7 measurements spanning 30 days, net +161. Dots are measurements; the straight line between them is drawn to join them, not to claim we know the path taken in between — snapshots are recorded only when a count changes, so gaps mean “no change observed”, never “interpolated”. The vertical axis spans 758–967 and does not start at zero.
Measurement log — every subscribers count we have recorded
Measured (UTC)
Subscribers
Change
5 Sept 2026, 11:56
943
+18
28 Aug 2026, 22:58
925
+75
22 Aug 2026, 06:48
850
+55
14 Aug 2026, 08:36
795
+12
7 Aug 2026, 03:14
783
+1
6 Aug 2026, 15:35
782
no change
6 Aug 2026, 03:52
782
first reading
Engagement
18 posts held, back to 15 June 2026 — the reader has not yet reached the start of this channel’s public history, so older posts may sit further back, unread. Read across 1 page of Telegram’s post history, 20 posts per page.
Nothing published in the last 30 days. ERR and ER are rolling 30-day measures, so there is nothing to compute — we hold 18 posts for this entry, the most recent from 5 August 2026. An engagement rate over an empty window would be a number about nothing.
Reaction mix
377 reactions across 17 posts, in 17 distinct kinds. The most used accounts for 45.9% of them.
Every reaction kind recorded on the sample, most used first
Reaction
Count
Share
Share, drawn
❤
173
45.9%
🔥
50
13.3%
💯
42
11.1%
👍
41
10.9%
😱
21
5.57%
🥰
14
3.71%
👏
12
3.18%
👌
4
1.06%
✍
3
0.796%
🏆
3
0.796%
💔
3
0.796%
🙏
3
0.796%
⚡
2
0.531%
🍾
2
0.531%
🤓
2
0.531%
👨💻
1
0.265%
🦄
1
0.265%
No sentiment is inferred, and none should be read in. This table is ordered by count and by nothing else. Emoji do not carry stable meaning across languages or communities — 🙏 is thanks in one channel and mourning in another — so we publish which ones were pressed and how often, and pass no judgement on what an audience meant by them.
Precision. Telegram publishes reaction counts per emoji and short-forms each one — 4.34K, 1.2M — so any single kind at or above 1,000 reaches us at three significant figures, and only counts below 1,000 are exact. The shares above are ratios of those figures and carry the same error. This is also why the total here can differ slightly from a reaction total printed elsewhere on the page: both are sums of the same rounded parts, taken over samples with different edges.
Coverage. Reactions were read on 17 of the 18 sampled posts in this sample. Summed by Telegram’s own count on each post — not by adding up the per-emoji breakdown above — those same posts carry 377 reactions in total: the kind of figure the paragraph above means by “a reaction total printed elsewhere on the page”.
Measured over the 18 most recent posts we hold, published 15 June 2026 to 5 August 2026, using the newest reading held for each. Telegram Stars are excluded: they are a payment, not a reaction, and they have their own section.
Telegram Stars
Stars received
553
across the posts below
Posts paid on
17
of 18 we hold a reading for · 94%
Most on one post
397
single highest reading
A paid reaction is a reader spending Telegram Stars — bought with money — on a post by @okmlai. Telegram publishes the count on the public post preview alongside ordinary reactions, and this register reads it there. It is the only figure on this site that measures money moving rather than attention.
Stars are not reactions, and the two are never added. They are rendered in the same strip on Telegram and counted in the same shape, but one is a tap and the other is a purchase. The reaction totals and the engagement rate elsewhere on this page exclude every figure in this section, and no rate here is computed against a reaction count.
This is not revenue, and we publish no currency figure. What a Star costs a reader and what it pays a channel are different numbers, Telegram takes a share we cannot observe, and the terms have changed. Converting a Star count into money would be an estimate dressed as a measurement, so the count is where we stop.
Counted over the 18 most recent posts we hold for this entry, published 15 June 2026 to 5 August 2026. Star counts above 1,000 reach us in Telegram’s short form and carry the same three-significant-figure rounding as everything else on this page.
