About the Observatory
The Observatory has three desks: AI, Ethiopia and Interesting Politics. Between them they follow subjects that move every day: AI in management and in software engineering, the research behind it, Ethiopia, and how some governments speak to their own people. It reads its sources the way a wire service would, keeps what falls within each desk’s beat, traces every story to where it first appeared, and publishes what it found each morning.
Every step is done by software following written editorial rules: what is in scope, what matters most, how a story is checked and how an issue is written. The rules are changed by hand, and every version is kept.
Video · How the Observatory works
01 · Collecting
6,627
items read from 575 sources
9 readers, one for each way a source can be read: feeds, pages without feeds, X accounts, newsletters, Telegram channels, WeChat, model hubs and paper archives, podcasts and video. Each source is read on its own rhythm.
79 scans in the last 24 hours
02 · Languages
5
source languages
Items in Chinese and Amharic are translated before they are judged, so a Chinese lab or an Ethiopian ministry is held to the same standard as a wire service.
English, 中文, አማርኛ, French, Russian
03 · Relevance
3,748
of 5,337 new items kept
Every new item meets its desk’s written scope rule. On X posts a second model looks again at everything the first dropped.
Each model benchmarked on our own material
04 · Editing
138
notes written
Accounts of the same event are compared and the best is chosen, given a type and tags, and written up as a note.
Ranked by each desk’s written priorities
05 · Checking
{labeled}
of 1,025 notes this week passed with a label stating their origin
Each note is traced to where it first appeared and to who stands behind it. Outlets repeating a story do not confirm it. We check origin and credibility, not facts.
Telegram and X posts count as their own originals
The archive
2,403
notes, and 115 issues, kept for good
Every note with its sources, and every issue sent. People here would rather read the originals, so the Book is not shown on this site; it is kept as the Observatory’s record.
Every version of every rule kept with it
06 · Publishing
12
issues sent in the last 24 hours
Each publication chooses its stories from the Book, groups them, writes the issue and looks it over.
8 publications
07 · Readers
1
copy of each issue per reader, at the hour they choose
By email or on Telegram, in English, Russian or Amharic. Nobody receives the same issue twice.
Email, Telegram, LinkedIn, YouTube and this site
08 · Control, over every room
20 benchmarks
No model is used because it is famous: each step has its own test set drawn from our material, and a model earns the step by score, cost and speed. Spending limits for each desk, and a morning report on anything that went wrong.
One day of work, drawn to scale: everything read in the last 24 hours, what was new, what fell within a desk’s scope, and what became a note in the Book.
Each step has a test set drawn from the Observatory’s own material, with right answers written and read again independently. A model is given a step for its score, its cost and its speed, and is tested again when a new one appears.
| Step | Chosen | Measure | Result |
|---|---|---|---|
| Relevance · X posts | DeepSeek V4.1 Flash + Jev | relevant posts lost | 2.1% |
| Relevance · feeds and pages | Jev | relevant items lost | 0.2% |
| Relevance · Chinese | DeepSeek V4.1 Flash | relevant items lost | 2.1% |
| Relevance · Ethiopia desk | DeepSeek V4.1 Flash | relevant items lost | 0.3% |
| Credibility check | DeepSeek V4.1 Flash | right outcome | 90.6% |
| Choosing the day’s stories | MiMo V2.6 Flash | overlap with the editors’ choice | 71.2% |
The subjects of the notes written in the last 7 days: how many notes, and what share of the desk’s notes, with the change from the 7 days before. A note can carry more than one subject.
This desk writes no notes: these are the themes of the stories in its issues of Dispatches from Others, how many stories, and what share of them.