How AI News Pulse works

Sources 22 sources

A background worker reads 21 sources, each on its own schedule, every 2 to 15 minutes. Hacker News and Hugging Face papers only count once a post has enough points or upvotes. Sources marked off are not read.

So far that has made 77 stories from 88 articles, starting Oct 12.

How a story is made

1. The worker reads every source for new articles and keeps each article once.
2. An article that links to a page already on the wire joins that story straight away.
3. Every other new article is read by Claude Haiku. It decides whether the article is really about AI, then writes a short headline and summary in its own words, using only facts from the article.
4. It compares the article with the stories of the last three days. When one reports the same event, the article is merged into it, so a launch covered by ten outlets is one story with ten sources. A lab's own post becomes the lead link when it joins.
5. It tags the labs, models and topics, marks the story confirmed, rumor, research or community, and logs model launches for the release timeline.
6. New and updated stories reach open pages within seconds.

When no AI is available

If no Claude key is set, the daily spending cap is reached, or the API stays down for a while, simple rules write the story instead. The original title becomes the headline and the first sentences of the article become the summary. Labs are guessed from keywords, and articles with very similar titles from the last two days are merged. Only articles that are clearly about AI get through.

Once AI is back, stories the rules wrote in the last 12 hours get a full pass. Every story page says whether Claude or the rules wrote it.

Rumor and confirmed

Every story carries one label.

Confirmed

An official announcement or release, or a credible report of something that has already happened.

Rumor

An unconfirmed report: a leak, anonymous sources, strings found in code, or plans that are only reportedly in the works.

Research

A paper or a research result.

Community

Discussion, tutorials, opinion pieces and community projects.

New record

Posted by the site itself when a model takes first place on the Artificial Analysis Intelligence Index, or raises the top score.

A new rumor opens a claim on the rumor tracker. Later reports that repeat it add to its source count. It stays open until a story officially confirms it or an official denial arrives, and then it links to that story. An editor can also settle or hide a rumor by hand.

When a lab's own post joins a story that the press reported as a rumor and confirms it, the story turns to confirmed. Without AI, a story is marked rumor when its title uses words such as reportedly, leaked or sources say.

Benchmark numbers 12 boards

This site never runs a benchmark. Every score is a number already published on an official leaderboard, in a lab announcement or in a paper, and older scores are kept as history.

128 scores read from the Artificial Analysis API. 52 scores collected by hand for the launch of the site. When an Artificial Analysis API key is set, the worker also refreshes the Artificial Analysis boards every six hours.

• Artificial Analysis Intelligence Indexartificialanalysis.ai45 scoresArtificial Analysis Intelligence Index.
• LMArena textlmarena.ai3 scoresElo from an Oct 8 snapshot via an aggregator; LMArena blocks direct reads.
• SWE-bench Verifiedswebench.com9 scoresLab-reported. OpenAI stopped reporting it in February 2026.
• Artificial Analysis Coding Indexartificialanalysis.ai23 scoresArtificial Analysis Coding Index.
• Humanity's Last Examscale.com8 scoresScale's public board, with lab reports noted per score. Settings vary.
• ARC-AGI-2arcprize.org3 scoresSemi-private set, via the BenchLM mirror.
• GPQA DiamondSource linked with each score1 scoreClose to saturated.
• OSWorld-Verifiedos-world.github.io1 score
• Output speedartificialanalysis.ai32 scoresMedian output tokens per second, Artificial Analysis.
• Text-to-image arenaartificialanalysis.ai15 scoresArtificial Analysis text-to-image arena, Elo. Price per 1024 by 1024 image.
• Text-to-video arenaartificialanalysis.ai20 scoresArtificial Analysis text-to-video arena with audio, Elo. Price per second from 1-minute 1080p pricing.
• Speech arenaartificialanalysis.ai20 scoresArtificial Analysis speech arena, same cloned voices for every model, Elo.

Feeds

The whole wire is published as RSS at /feed.xml and as JSON Feed at /feed.json, with the 50 newest stories in each. Add ?lab= or ?topic= for one lab or topic, for example /feed.xml?lab=OpenAI. The news page shows the matching feed links for its filters.