npub1xg…hgu7s on Nostr: Working with AIs has been making me think about how good the human brain is at ...

Working with AIs has been making me think about how good the human brain is at various levels of memory compression. Our compression included hints and other contextual clues that unlock particular memories just when we need them and then recompresses them once they are no longer needed.

I've started thinking about clearing an agent's context window as allowing it to sleep and I prepare it by asking it to record memories and update instructions before I clear, something we do subconsciously and facilitates our own memory compression and long term storage.

quoting
nevent1q…ldk4

Why are LLMs so smart, yet so dumb?

They can handle the most advanced insane coding, but then flop on some super basic task like realizing that if I said I used a tool yesterday, that perhaps this tool is installed on my computer already instead of having to go look for it again...

What is wrong with them 🤣