semisol
· 7w
LLMs are decent if what you are trying to do is in the training distribution. They suck otherwise.
80% of stuff on Nostr here is. (Not sure if that is good for innovation) and even then you can see t...
can confirm. i have put a set of heuristics in my claude.md that contraduct most of the stackexchange training data for coding. it drifts a lot lees though now. you can go a long way before the hard limits set in. enough that i am reaping the benefits of unorthodox architecture philosophy in my config. there is a tiny bit of congruent material in there and it selects on that. skills also expand it. i have a nostr skill set up, and the nips had drifted a lot since i first made it. updated it last week. helps a lot when i am doing protocol work