How we pick
Hundreds of candidates in.
The best few dozen out.
- 01
Collect
Every morning we pull a few hundred fresh candidates: the Hacker News front page and best list, Lobsters, fast-rising new GitHub repositories, Hugging Face daily papers and trending models, Hacker News searches for AI and systems topics, release notes from the inference and agent frameworks engineers run in production, and engineering blogs from AI labs, infrastructure teams and writers worth following. Anything already published in the last week is dropped, and so are science and medicine journal links that are not about ML.
- 02
Read
For each candidate we fetch the linked page and keep an excerpt of its own text, plus the preview image the page itself declares.
- 03
Judge
A language model gives every candidate two separate marks. Quality, from 1 to 10: would a senior engineer want to read this? Depth, evidence and hard-won production lessons score high; a bare title with nothing behind it cannot score above 5. Relevance, from 0 to 3: how close it is to our beat of agents, inference, ML and system design. Off-beat stories (biology, politics, gadgets, retrocomputing) are dropped however good they are.
- 04
Write
The model writes a plain headline and one reason to read it, using only the title and excerpt. For the stories that make the cut it adds two or three brief points and, when the article describes a concrete architecture, a small diagram of the components it names. Once the edition is picked, it writes “Today in one minute”: three or four sentences on what mattered and how the stories connect, pointing only at stories in the edition. When fewer than five stories rate 8 or more, it says plainly that it is a quiet day. It never chooses links or images, and anything it writes that contains a URL is thrown away.
- 05
Balance
Up to 50 are published. Nearly half the edition is reserved for AI (agents, inference, models, research) and a quarter for systems and architecture, filled only with stories scoring 7 or more. General security is capped at 6 and launches at 6, so they cannot crowd out the beat. Must-reads are the best AI story, the best story from another desk, and the next best. Near-duplicates from different sources are merged, and on a thin day the edition is shorter rather than padded.
- 06
Keep
Every story stays findable. Search looks through every edition by headline, summary, source and author, filtered by topic and kind. The weekly edition gathers the ten that mattered most each week. The AI Radar lists the models trending on Hugging Face and the new repositories rising fastest on GitHub, exactly as those sites report them.
- 07
Illustrate
Lead stories carry the source's own preview image, downloaded once and re-encoded so your browser never contacts a third party. Stories without one get cover art generated from their kind and topic. The pulse line and topic mix are drawn from the day's scores.
Blogs we read
Cloudflare Blog · Netflix TechBlog · GitHub Engineering · Engineering at Meta · Google Research · OpenAI · Stripe Engineering · Dropbox Tech · Hugging Face · PyTorch · Kubernetes · LWN.net · Simon Willison · Julia Evans · Lilian Weng · Sebastian Raschka · Rust Blog · Go Blog · V8 · WebKit · Mozilla Hacks · The Old New Thing · Google DeepMind · SemiAnalysis · Chips and Cheese · Brendan Gregg · Marc Brooker · Metadata (Murat Demirbas) · Martin Fowler · Postgres Weekly · DuckDB · PlanetScale · Trail of Bits · AWS Architecture · Google Cloud · Fly.io · Discord Engineering · Figma Engineering · Jane Street Tech · Shopify Engineering · Airbnb Engineering · Slack Engineering · NVIDIA Technical Blog · Latent Space · Interconnects · Import AI · Chip Huyen · Eugene Yan · Hamel Husain · High Scalability · ByteByteGo · InfoQ · Pinterest Engineering · The Pragmatic Engineer · Armin Ronacher · Will Larson · Andrej Karpathy · Microsoft Research · AWS Machine Learning · Berkeley AI Research · Lobsters (AI, distributed, performance)
Summaries are machine-written from each source's own text. Treat them as a reason to click, and follow the link before relying on a claim.
