How we pick

AI Worldwire publishes one article a day and, from November 2026, up to three Builder's Signals: papers, models or code repositories worth a builder's attention. This page explains how both are chosen. Verification rules are on the Methodology page.

Where candidates come from

• News: primary and regional sources in every region we cover, from the last 24 to 72 hours.
• Papers: new submissions to arXiv in AI, computation and language, and machine learning (cs.AI, cs.CL and cs.LG), and the papers the Hugging Face community upvotes in its Daily Papers list.
• Models: models trending on Hugging Face, and official model announcements.
• Repos: GitHub repositories created in the last seven days with an AI topic and at least 20 stars.

An automatic scan collects these candidates every morning through public, machine-readable interfaces. It does not scrape sites that ask automated readers to stay away, and it does not scan X.

The rubric

Each candidate is scored from 1 to 5 on:

1. Significance to people who build with or decide about AI
2. Novelty: is it genuinely new, or a repeat?
3. Verifiability: is there a strong primary source (the paper, the model card, the repo, the official notice)?
4. Reader demand: are people looking for it?
5. Cross-region relevance: does it matter beyond one country?
6. Fit with the day's article type and our open topic series

The highest total becomes the lead story. For Signals we also ask: can a reader use it now (published weights, runnable code, a readable paper)?

Rules for Builder's Signals

• Three at most a day, inside the daily article and newsletter, never as separate thin pages.
• At least one from outside the US whenever a qualifying item exists.
• Primary link only: each Signal links to the paper, model card or repository itself.
• One line on why it made the cut, for example "open weights, Apache 2.0 licence, runs on one GPU".
• The authors' claims stay theirs. Benchmark results are "according to the authors" until an independent evaluation exists.
• Popularity is not proof. Stars, likes and upvotes can be gamed. We may report them as "N stars as of [date]" but we never pick an item on popularity alone.
• No paid placement. Nobody can pay to appear in Signals or in the news. Any sponsorship is labelled and kept separate (see Advertising and disclosure).
• A human editor approves every Signal with the article it appears in.

The weekly numbers

The Signals page shows how much our scan looked at in the last seven days and how many items made the cut. The counts come straight from the public interfaces listed above; we do not round them up or add sources we did not scan.

Open data

Signals and the open-weights tracker can be downloaded as CSV or JSON and reused freely with a link to AI Worldwire. Spotted an error? Tell us.