[alexitorg]

I like how they summarise the content history, I wonder if you could get the Large language model (LLM) like ChatGPT-4 to do the summarisation once the token window was full?

I was thinking about how you could do something like this. Prime the LLM with description of the character and feed the dialogue back into it to generate realistic NPC dialogue. Have some logic behind it to update the chat log with events, perhaps quest prompts. Doable with the current price of$0.03/1k prompt tokens. Not sure you could do this on a triple A game with the dependency on an outside provider that might jack the prices and possible impossible to predict behaivour. But it might be fun in a small indi game?

I wasn't anticipating the interaction between the NPCs. That would make it even more immersive. I am also curious how they got the output in chat-gpt to influence the NPCs movements and actions?