AI Daily/How it works

How AI Daily works

Every weekday at 17:10 Pacific a machine reads the AI industry and publishes three things worth knowing. The whole filter fits in one picture.

Why three days and not twenty-four hours

Research labs and independent analysts do not publish daily. Give it a one-day window and most of them are empty, so the issue collapses into whatever the news sites ran that morning. That is the noise this exists to filter. Three days is the shortest span where the people who actually make the news reliably have something to say.

The full source list, with the group each one sits in, is published here, six broken ones included.

Why the group a source sits in matters

Nothing gets published unless at least one source can be quoted for it. Every quote is copied word for word, never paraphrased, and tagged with the job it does. A company announcing its own work is strong evidence that something happened. It is weak evidence that it matters.

• Company the organisation announcing its own work
• Independent research and analysts with nothing to sell you
• Press forums, repos and the trade press
• Slow lane podcasts and long-form

This is why a trend tells you what it rests on, in the form well sourced, 1 company, 3 independent, instead of showing a confidence score. You can look at the mix and decide for yourself.

When it breaks, it says so

Three gates can kill an issue. If the sources return nothing, the run stops instead of publishing an empty issue. From the outside a genuinely quiet window and a broken connection look identical, and guessing between them once produced an issue with invented content in it. If nothing significant turns up, it skips. If the writing step fails, the issue is withheld rather than sent half empty.

Days that produced nothing stay in the archive with the reason attached.

Each issue also carries a short Before you act on this note. These are things the writing step flagged about its own output: an unconfirmed deal value, a lawsuit allegation that should not be read as settled fact. Roughly six issues in seven have one, and the contents are different every time.

What it will not do

• It does not rank by importance. Ordering comes from how recent something is and which group the source sits in. There is no model of what matters.
• Roughly half of a day's subjects never reach the writing step. That is the second pale block in the picture. If something is missing, that is weak evidence at best.
• The source list is one person's judgement. Mine. It leans English-language and toward the engineering and product conversation. If a source is not on it, that means I did not pick it, not that I assessed it and said no.
• "Well sourced" is not "true". It describes what a claim rests on and nothing else.
• The same day cannot be re-run to give the same issue. The writing step is not deterministic. What is fixed and published is the input and the instructions, not the output.
• Three a day is a hard cap, chosen so this stays readable. A busy day still publishes three.
• A source that went quiet and a source that broke look the same from the outside. Total blackouts get caught. One quietly dead source can go unnoticed for a few days.
• The numbers describe this pipeline, not the industry. They say what these 93 sources published on this schedule, and nothing wider.

These eight are the canonical list. A technical write-up of the scoring, the de-duplication and the evidence model is not published yet; when it is, it will say the same eight things in the pipeline's own vocabulary.

Using it

Every issue is published as data too: a file per day plus a feed, so you can read it in a reader or hand it to your own tools. Get it as data →

MIT licensed. Use it, copy it, modify it, redistribute it, build a product on it. The only condition is that the copyright notice travels with it. Full terms →