It can be any game that has dialogs. Not necessarily only text one. I will say even more: people are already making games like this. LLMs are actually very good to give some life for NPC, since their dialogs will be different from time to time and actually dependent on what gamer typed into their text field during dialog.
Ah, yes, for dialogue, that makes a lot of sense. I thought you meant physical movement, like using LLM tech to replace traditional NPC AI.
It would be cool if it could be used for custom character names. It really bugs me when the main characters first name is never used in dialogue, because you created it. Especially in the recent Hogwarts game. Maybe the creator could generate the spoken name from what you enter, and let you fine tune pronunciation somehow.
I don’t know, maybe. But avoiding first names sucks. Humans use first names. And the idea I proposed would use the AI initially, but then be fine tuned by hand to sound right, and then saved. It wouldn’t use the AI in real time. If that could be made to work, I see absolutely no problem, but it’s just an idea.
Using LLM for NPC movement is a very bad idea. Normally, people use special neural networks that are trained to decide which action NPC should do next. They are much smaller and don’t work with pure text, but rather with strictly categorized values like position of NPC, it’s stats, surrounding objects, world values(like weather) and etc.
Uh, as far as I know, the concept doesn’t even make sense, which was my point. LLMs have absolutely nothing to do with that type of system, no? It’s a bad idea because it’s impossible, and also I think not what you meant?
What do you mean by system? LLMs are Large Language Models. They are by definition work with text. And they are not designed to work normally with what they are not trained for. But, LLMs are only 1 family of the whole range of AI models. For controlling NPC actions, decisions and other stuff you can use small AI models to make their actions more dynamic. For example, a model that has like 10 input parameters, 3-4 hidden layers and the output layer with a few results. And this thing will actually work for many things, not only for controlling npcs. Even on a microcontroller this thing can be used for operating with fuzzy values. These features for creating AI models you can find in game engines like Libgdx or as a separate framework like Neuroph(for Java).
The architecture of a model and what it is designed for are not strictly connected to the method of learning. For small AI models for decision making you can absolutely easily write a few hundred examples of actions by your hand and then have a tiny model, that will decide what NPC should do in a simple shooting game. Reinforced learning is just a method of learning, where you can’t or don’t want to provide enough training data, so that you provide and an ability for AI model to learn in a sandbox.
Can’t imagine handing control of anything over to an LLM.
Nah, you are wrong. Since LLMs are for entertainment, making it control an NPC is totally fine, since hallucinations will only make this NPC funnier.
That’s in an imaginary world. Not really what I was getting at.
Yeah, I know. But LLMs are an imaginary artificial intelligence, so my point is also can be considered valid.
How are you imagining that go down? Like a game with text based control? I don’t understand.
It can be any game that has dialogs. Not necessarily only text one. I will say even more: people are already making games like this. LLMs are actually very good to give some life for NPC, since their dialogs will be different from time to time and actually dependent on what gamer typed into their text field during dialog.
Ah, yes, for dialogue, that makes a lot of sense. I thought you meant physical movement, like using LLM tech to replace traditional NPC AI.
It would be cool if it could be used for custom character names. It really bugs me when the main characters first name is never used in dialogue, because you created it. Especially in the recent Hogwarts game. Maybe the creator could generate the spoken name from what you enter, and let you fine tune pronunciation somehow.
Mixing generative AI with voice acting sounds to me more immersion-breaking than using ways to avoid naming you directly.
I don’t know, maybe. But avoiding first names sucks. Humans use first names. And the idea I proposed would use the AI initially, but then be fine tuned by hand to sound right, and then saved. It wouldn’t use the AI in real time. If that could be made to work, I see absolutely no problem, but it’s just an idea.
Using LLM for NPC movement is a very bad idea. Normally, people use special neural networks that are trained to decide which action NPC should do next. They are much smaller and don’t work with pure text, but rather with strictly categorized values like position of NPC, it’s stats, surrounding objects, world values(like weather) and etc.
Uh, as far as I know, the concept doesn’t even make sense, which was my point. LLMs have absolutely nothing to do with that type of system, no? It’s a bad idea because it’s impossible, and also I think not what you meant?
What do you mean by system? LLMs are Large Language Models. They are by definition work with text. And they are not designed to work normally with what they are not trained for. But, LLMs are only 1 family of the whole range of AI models. For controlling NPC actions, decisions and other stuff you can use small AI models to make their actions more dynamic. For example, a model that has like 10 input parameters, 3-4 hidden layers and the output layer with a few results. And this thing will actually work for many things, not only for controlling npcs. Even on a microcontroller this thing can be used for operating with fuzzy values. These features for creating AI models you can find in game engines like Libgdx or as a separate framework like Neuroph(for Java).
Yes, you are talking about reinforcement learning, which strictly speaking is not a type of neural network.
The architecture of a model and what it is designed for are not strictly connected to the method of learning. For small AI models for decision making you can absolutely easily write a few hundred examples of actions by your hand and then have a tiny model, that will decide what NPC should do in a simple shooting game. Reinforced learning is just a method of learning, where you can’t or don’t want to provide enough training data, so that you provide and an ability for AI model to learn in a sandbox.