I'm a little surprised that while using language models for generating text is suggested, there's no suggestion at using other ways to parse input. It feels like there is some much more advanced technology here that could be put to use.
There's a certain danger to letting the AI "find" the answer, that is you are trying to find an overlap between what the user entered and the options that are available in a situation, and since the program can enumerate the options it might find solutions that the user didn't _really_ identify. But if you are comfortable with a game that might err on the side of being easy... at least I would prefer that, as I tend to find IF too frustrating to play because it's like I'm trying to get in the author's head and figure out what they want me to type.
It would be interesting to think about how GPT-3/etc can be used to generate stable non-hallucinatory descriptions. It could solve some of the combinatorial problems of state and generating descriptions, and perhaps even raise interesting and unplanned events. But to do that you'd almost have to parse the results and then incorporate them into the world (or reject the results and regenerate). That's... actually how a lot of Inform works, so maybe it's a real option? It seems particularly interesting for NPCs, where perhaps you could use a language model for responding but _also_ allow for the model to suggestion NPC actions. By parsing them similar to how you parse player input you can keep the NPCs from being able to do impossible things (a big problem in AI Dungeon) but allow them to attempt actions. Then inform becomes the stable model of the world in which both players and bots are doing things. (This idea actually has me kind of excited...)
There's a certain danger to letting the AI "find" the answer, that is you are trying to find an overlap between what the user entered and the options that are available in a situation, and since the program can enumerate the options it might find solutions that the user didn't _really_ identify. But if you are comfortable with a game that might err on the side of being easy... at least I would prefer that, as I tend to find IF too frustrating to play because it's like I'm trying to get in the author's head and figure out what they want me to type.
It would be interesting to think about how GPT-3/etc can be used to generate stable non-hallucinatory descriptions. It could solve some of the combinatorial problems of state and generating descriptions, and perhaps even raise interesting and unplanned events. But to do that you'd almost have to parse the results and then incorporate them into the world (or reject the results and regenerate). That's... actually how a lot of Inform works, so maybe it's a real option? It seems particularly interesting for NPCs, where perhaps you could use a language model for responding but _also_ allow for the model to suggestion NPC actions. By parsing them similar to how you parse player input you can keep the NPCs from being able to do impossible things (a big problem in AI Dungeon) but allow them to attempt actions. Then inform becomes the stable model of the world in which both players and bots are doing things. (This idea actually has me kind of excited...)