One issue a day, at 07:30. Below is what the system does before that.
The digest does not retell a feed. Of everything that went past the sources in a day, a few topics make the issue: the ones with a primary source under them that change something in practice.
The system walks X lists, RSS, Hacker News, lab blogs (OpenAI, Anthropic, DeepMind) and newsletters. About 40 sources per run, and I pick what is on that list.
Rumours, reposts and marketing are dropped by rules I set. What stays has a primary source under it: a lab post, a paper, a repository, a document.
The system opens every link and compares it with the text of the issue. A claim that does not match what the source says never reaches the issue.
The issue is written in Ukrainian and translated into English. Numbers and links stay as the source has them, with no rounding and no retelling.
The finished text becomes a podcast, the same selection for listening instead of reading.
At 07:30 the issue, the PDF and the podcast appear on the site.
The list I picked, and what step 1 walks. It is open: every topic in an issue can be checked against where it came from.
Lists I put together by hand: researchers, founders, engineers, around 190 follows. The main feed goes unread: ads and reprints. Below are the 40 accounts that made it into at least three issues, by frequency. Across all issues 181 accounts are cited.
Independent write-ups and newsletters. A layer on top of X: here an event gets taken apart on the merits, with numbers and consequences.
Tech desks. Needed where an event leaves the industry: courts, regulation, money.
Stories from the last day that passed 80 points. Shows what engineers are actually arguing about, comments included.
Blogs and releases from the labs and projects themselves. An announcement is read where it came out, in its original form.
The rule I set: a claim goes into an issue when a primary source backs it, or when two independent sources confirm it. One post is not enough; where there is a single source, the issue says so.
An X account, a blog, an RSS feed, a newsletter.
Every issue is text on the site, a PDF with clickable sources and a podcast. Links in the text go to the primary source. Any issue in the archive can be checked on that.