Project · Live
Fellow, a live AI news watch
Fellow is the French-language artificial intelligence news watch I designed, built and edit: every day it reads the articles of sources chosen one by one, translates them, sorts them and judges them against a public rubric. It is for people who want to follow AI without reading everything, and for people who write about it.
The figures
Read on fellow.news on , at 3:35 am, Paris time. Each figure links to the page that gives it (in French).
- Sources read
- 4522 RSS or Atom feeds, 11 newsletters, 6 APIs, 6 index pages fellow.news/methodologie
- Editions a day
- 4at 9 am, 1 pm, 5 pm and 10 pm, Paris time fellow.news/methodologie
- Edition at the top of the site
- No. 638that of 9 November 2026 fellow.news, top of the page
- First edition
- 20 April 2026No. 1, in the morning fellow.news/methodologie
- Articles in 35 days
- 1,580 fellow.news, page footer
- Articles judged
- since late August 2026 fellow.news/llms.txt
The need
The starting point fits in one sentence of Fellow’s manifesto: “AI has made the cost of producing content collapse… but not the cost of your attention.” There has never been so much to read about AI, and every headline is written to be clicked.
Following the subject therefore means reading everything, then sorting it and saying why. I built it to do that work every day, showing every judgement.
The pipeline
Six steps, according to Fellow’s public methodology:
- The sources. 45 sources chosen one by one: 22 RSS or Atom feeds, 11 newsletters read for their links, 6 APIs (arXiv, Hacker News, Hugging Face) and 6 pages of labs with no feed, read by an agent.
- Collection. Four times a day, the sources are read in parallel and duplicates removed by address.
- Enrichment. Each article goes through a Claude model, with Mistral as a fallback, which returns strict JSON: a French title and summary, a category, tags and three scores. Reprints of the same announcement are then grouped by meaning; when in doubt, a judge keeps them apart.
- The editorial pass. Thirty minutes later, the model rereads the actual body of the article, when the site allows it, and gives a verdict (major, notable, minor or noise) with a sentence that cites the deciding criterion. A headline that claims more than the article is flagged “headline to be qualified”.
- Publication. Four editions a day, the live feed, the 9:45 am newsletter, the archives and an MCP server for assistants.
- Revision. A verdict can change; each change is dated and says where it came from. The opinion signed “La rédaction de Fellow” (Fellow’s editors) is read by a human before it appears.
The choices
- A rubric written by a human, applied by an AI. The criteria are public and open to challenge (manifesto).
- Two passes, the second on the full text when the site allows it. “We read the body of the article, never just the headline” (manifesto).
- Strict JSON. The model fills in named fields; for lab pages, only addresses present on the page are accepted (methodology).
- The headline gets the benefit of the doubt. “Better a missed flag than an unjustified one” (methodology).
- A crawler that gives its name. FellowBot never passes itself off as a browser; before reading a page on another site, it checks robots.txt and refusals addressed to AI crawlers, and a page that refuses it is not read (our crawler).
- Noise marked, not hidden. “With noise deleted, you can’t check that we judged well” (manifesto).
- An open MCP server. Free, no account, read-only (fellow.news/mcp).
The limits
The manifesto says so itself: “Fellow has no access to the truth. Our verdicts are given by an AI, with a rubric written by a human: it is a situated judgement, open to revision, and sometimes wrong.”
The methodology names a “known limit”, an article’s heat, a subjective estimate by the model; and no human decision enters the collection loop: “errors are inevitable”.
What it shows
Fellow is a personal project, but it is the same work as for a company: start from a real need, here a stream too large to read, and build a pipeline that runs every day, with its safeguards. That is the heart of AI solution development.
Deciding where the AI decides and where a human keeps control is advice. Understanding what a model does and what it doesn’t, and writing rules clear enough to be discussed, is what I pass on in training.