Less noiseMore signal
Know Before You Click.
Winnow tells you whether an article or video is worth your time before you open it. Read, skim, save, or skip, with a confidence, tuned to what you actually care about.
youtube.comSimulated
Deep dive into KV caching for LLMs
READ· 92%
- High insight density
- Mostly new to you
- Not clickbait
- Best part: 04:32 → 08:10
- Deep dive into KV caching for LLMs youtube.comREAD· 92%
- 10 productivity hacks that changed my life medium.comSKIP· 91%
- Why AI changes EVERYTHING youtube.comSKIP· 95%
- Notes on Raft, twelve years later usenix.orgSKIM· 84%
- The one tool every developer needs in 2026 sponsoredSKIP· 89%
- Is this the end of programming? substack.comSKIP· 87%
- A field guide to distributed consensus arxiv.org · 48 minSAVE· 81%
- You won't believe what this model can do youtube.comSKIP· 93%
What it costs
- $0.0015
- Per Hacker News front page, 30 links*
- 0 tokens
- On reload. Verdicts are cached for seven days.
- 1.2 s
- Per judgment, measured*
*Measured 2026-09-18 against jev-1.13.0. Proof: docs/eval.md and docs/jev-contract.md
It works where you already browse.
- Deep dive into KV caching for LLMs 41:05 High insight density · Mostly new to you · Best part: 04:32 → 08:10
- Building RAG from scratch 24:18 Good explanation, partly familiar · Best part: 09:20 → 14:00
- Why AI changes EVERYTHING 18:42 Low new-information density · Likely engagement bait
- Transformers explained from scratch 32:10 You already know most of this · Overlaps with what you read this month
- Notes on Raft, twelve years later usenix.org High technical density · New to you · Matches your systems interest
- The one tool every developer needs in 2026 blog.example.com Likely an undisclosed sales pitch · Reads as an advertorial
- A field guide to distributed consensus arxiv.org Original research · Long read: ~35 min · Serves your goals
- Show HN: I rewrote my blog in Rust github.com One useful idea · Mostly familiar ground
- What we learned running Postgres at scale engineering blog Dense, specific, non-obvious · Serves your goals
- 10 productivity hacks that changed my life medium.com Listicle · Low new-information density
- Is this the end of programming? substack.com Opinion · Claims not supported · You have read this take before
- A practical guide to Raft leases personal blog Tutorial · Payload sits in the second half
Hacker News, YouTube, Any Link-Heavy Page
In feed mode Winnow reads the titles and snippets already on the page, judges up to twelve links in a single call, and drops a pill next to each one. Reload and nothing is sent: verdicts are cached for seven days.
Hover For The Card. Nothing On The Page Moves.
Pills are rendered in a shadow DOM, sized to the line they sit on. Hover or focus one for the reasons. The page you are reading keeps its layout, its scroll position and its scripts.
Skip to the part that matters.
Winnow asks one typed question about eight caption segments. The answer is a probability per segment. The tallest one is where you start.
Tap a column for its caption.
Jump to payoff09:20 → 14:00 · 4 min 40 s of 18:42
- 4:40
- worth watching, out of 18:42
- 25%
- of the runtime
youtube.com · 32:10Simulated
Transformers explained from scratch
SKIM· 67%
- General signal quality
- No goals set yet
- Nothing in your history on this topic
What Reader State Is
Three things: the goals you type into the options page, a rolling window of topics you have recently read, and the read/skip feedback you give on cards. All of it stays in the extension's local storage.
Change Your Goals, Change The Verdicts
Reader state is part of every question, so editing your goals invalidates the cache. The next page you open is judged fresh against the new you.
Some things deserve a quiet skip.
Rage Bait
SKIP· 96%
Possible engagement bait.
Low new-information density.
Possible Undisclosed Ad
SKIP· 89%
Steers you to a product without saying so.
Reads as an advertorial.
Likely AI-Generated Filler
SKIP· 85%
Uniform rhythm, no specifics.
Insight density 0 of 2.
Simulated. Junk is a typed question too; the answer is a skip, not an alarm.
Open source from day one.
What You Get
A Chrome MV3 extension, MIT licensed, at github.com/ThinkyMiner/Winnow. No backend. No analytics, no telemetry. Your key stays in local storage.
Where Your Text Goes
Page text goes to api.typesafe.ai and nowhere else. This is early access: expect rough edges, and read the privacy note before you install.
Install
-
01 Download
Get the latest release zip and unzip it anywhere you will not delete.
-
02 Load Unpacked
Open chrome://extensions, turn on Developer mode, click Load unpacked and pick the folder.
-
03 Paste Your Jev Key
Create one at console.typesafe.ai. About a tenth of a cent per Hacker News page.
- Chrome Web Store: coming soon.
Questions
How much does it cost?
About $0.0015 for a 30-link Hacker News front page, measured against jev-1.13.0. A single article or video is a fraction of that. Verdicts are cached for seven days, so reloads cost nothing.
How accurate is it?
80% verdict agreement on our 40-fixture held-out split. The goldens are under review, so treat that as a floor we are still checking, not a benchmark.
What about paywalls?
Winnow judges what it can see. On a paywalled page that is the teaser, so the verdict is about the teaser and says so.
Videos without captions?
You still get a verdict from the title and description, but there is no skip-to. Winnow does not guess timestamps it cannot read.
Other Chromium browsers?
Anything that loads MV3 unpacked extensions should work: Edge, Brave, Arc. We test on Chrome.
Firefox?
No, not yet.
Does it slow pages down?
No. Pills render in a shadow DOM and every call is asynchronous. The page finishes loading exactly as it would without Winnow.
Can I change the questions?
Yes. They live in src/jev/questions.ts. Edit, rebuild, reload the extension.