Skip to content

Comment on StreamingLLM: tiny tweak to KV LRU improves long conversations

Comments

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.

Here they simply used different models for different turns and apparently it gave more "engaging" results:

https://arxiv.org/abs/2401.02994

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.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.