A crawl of 200,434 Telegram nodes, narrowed after verification to 728 physics channels and groups. Each one's post archive was kept — 1,578,436 messages, roughly 455 MB of text spanning 2015 to 2026. Every description was written from posts sampled evenly across a channel's whole lifetime rather than its recent output: a channel that changed direction would otherwise be described as whatever it happened to post last week.
The original labelling could not be reused. Its "STEM" and "physics" flags were true for all 728 entries; "olympiad" and "physics olympiad" covered exactly the same rows, as did "research" and "physics research". Different tags returned identical lists — precisely the failure a catalog like this exists to avoid.
So the sections were rebuilt: the model described what each channel is for in its own words, those descriptions were grouped, and a section was published only if enough entries fell into it. Sections below that floor were merged or dropped.
It counts how many other catalog members link to that entry. A link means the channel's bio or its recent visible posts reference another channel, or that Telegram's own "similar channels" list names it — so part of this graph reflects Telegram's recommendation algorithm rather than anyone's editorial judgement. That is worth knowing when reading the number.
PageRank over the same graph is also computed and used as a secondary sort. It is not the headline number: PageRank sums to one, so every score falls as the catalog grows even when nothing about the link structure changes. A plain citation count behaves predictably.
Not by citations alone, which would put the largest institutional accounts at the top of every section. The order also accounts for how distinctive an entry is to its section, and for level and language: this site is read mostly from Russia, Kazakhstan, Belarus, Ukraine, Uzbekistan and Kyrgyzstan. Entries in other languages remain in the catalog and are reachable through the language filter — they are simply not the default.
Eight pseudoscience entries, including a 20,300-subscriber channel predicting earthquakes from planetary alignment; three broken ones; and job boards, funding notices, immigration consulting and personal brand channels — anything that is not science. Mathematics and informatics were kept, in their own sections. Private invite links are never published.
Entries judged to be general-audience are held separately so they do not sit beside topic-specific sections.
Relevance judgements come from a language model, not from human annotation. Selection accuracy was measured on a sample, not across the whole set. Some channels carry too little text to classify confidently and are marked "thin data".
| Channels described | 718 |
| Published | 600 |
| Removed as not science | 82 |
| Removed as not physics education | 36 |
| Flagged general-audience | 68 |
Snapshot 2026-08-02.