I’ve thought a lot about that. I think it definitely does, but I also think there’s a lot of unfulfilled potential left in more “traditional” AI.
There are pros and cons to the GPT approach, with down-sides such as the (current) limit on the context, and difficulty establishing consistent “facts”. These are generally outweighed by the obvious up-sides, but a GPT-based AI will feel different than a more hard-coded one.
All this to say: There will be a hundred better GPT-based games released before I can ever hope to release my project.
Oh, also: the way I’m writing my project, I’m intentionally making the AI modular, so I can “slot in” something like ChatGPT and play around with that.
ChatGPT is amazing for generating reasonable-sounding prose. That means you can have an NPC say things that largely make sense… If you know what they need to say.
For instance, let’s say you ask an NPC for directions to the mayor’s house. In this scenario, you need your NPC to
Parse and understand the player’s speech. (This is a question. The question is about reaching a location. The location is the mayor’s house.)
Modeling the NPC’s knowledge. (Does this NPC know where the mayor lives? Do they know someone else who does?)
Disambiguating conversation. (Do we have a mayor? Are they asking about our mayor, or - from context- someone in a different town?)
Constructing an answer. (Left, right, right, straight…)
Converting the answer to a conversational tone. (Well, if ya’ head south down wewhauken, turn right on glottis st…)
Chat GPT solves #5 for sure. But 1-4 are…iffy. Sometimes, you can give Chat GPT a list of factoids and it’ll reply with the data you gave it. Other times it’ll “hallucinate” an answer, especially if a player asks something you didn’t expect. (And people will definitely come up with stuff you don’t expect.)
Still, LLMs do solve a really hard piece of the dynamic-NPC puzzle. I’m sure we’ll see them in use. It’s just not necessarily even the hardest part of this problem.
I wonder if AI’s like ChatGTP make this more achievable
I’ve thought a lot about that. I think it definitely does, but I also think there’s a lot of unfulfilled potential left in more “traditional” AI.
There are pros and cons to the GPT approach, with down-sides such as the (current) limit on the context, and difficulty establishing consistent “facts”. These are generally outweighed by the obvious up-sides, but a GPT-based AI will feel different than a more hard-coded one.
All this to say: There will be a hundred better GPT-based games released before I can ever hope to release my project.
Oh, also: the way I’m writing my project, I’m intentionally making the AI modular, so I can “slot in” something like ChatGPT and play around with that.
Yes… But actually no.
ChatGPT is amazing for generating reasonable-sounding prose. That means you can have an NPC say things that largely make sense… If you know what they need to say.
For instance, let’s say you ask an NPC for directions to the mayor’s house. In this scenario, you need your NPC to
Parse and understand the player’s speech. (This is a question. The question is about reaching a location. The location is the mayor’s house.)
Modeling the NPC’s knowledge. (Does this NPC know where the mayor lives? Do they know someone else who does?)
Disambiguating conversation. (Do we have a mayor? Are they asking about our mayor, or - from context- someone in a different town?)
Constructing an answer. (Left, right, right, straight…)
Converting the answer to a conversational tone. (Well, if ya’ head south down wewhauken, turn right on glottis st…)
Chat GPT solves #5 for sure. But 1-4 are…iffy. Sometimes, you can give Chat GPT a list of factoids and it’ll reply with the data you gave it. Other times it’ll “hallucinate” an answer, especially if a player asks something you didn’t expect. (And people will definitely come up with stuff you don’t expect.)
Still, LLMs do solve a really hard piece of the dynamic-NPC puzzle. I’m sure we’ll see them in use. It’s just not necessarily even the hardest part of this problem.