A Nostr relevance feed. Fetches posts from your follows and ranks them by what you actually care about, using an LLM instead of reverse chronological order.

Live demo: https://cypherpunk.today/nalgorithm/

Here's a description of how it works:

You write a short profile describing your interests ("I like cypherpunk culture, Bitcoin, cats, thoughtful longform writing. I don't care about price speculation or GM posts"). The app sends batches of posts to your chosen LLM along with this profile, and each post gets a 0-10 relevance score.

There's also a "learned prompt" that gets built automatically from your Nostr likes (kind 7 reactions). You keep using your regular Nostr client, like the posts you enjoy, and nalgorithm picks up on those patterns. The learned prompt evolves incrementally -- each run only looks at likes since the last run and refines the existing prompt, rather than regenerating from scratch. It only affects future rankings; already-scored posts keep their scores.

Scores are cached locally (localStorage in the web app, a JSON file for the CLI), so re-running only scores posts the LLM hasn't seen before. No wasted tokens.

I wonder if this gets a little expensive, but I'm not enough of a nostr boss to actually use it. I tend not to use nostr for a feed.

Juraj seems to have some interest in running it as a service: