• by Archelaos on 10/6/2025, 7:43:54 PM

    Isn't the whole system somewhat too limited? An interaction can be unpleasant and pleasant at the same time, like getting food offered I dislike when I am very hungry. Instead, shouldn't it be preferable to model the inner state of NPCs with a high dimensional vector space plus fading memory? Folk psychology emotions could than be interpreted as reductions of this high dimensional vector space. We also could use a part of the vector space for a dynamic feedback system, such as adjusting the fading of a particular memory depending on particular values (intens events remain longer in memory, ...) an overall condition (repeated recent events remain longer in memory; certain recent events in close proximity in the vector space improve their memorization, ...), other memories, etc.

  • by 7373737373 on 10/6/2025, 5:47:23 PM

  • by mvieira38 on 10/6/2025, 4:46:18 PM

    Excited for stuff like this as a CRPG player. A possible future for designing a crpg NPC might be to write a bunch of memories, descriptions, likes and dislikes, etc. in text instead of trying to convey those through branching dialogue.

  • by bbminner on 10/6/2025, 7:09:04 PM

    Could you provide an example conversation that showcases the quality? For example either inferring states of npcs and re-ranking response options?

    Kind of reminds me of the social interaction puzzle in oblivion :)

  • by nathan_douglas on 10/6/2025, 7:48:26 PM

    This is really cool - the sort of thing I love seeing on HN. I'm very deeply interested in developing/interacting with realistic characters within a game world and watching them interact with another, e.g. intertribal/interpersonal squabbles in a kobold tribe, that kind of thing. I think the potential for this sort of experience in gaming is really untapped and is going to be incredible.

  • by Arch485 on 10/6/2025, 7:22:13 PM

    Is there any specific reason to use this over e.g. manually setting an "emotional score" for specific dialogue? Maybe I'm misunderstanding how this works, but from a game design perspective I'd be a bit worried about the model inferring the emotion incorrectly, or more generally having a lack of control over the specific values being used.

  • by oneandonley1 on 10/5/2025, 5:44:32 PM

    >call the matrix functions directly in C++

    Sweet so it can be used in unreal engine it would be awsome to see this used for a local llm game that can generate it's own unique NPCs

  • by Charon77 on 10/7/2025, 8:49:38 AM

    Looks really cool but I'm wondering if the whole Docker setup was necessary just to get onnx?