How we score

30 RSS and Atom feeds, in three tiers. We take the 10 newest items from each feed on every run and ignore anything more than 10 days old. Some feeds are unavailable on a given run.

• Tier 1, organisations' own blogs and newsrooms: OpenAI, Google DeepMind, Google (AI), Google Research, Meta (AI research), Microsoft, NVIDIA, Hugging Face, Mistral AI and AWS Machine Learning.
• Tier 2, news outlets: TechCrunch, The Verge, Ars Technica, Wired, VentureBeat, MIT Technology Review and The Decoder. We use each outlet's AI feed; The Decoder covers only AI.
• Tier 3, research feeds, newsletters, forums and personal blogs: arXiv (cs.AI and cs.LG), Apple Machine Learning Research, Microsoft Research, Hacker News (AI posts with 100 or more points), Import AI, Simon Willison (AI posts), Latent Space, Interconnects, the AI Alignment Forum, LessWrong (AI), The Gradient and Ahead of AI.

We do not read social media.

• Titles that look like podcasts, webinars, newsletters, weekly round-ups, daily digests or job adverts are skipped.
• For tier 3 feeds and the blogs that publish a lot of how-to material (Hugging Face, AWS, Microsoft Research, Google Research), titles that look like tutorials, explainers, listicles or comparisons are skipped.
• A story whose title is almost the same as one we saw in the last 7 days is skipped or added to that event.
• At most 5 stories per feed per run (3 for tier 3) and 60 stories per run go any further.

A small, fast model (Claude Haiku) rates the title and the first 600 characters from 1 to 10. Opinion, tutorials, rehashed news and marketing should score low. Anything under 5 stops here.

If the feed only gave us a short summary, we fetch the article page. If we still have fewer than 500 characters, the story is not scored, because a teaser is not enough to judge. This happens most often with OpenAI's own news posts, which block our fetcher. The same event usually reaches us from an outlet whose full text we can read.

Claude Sonnet reads up to 8,000 characters of the article and writes down the event underneath it: a plain headline, what concretely changed, who is involved, the signal tags, whether there is primary evidence, and whether the piece is only commentary. Commentary with no evidence behind it is dropped.

We compare each new event with the events of the last 7 days that are dated within 72 hours of it. It counts as the same event if any of these is true:

• they share a primary evidence link;
• they share a named company, person or product and have similar headlines;
• a small language model that runs on our own server (all-MiniLM-L6-v2) rates the headline and description as very similar: 0.80 or more on a scale of 0 to 1, or 0.74 or more when both name the same numbered product first, such as a model version.

If the matching event is published, the new story is added to it as more coverage and is not scored again. If the matching event was not published, the new story gets one scoring attempt as an event of its own, because the first article may simply have been a poor one. After that, further write-ups are added to the best attempt.

The number of sources shown on each event is the number of stories attached to it. Several outlets rewriting one announcement each count once, so this number measures coverage, not independent confirmation.

Claude Opus scores the event against the rubric below. It sees the extracted event, the article title, the source and its tier, and the first 4,000 characters of the article. The score runs from 0 to 100.

Points for signal (up to 100)

• Evidence quality0 to 2015 to 20 when we are reading the primary document itself (the organisation's own announcement, the paper, the filing, the release notes) or a report that quotes it in detail. 10 to 14 for a news report with named sources. 0 to 9 for anonymous sources, a summary of someone else's report, a podcast or a social media post. A company's own post proves it announced something, not that its own benchmark or customer claims are true.
• Concreteness0 to 15Exact numbers, dates and specifications score high. Vague claims score low.
• Real-world impact0 to 20How much this changes what people can do or how markets work. An existing model arriving on one more platform, how-to guides, customer case studies and minor version updates get 0 to 8, however well documented.
• Falsifiability0 to 10Could the claim be checked and shown to be wrong? “AI will transform X” scores 0.
• Novelty0 to 10New information rather than recycled news.
• Actionability0 to 10Can someone make a decision based on it? A price change scores high.
• Longevity0 to 10Will it still matter in six months?
• Power shift0 to 5Does it change who has power or access?

Points off for noise (up to 25)

• Vagueness0 to 5Words like “revolutionary” or “could transform”.
• Speculation0 to 5Claims about the future without evidence.
• Packaging0 to 5PR framing that overstates what actually changed.
• Recycling0 to 5Rehashed news, or small updates presented as breakthroughs.
• Engagement bait0 to 5Clickbait, fear and hype.

Signum covers AI. If AI is absent from an event or only mentioned in passing, the model is told to score it 30 or less.

The final score is the lower of two numbers: the model's own overall score, and the points for signal minus the points off for noise.

• No concrete change identified: the score is capped at 25.
• No primary evidence: evidence quality is capped at 10.
• Commentary only: real-world impact is capped at 5.
• No signal tags: the score is capped at 20.

Confidence is then set by code, not by the model:

• High: primary evidence, evidence quality 15 or more and concreteness 10 or more.
• Medium: evidence quality 8 or more and concreteness 6 or more.
• Low: anything else.

• 60 or morePublished straight away if confidence is medium or high.
• Held back60 or more but low confidence, or 85 or more with no primary evidence. These are not shown and nobody reviews them.
• Under 60Rejected. Kept in our database, not shown.

Only published events get a written summary. At most 15 new events are scored on each run; the rest wait for the next run.

The front page shows the 60 most recently added published events from the last 60 days. By default they are ranked by 65% score, 25% freshness and 10% coverage.

• Freshness: 100 if under 6 hours old, 85 under a day, 70 under 3 days, 50 under a week, 30 after that.
• Coverage: rises with the number of stories attached (1 story 24, 2 stories 38, 3 stories 49, 5 stories 63, never more than 100).

You can also sort by score alone or by date.

• capabilityWhat AI systems can do
• economicsPrices, costs and markets
• distributionWho gets access (shown as Access)
• regulationLaws, policy and governance
• infrastructureCompute, chips and data centres
• adoptionUse in the real world
• labourJobs and work
• powerWho gains or loses influence

We also run a small statistical model (logistic regression on 15 features, using the same local sentence model) that estimates how likely an event is to be signal. It was trained in March 2026 on 166 events labelled by their AI scores, so it mostly learns to agree with the AI score. Its output is stored but does not change any score, ranking or publishing decision. It is retrained weekly and a new version replaces it only if it does better; none has since March.

The thumbs up and down buttons on event pages are recorded. They do not change scores. The weekly retraining reads them, but so far there are too few to make a difference.

• A score is an AI judgement, not a measurement. The same event written up by two outlets can score 10 or more points apart, so an event near 60 could have gone either way.
• Evidence links are only kept when the extraction step finds them in the article text, so many events show none.
• Summaries can contain mistakes. The linked sources are the record.

The signal and noise framing owes a lot to David William Silva's essay “What Is, Concretely, the AI Hype?”.