HN user
This is so cool, I'm always speaking to people about how the advancement in the SOTA hosted AI's is also happening in the local model space, i.e. the SOTA hosted AI models 6-12 months ago are what we're seeing now being able to run locally on average hardware - this is such an amazing way to actually demo it.
This looks interesting, I enjoyed the explanation of how RAG works vs this, found it easy to follow. Would like to try this in some projects or claw assistants to see if there's any meaningful improvement in context handling.
I'm a bit confused, it claims to restore full context after compaction but it reads like it's doing it's own form of compaction; if it's restoring the full context how is it avoiding instantly filling up again? If not, why is it's own method of compaction superior to Claude's natives compaction?
Well I certainly got sucked in by some cats staring at the camera with an empty bowl, got me buying them kibble.
I generally agree that strict moderation is the key but there's obviously a certain threshold of users and activity that is hit where this becomes unfeasible - ycombinator user activity is next to nothing compared to sites like Facebook/twitter/reddit. Even on Reddit, you see smaller subreddits able to achieve this.
But just like a public park, if 2 million people rock up it's going to be next to impossible to police effectively.