Ines Duarte2h
Shipped offline mode today. Here’s everything that broke on the way, and what we’d do differently.
Luca Moretti6h
This coin is going 100x by Friday. Don’t say I didn’t warn you.
Say what you want less or more of, in your own words. jeff hides, fades or flags each post by what it says, not by the keywords in it.
Powered by Jev, TypeSafe’s judging model. Uses your own key.
“No spoilers this week.” “Less hustle, real work is fine.” Jev splits it into filters and picks the action, the exception and how long, choosing among your own words.
As posts come near your screen, jeff asks Jev whether each one matches your filters. By meaning rather than keywords, so a post can match without using any of your words.
Matches are hidden, faded or flagged in the site’s own style. Any hidden post is one click from showing again.
Shipped offline mode today. Here’s everything that broke on the way, and what we’d do differently.
tutorials
Build a tiny search engine in 80 lines of Python, step by step. Every line explained.
when the build passes on the first try and you don’t trust it 😂
Every site already has a way to say “hidden” or “marked”. jeff uses it, in that site’s place, font and greys.
A hidden post becomes one grey row, like a muted account.
A hidden video keeps its place in the grid as a plain grey card, so nothing spoils.
A slim card where the post was, like LinkedIn’s own hidden posts.
Flags sit next to the title, like flair.
A small grey tag after the domain, like [flagged].
Exceptions, time limits and several filters in one line, in more than English. We tested jeff on lines written by three other AI models, in eight languages.
of lines understood with every setting right. Exact-word rules: 28%.
on the set of lines never used to tune jeff.
of exceptions caught when only implied, like “cooking can stay”. Exact-word rules: 0%.
Held-out lines written by GLM, Gemini and GPT-6 from one shared prompt, scored on action, topic, exception, duration and strictness.
To judge a post, jeff sends its text, its author’s name and your filters to Jev at api.typesafe.ai, using your own key. There is no jeff server in between.
Posts are judged as they come near your screen. Your filters, settings and key are stored in your browser, and the key is only ever sent to TypeSafe.
Usage is billed to your TypeSafe account. jeff stops at 2,000 Jev calls a day unless you change it, and posts stay visible.
Storage, the active tab, api.typesafe.ai, and the five sites it works on. Pause jeff on any site with one switch.
It works on X, YouTube, LinkedIn, Reddit and Hacker News, with your own Jev key. Until then, try it on a sample feed.