Из одного большого фреймворка — в экосистему специализированных библиотек. Рассмотрим NeMo от NVIDIA
NeMo по умолчанию воспринимается как фреймворк для обучения LLM, мультимодальных и speech-моделей, а теперь NVIDIA делает ставку на модульную архитектуру (из-за монолитности там было тяжело что-то найти, да и контейнер разросся). Реструктуризация анонсирована официально, исходный репозиторий NeMo 2.0 теперь сфокусиро…
Первый опыт запуска локальной LLM на Jetson Orin Nano
Проверим, насколько современные edge-устройства готовы к локальному запуску языковых моделей. В качестве платформы — Jetson Orin Nano Developer Kit 8GB, в качестве модели — Qwen2.5-3B-Instruct через llama.cpp.
😼 В итоге модель успешно заработала локально на GPU, но по пути встретилось несколько нюансов, о которых стоит знать заранее, тк придется немного 🤔.
Посл…
Математика категорий и ИИ
Любишь читать академичные лонгриды с сомнительной практической пользой? Тогда этот пост для тебя!
О теории категорий обычно говорят как об одной из самых абстрактных областей математики. Довелось прочитать популярную книгу «Восторг абстрактной математики» Юджении Ченг (вслух тебе её прочитают на ютубе, можешь купить на озоне за 4к и в целом за год достаточно прочитать только ее, чтоб собой…
Мониторить агентов — хорошая идея. Но монитор тоже придётся мониторить
Работа Preventing Rogue Agents Improves Multi-Agent Collaboration предлагает вместо того, чтобы улучшать самих агентов, добавить над ними отдельный слой runtime-мониторинга, который отслеживает признаки того, что один из агентов начинает вести себя девиантно.😐
Механизм простой и интересный. Монитор замечает, что агент начинает путаться 😬, и сист…
Эпоха AI vs AI
Hugging Face рассказали о первом известном случае полностью автономного AI-взлома.
Точка входа — классическая для AI-платформ. Вредоносный датасет эксплуатировал два пути выполнения кода в пайплайне обработки данных (remote-code loader и template-injection в конфигурации) и запустился на процессинг-воркере. Дальше шла эскалация до уровня ноды, сбор облачных и кластерных учёток и боковое перемещение п…
С чего начать изучение AI Security?
Одна из многих проблем LLM — prompt injection. Но свет на ней клином не сошелся! Это лишь один из множества классов атак. 😲
Чтобы разобраться в ландшафте угроз, рекомендую начать с OWASP Top 10 for LLM Applications — самой известной единственной таксономии рисков для GenAI-приложений. Отличная точка старта.
🙂 При этом важно понимать, что это не полный список угроз. Документ в пер…
Почему внимание — не всегда лучший критерий?
Исследователи из LUMIA Lab (SJTU) и Edinburgh предложили InfoKV — фреймворк сжатия KV-кэша, который смотрит в будущее! Это как вообще? 🎂
Существующие методы (SnapKV, PyramidKV и др.) выбирают токены по весам внимания, то есть по тому, насколько недавние токены смотрят на прошлые. Это работает для ближнего контекста, но в длинном ризонинге токены важны и для будущих шаго…
Как измерить качество сгенерированного текста?
Конечно, кажется, что можно использовать только метрики, пришедшие из машинного перевода (например, BLEU - считает n-граммы от 1 до 4 слов (униграммы, биграммы, триграммы, 4-граммы) плюс штраф за краткость, ссылка на интересную статью про это).
Кстати, только с 2010 по 2020 было предложено 100+ новых метрик — все мы тут не рассмотрим. Но в эпоху LLM простого сравнени…
Фейковый баг-репорт угоняет ИИ-агентов для кодинга
Исследователи из Tenet Security показали новый класс атак на кодинг-агентов. Суть в том, что атакующий может заставить ИИ-агента выполнить произвольный код на машине разработчика, просто подбросив один фейковый отчёт об ошибке в публичный баг-трекер. 💡
Атака построена вокруг Sentry — популярного сервиса для отслеживания ошибок и мониторинга приложений. Исследовате…
Похоже, в AI вайбкодинге появился новый модный термин! Сегодня про Harness Engineering 🤲
Команда Nexu опубликовала открытый Harness Engineering Guide — практическое руководство по созданию среды выполнения (harness) для AI-агентов.
