News for hackers. Stuff that matters.
Re:StreamingLLM: tiny tweak to KV LRU improves long conversa... by gremlinsinc on Tuesday February 13, 2024 @06:54AM (#39356827)
I wonder if it could make sense to maybe have break away bots, where at 10k tokens a new one launches with the first 2k, and the last 1k and a table of contents such that when you go back to something you're handed off to a model where that data is stronger reinforced or something like that. Sort of like mixture of experts but they're only an expert about individual snippets of a long conversational thread.
Re:StreamingLLM: tiny tweak to KV LRU improves long conversa... by joshspankit on Tuesday February 13, 2024 @08:58AM (#39357638)
You’re right: A lot of the conversation can be condensed, especially if there are enough cues for the AI to arrive in the same “neuronal neighborhood” as the previous conversation.
Re:StreamingLLM: tiny tweak to KV LRU improves long conversa... by kgeist on Tuesday February 13, 2024 @10:09AM (#39358282)
Here they simply used different models for different turns and apparently it gave more "engaging" results:

https://arxiv.org/abs/2401.02994