Reply to thread

Message: <blockquote data-quote="Cergorach" data-source="post: 9621025" data-attributes="member: 725">Depends on how you define LLM memory. You either train on it, use RAG, or use the existing context window.Most of the mainstream LLMs offer context windows around the 128k token size, which should fit about a novel. One of Google's LLMs offerts a million tokens worth of context window (~7.5 novels). And there's one odd duck that offers 100 million tokens context window (~750 novels).But it's kind of inefficient of logging all the interactions of the users. Humans tend also not remember all the interactions in exact detail, certainly not a week later. You could easily summarize the interactions, something certain LLMs are very proficient in. Thus making a novel a LOT shorter.You could also log all the interactions in independent files and RAG them, which pretty much allows the LLM to reference the documents without being trained on them.You could also train the LLM on last sessions interactions for next weeks session.OR you could do all three....And while this might take enormous resources if you did this on mainstream generic models, this would be hugely expensive. But keep in mind, that for a specific pnp RPG, the LLM doesn't need all the knowledge in the world/Internet. If you use multiple, highly specific LLMs and chain them together you could get very efficient, especially with what the DeepSeek people have made publicly available.25 years ago the state of VTT was abysmal compared to what it now is. Some people make LLM out to be the 'ultimate solution' tomorrow, others see it as a useless evil. The reality is somewhere in between and certainly not tomorrow or next week. But people have been building some really interesting stuff with LLM and am very curious how that will evolve in the future.</blockquote>

Verification