seeing some generated text adventures that had more object manipulation puzzles (looks like this one only had a key and door puzzle) that you couldn't talk your way around.

I've looked into procedurally generated text adventures before and one challenge for example is not to give players access to objects that can be used to break/skip a current puzzle e.g. if the player needs to get money to buy a knife to cut a rope, don't give the player access to a match before they cut the rope or they could burn through the rope instead. I'd also want to stop ChatGPT from letting you bite through the rope or other solutions.

Tangentially relevant: I made a [Text Adventure GPT]

https://chat.openai.com/g/g-EDKrXhiwo

A while ago and never really use it. I also have found text adventures to be critical to achieve human level (embodied) cognition in LLMs but don’t feel like sharing the Neural Nexus because, well, why the fuck would I help OpenAI when they seem to be monopolistic, unfair, unreasonable, and anticompetitive?

Sadly, I have lost faith in OpenAI’s fundamental alignment safety ethics due to their customer noncompete clause. Here’s a Google drive folder full of evidence of evil at various anticompetitive monopolies including Microsoft and AWS and Anthropic and Inflection, enjoy https://drive.google.com/drive/folders/1n_L_cTUZbPuOiA5Snf6S...

welcome to the real world, it's cruel and unjust, but with some bright spots.

those agreements are as good as they are enforceable. it should be hard to enforce when the source is a derivative from stolen data.

Here is the one I made: https://chat.openai.com/g/g-PoQSljlT5-dungeon-quest

It's inspired by Fighting Fantasy. Uses Python to do dice rolls such as for creating the character.

Generates images and tracks things like inventory and stats consistently.

Played it for a bit - very cool! Is there a way in GPTs to use the Python variables to control the prompts? Curious how that part works.

Thanks.

You just check the Analysis/Interpreter box and tell it to how and when to use Python in the GPT instructions.

I put a mini Python lib for rolling dice and skill checks in the GPT instructions and just enabled the Analysis or whatever checkbox. And I told it to run the code in the beginning to initialize and use the functions for dice rolls etc.

It can write and call functions on the fly, but I wrote them ahead of time and have it call the library functions to reduce the amount of code needed to the minimum to try to speed things like dice rolls up.

This is pretty neat but my first thought was "is this producing valid Inform7 code?"

Holy smokes, I've never heard of Inform7, but I wouldn't be surprised if ChatGPT could output it for an adventure it creates

It most probably can, but every time I asked it to create/be a text adventure it lost track of state pretty soon wich rendered the game unplayable.

I think it would be interesting to put an LLM in front of a text adventure engine so that it could interpret user input to try to match with available commands.

Here are my experiments in this domain concerning complex interaction between the plane of narrated action and the plane of narration, i.e. having characters telling you stories ... within stories ... within stories.

But stories you can interactively drive nonetheless, as if you were in their shoes, while you listen them tell you what happens in reaction one narration level above. You can even argue with the narrator about what really happened rather than act in the story they tell.

https://pastebin.com/fK07dNhD

-----------------------Player input ---------------------------

//"Nah, Ol Rythm, I know Nancy story's by heart, this isn't what happened. It wasn't a sandstorm in the desert. It was a rainstorm in the jungle. You're mixing up everything in your boozed up mind bro"

--------------------------------------------------

Level 0: Ol' Tales – Player Challenges Ol' Rhythm "Wait, hold up," you correct Ol' Rhythm with a playful jab, "You're mixin' up the stories. It wasn’t no desert – it was the jungle, man. Rain pourin' down thick as mud, not that scorchin' sun and sand. You sure that whiskey ain’t messin' with ya memory?"

Ol' Rhythm throws his head back and laughs, a deep, belly rumble. "Aight, aight, you got me. Maybe I did have one too many. Alright then, the jungle it was," he concedes, his grin acknowledging the mix-up. "Nancy's convoy was knee-deep in a downpour, visibility nil, the whole deal. So there they were..."

Level 1: In the Jungle – The Convoy Encounter "...mud up to our axles and water pourin' down like we’d angered the gods," Nancy's voice takes over, now steeped in the greens and browns of the jungle, a torrential rain replacing the arid breeze, "And just when we thought it couldn’t get no worse, we heard it over the rain – the distinct chatter of AKs. We were about to have company, and they weren't there to share umbrellas."

Your eyebrows shoot up in intrigue, and you lean in a little closer, the bar around you fading into the shadows of a rain-soaked jungle on the edge of chaos. gpt-4-1106-preview using 2392 tokens ~= $0.030120

cool, tried it with chatgpt-4. it's more verbose, and (!) you can ask for illustrations along the way.

Reply on news.ycombinator.com