About and method
AI Radar is a once-a-day summary of what happened in AI, written for developers. It is generated automatically and maintained by @askbatu.
How a day is made
- Collectors read Hacker News, official lab blogs and changelogs, OpenRouter, Hugging Face and GitHub every few hours.
- Items about the same event are grouped into topics. Each topic gets a heat score from recency, engagement and how many independent sources covered it.
- Once a day Claude writes a short title, a one or two sentence summary and the "Today in AI" digest from item titles and short snippets.
- Code checks every link: the model can only point at collected items, never write a URL. Anything referring to a link outside the collected set is dropped.
- The result is published as JSON and this static site is rebuilt from it. Visitors never trigger a model call.
Models used: claude-haiku-5-5, claude-sonnet-5-5, claude-opus-5-5. If the daily budget is reached or the model is unavailable, the edition falls back to titles and links only and says so at the top.
Evidence levels
- high evidence
- An official source, or two or more independent domains.
- medium evidence
- Several community reports without official confirmation.
- low evidence
- A single community post. These never enter the digest and are phrased as claims.
Limits
Claims are not checked by hand. The pipeline checks that each item is tied to a real source, not that the source is right. Summaries can be wrong; follow the link before relying on anything here. Benchmark numbers are copied from each leaderboard as published and are not comparable across boards.
What is published
Only our own short title and summary, the source's original title as link text, its domain, date and engagement counts. No article bodies, comments, images or usernames. The site sets no cookies and has no forms.
Corrections and removal
To correct or remove an item, contact the maintainer (by @askbatu). Removed topics and links drop out of every page, including past days, on the next run.
Data
Everything on the site comes from three files you can download: latest.json, models.json and benchmarks.json. There is also an Atom feed with one entry per day.