[ot][spam][crazy] middle smart adventure game
1. for some time people wanted to make text adventure games that were open-domain, where the player could do anything. some people actually did this, although i don't know where to find their work. such a game featured in the book "ender's game". 2. when transformer language models got big, AI dungeon happened. basically they just told a language model it was an adventure game, and gave the user its output, the same way chatgpt works. the adventures had absolutely no logical continuity, but the user could do anything they wanted to. aidungeon has since worked more on their product. a day or two ago, as part of a trick around my behavior, to try to be productive during a difficult time, i started work on a middle-of-the-road approach, that uses a small language model to generate content and interpret user actions, but can do it more intelligently. i'm aware that others have done this, although again I don't know where to find their work. it's very hard for me, and it was interesting to have a region of it become more easy. NOTE: people call these "text adventures" but nowadays building such a "text" adventure game is _very roughly_ equivalent to building a fully interactive virtual world, because of how AI is being developed. Personally, I'm still just interested in text!
here are files basically i just spasmed at my computer with the goal of making prompts that could do basic things that would be needed for the approach. the work i did is stuff that people using machine learning models have been doing for some time. it's notable that typing a sentence into a language model like "guess the weight of an unripe apple" can actually be (a) condensed into numbers representing the model's state after processing the text and (b) used as seed data to finetune a model specific to the question; so if you're struggling to move forward at all, it can be a reasonable prototyping step. the model is stlil huge and bulky though. i didn't link in an approach to make loading fast, like using a decentralized model. - try_multigen.py is the latest thing i worked on. i kind of want to tune it more because i encounter ram exhaustion when i let it run. it also makes it clear that there's value to running things in parallel, which i didn't implement. pretrained.py was for exploring listing properties of things. try_multigen.py was for exploring creatively generating things. both work enough to move forward if needed, although stimulating that wasn't necessarily my intent. the content is a funny and rare small area of my behaviors. it's confusing, because i am used to taking a more logical approach: as i touch on repeatedly in comments, there's no need for the models, you can just have the users generate content in a concept graph. but the models help one do things, it's emotionally easier to use language models to do things nowadays. the simple idea focus is basically verbs acting based on properties. everything having many properties, having language models able to guess the properties (if the user provides a property it can significantly strengthen the output of a model to add it to the prompt or training data), and using the properties to respond to the user as well as possibly form content. basically it's pretty obvious and the barrier is doing the work.
properties of behavior: - avoids model training this was scary, i was near it and thought about it but refrained from triggering the inhibition; i might have done a little if things had gone the way to - uses python and huggingface - includes concept of 'reacrational game' this was what a lot of energy was around, and i got myself pretty confused with the slide of "wasting time" into "game development". very satisfying, but i didn't realize that i was taxing some of my cognitive traits (in exchange for the opportunity to work others!). - includes pretrained language models toward reducing software design this helped strengthen the "wasting time" and "game playing" concept, resonating around support of mainstream business industry. this really helps me have less internal suffering from my brainwashing. - includes software development with physical result this is really nice for me, spending so much time wasting time, often dissociated from a project i consider important, and always valuing developing my skills, it's quite nice to build something. i really appreciated that. - no stable product users would appreciate this is kind of a secret, it's really very hard for me to meet that goal given my situation, so i think often my parts use our dissociation to pretend it is not the case, and i like to tell the truth always so this is probably hidden from me many more properties exist in general, we are holding thankfulness that the weird and confusing pull that made this left us with ability to think of object-properties, which is a very useful skill for thoughts that have to figure out how to function in very difficult envirnonments. but i am still in a weird state that is a little delusional or trancelike so i have been avoiding the project to figure out if i can do other more familiar things. i missed an appointment and travel plans, i'm having some trouble navigating my coping strategies for eating how i am used to, and i've had some more difficult motor and memory troubles when away from the computer. it is a little similar to waking up in a hospital drugged.
i reskimmed hte source and noted i left out the inhibition around generality. it was interesting to dance around that inhibition, and i also made big strides on small parts of it, somewhat similar to the properties inhibition. content comment there's a lot of mess at the bottom of pretrained.py around discerning whether or not an action was possible. the small models i was using were so poor at that, that i ended up deciding on doing it based on properties rather than action, breaking it into two parts: first associating a verb with what properties are needed to make it possible or not, and then associating what it is done on, with those properties. really satisfying change because of my past habit of using properties to think through confusions, after trying to build a hobby ai as a freshman, and having that inhibited as a coping skill and stuff. the change to properties also makes it more clear how to handle the player changing: e.g. you can only fish if you have a fishing pole, that's a property of the player rather than the fish. since a language model is involved, you likely wouldn't have to make a logical model of the world, you could figure some structure that lets it do so itself.
participants (1)
-
Undescribed Horrific Abuse, One Victim & Survivor of Many