Если очень упростить, то LLM — это мозг, а harness — всё остальное, что делает агента агентом действительно автономным:
🥹 управление тулами, mcp;
🥹 сбор и суммаризация контекста;
🥹 п…
❤12🔥6💯4👏2💔2🤓2🥰2
Showing the 12 most recent of 18 posts we hold for @okmlai. View and reaction counts are the latest single reading for each post, not a live figure, and a recent post is still accumulating both. A view count marked ≈ was rounded by Telegram before we ever saw it — t.me prints views in full below 1,000 and to three significant figures above, so ≈1,200,000 means somewhere between 1,150,000 and 1,249,999. Unmarked counts are exact. Text is reproduced from the public post preview and truncated for length.
Stars beside a post are paid reactions — Telegram Stars, bought with money and spent on that post. They are a different unit from reactions and are never added to them, here or anywhere else on this page.
Polls
The poll we hold for this entry, as Telegram rendered it when we read the post. A poll’s figures keep moving after that, so each one is dated.
Shares as published. No per-option vote count is published by Telegram, so none is shown.
Percentages only — there are no per-option vote counts here, because Telegram publishes none. The public post preview gives each option’s share and a single voter total, and nothing else. Multiplying one by the other would produce a per-option tally that looks measured and is not: the shares are rounded to whole numbers before we ever see them. We print what was published and leave the column that does not exist empty.
The shares need not add up to 100. Rounding alone puts many polls at 99 or 101. A poll that allows more than one answer per voter runs well past 100 by design, and several here do. The bars are drawn against a fixed 100% track at each option’s own percentage rather than normalised to the total, so a poll that exceeds it shows that it does instead of being quietly rescaled.
Read from the 18 most recent posts we hold, published 15 June 2026 to 5 August 2026. Telegram labels each poll by kind — an anonymous poll, a quiz, a closed set of final results — and that label is reproduced rather than paraphrased.
Forward network
Republished by
Channels on the register that have forwarded this channel's posts into their own feed.
Built only from forwarded posts we have actually read, on both sides. Coverage is early and deliberately incomplete: a missing link means we have not read the post that would prove it, never that the relationship does not exist. Counts are distinct forwarded posts observed, so they only ever go up as we read more.
Mentions
Named by 2 registered channels — every channel on the register whose own posts have named this one, by its current username or any other username it currently holds, merged from two separately captured readings of the same fact so a namer caught by only one of them is not missed and a namer both caught is not counted twice. A username this channel has since dropped is not matched — that handle may belong to someone else now, and crediting today’s namer to yesterday’s owner would misattribute it.
Named by
Channels on the register whose posts name this channel's handle.
A mention is a weaker signal than a forward and is counted separately for that reason — naming a channel is not republishing it, and a handle in a post body is easy to place deliberately. The post counts beside each row below are distinct posts in which the handle appeared, from posts we have read on both sides — the “Named by N registered channels” figure above is a different count, of distinct NAMING CHANNELS rather than posts, and is not the sum of the rows under it.
Cite this entry
A live page changes as we take new readings, so a citation should name the measurement it is based on, not just the URL. The line below cites the subscriber count as measured 5 September 2026 — this
entry's latest reading, not the date you are reading this.
“OK ML” (@okmlai), 943 subscribers as measured 5 September 2026. Telegram Register, tgregister.com/channel/okmlai.
Full measurement history, CC BY 4.0. Every reading this register holds for this entry, not just the latest one, as a dated, downloadable record: CSV · JSON. Free to use with attribution to tgregister.com. Each file carries its own generation timestamp, which is the figure to cite for exactly when the data was retrieved.
