AI: what to read

Старший Авгур@senior_augur · About artificial intelligence research, language models and the math behind them, for people following big tech and developer news. As of 14 September: 8.1K subscribers, ER at 81.5%, well above the category median, though subscriber count sits below the category median and only 11 posts went out in 30 days, less often than others in the category. Recent topics include the story of mathematicians Tristan Buckmaster and Levent Alpöge, whose paths crossed with OpenAI and Anthropic, a breakdown of Flow Matching as an approach behind Stable Diffusion 3.5 and FLUX.2 image generation, the launch of the Needle search benchmark, and Web Query Language being renamed Keenable SELECT. The author also covers the FlyWire map of a fruit fly's neurons and security incidents involving language models such as Kimi K3.8,289 subssubs70B

Language models, benchmarks and research: new evaluation sets and the loud incidents around models, taken apart one by one. A small channel — 8 thousand subscribers since 2024 — with the highest engagement in this collection: a post is read by 57 % of subscribers. 10 posts a month, 8 of them about models and neural networks. Score 69.

AbstractDL@abstractdl · About artificial intelligence research and large language model news, for those tracking the race between OpenAI, Anthropic and other labs. The author breaks down model releases, tests their capabilities, and shares observations on how subscription limits and API pricing keep shifting. As of 14 September: subscribers stand at 18.4K, above the category median, while engagement holds at 68.9 percent, well above the category median of 24.0 percent. Posts come out less often than at other channels in the category, but each one digs into a specific case: the multi-agent mode in GPT-6-astra, the Codex context window expanding to 1M tokens, and the backlash over Anthropic's Max plan limits. A separate piece covered the Ouroboros architecture, a preprint written together with Hope and Roman Yampolskiy that landed on GitHub Trending and passed a thousand stars.18,664 subssubs86A

News on language models, autonomous agents and coding benchmarks, and in between — the author's progress on an agent framework of their own. Posts rarely: 8 a month, 5 of them about models. But it gets read: 18 thousand subscribers, 54 % of them open a post, score 83 — the highest in this collection. Running since 2021.

Время Валеры@cryptovalerii · Engineering notes on machine learning and agentic systems: design docs, system prompt compression, and practical work with LLM agents. The author shares experience building GraphQL agents, reviewing technical documentation, and is co-writing the book Agentic AI System Design with collaborators from an ML System Design course. As of 14 September: subscribers stand at 30.8K — well above the category median of 14K — while posts over 30 days number 6, less frequent than most channels in the category. ER sits at 53.6%, well above the category median of 24.0%. Recent topics include the open-source skill lossless-doc-compress for compressing documents without losing information, a breakdown of Shopify's approach to gisting, and an announcement of the Yandex Scale 2026 conference. The channel mixes technical breakdowns with personal stories, including an invite to a FuckUp Night event with Valery Babushkin and other speakers.31,243 subssubs74B

Engineering of AI agents and system design of products built on language models: prompt compression, reviews of technical documentation, work on a book about agents. 31 thousand subscribers since 2021, and 53 % of them read a post — second in this collection. 8 posts a month, 6 on topic. Score 81.

RnD ML Team@rndml_team · For those following research in machine learning and language technology: the team shares updates on its scientific papers, conferences and projects. As of 15 September: the channel has 4K subscribers — below the category median, while engagement stands at 45.3%, well above the category median of 24.6%; posts over 30 days numbered 11, less frequent than the category median. Over the past month, the team reported that three of its papers on sign language research got accepted to EMNLP 2026, released the MERA Text 2.0 text benchmark, and published STARM, an open-source framework for recursive reasoning models from AI Forever. Posts also covered work on Omnimodal Fusion, combining audio, text and visuals in a single latent space, plus topics like Google DeepMind's agentic video understanding and prompt injection risks in language models.4,131 subssubs63C

Research and development from an R&D team: new benchmarks, open frameworks for reasoning models, paper breakdowns with the methods. The smallest channel in this collection — 4 thousand subscribers since 2023 — and one of the most read: 40 % open a post. 13 posts a month, 8 of them about models. Score 67.

Институт AIRI@airi_research_institute · Research at Russia's AIRI Institute and the work of its scientists get covered here, alongside breakdowns of papers, seminar announcements, and staff appearances in media and podcasts. As of 15 September: subscribers stand at 13.6K, on par with the category median, while ER of 37.3% runs well above the category median of 24.6%, even as the 15 posts over 30 days come in below the category median. Over the past month the channel covered AIRI researchers presenting at the ECML PKDD conference in Naples and IJCAI in Bremen, a new way to test diagnostic medical models developed with the Almazov National Medical Research Centre, and recordings from the Summer with AIRI 2026 school in Ust-Labinsk. It also ran pieces on crafting effective prompts and the legal side of AI-generated content.13,700 subssubs61C

The scientific life of the AIRI institute: conferences, publications, interviews with researchers on applying artificial intelligence in medicine, archaeology and other fields. 14 thousand subscribers since 2021, and 36 % of them read a post. 15 posts a month, 12 about models and research. Score 70. The only research-institute channel in this collection.

Love. Death. Transformers.@lovedeathtransformers · About artificial intelligence research, language models and the everyday texture of the AI industry, told with irony and without smoothing over the details. The author covers new architectures, experiments and releases from major labs, mixing in personal observations and jokes about AI safety debates. As of 15 September: 25.1K subscribers, ER 36.7%, above the category median. In the last 30 days the channel published 56 posts, well ahead of the category pace, while its subscriber count also sits well above the category median. Recent topics include the open-source GigaChat 3.5 Reasoning model built on a 432B-A28B MoE architecture, a new Yandex model with 1:70 sparsity, and a breakdown of a hierarchical sparse KV-cache indexing scheme. The author also announced a talk at an applied machine learning conference for cybersecurity in London.25,698 subssubs56C

Breakdowns of new models and benchmarks, notes on robots, links to papers and repositories with the author's comments; the tone is informal. 25 thousand subscribers since 2020, and 35 % of them read a post. Posts often: 46 a month, 24 of them about models and neural networks. Score 70.

Denis Sexy IT 🤖@denissexy · Covers artificial intelligence and tech with a personal voice — breakdowns of stories like the Hugging Face hack, fresh model releases, and the author's own experiments with coding agents. The tone is casual and often funny, mixing serious AI analysis with odd finds like a fly brain map or YouTube thumbnail tricks. Good fit for readers who want AI news with commentary instead of a dry digest.137,865 subssubs69B

Artificial intelligence and technology: the loud stories taken apart, fresh model releases and the author's own experiments with coding agents. The largest channel in this collection — 137 thousand subscribers, running since 2016 — and still 32 % of them read a post. 54 posts a month, 27 about models and neural networks. Score 75.

эйай ньюз@ai_newz · About artificial intelligence news: breakdowns of fresh models, releases and industry debates. The author covers Sber's GigaChat 3.5 Reasoning with open weights and a 432B-A28B MoE architecture, Krea Agents' file-based system for organizing neural content, Tencent's AuK model for speech generation and editing, and compares GPT-Image-2.5's Flare and Sunburst versions on speed and quality. As of 13 September: subscribers 96.8K, median post views 30.3K, ER 31.3%, with 45 posts published over 30 days. Subscribers exceed the category median, posting outpaces category peers, and ER runs above the category median. Over the 20-day observation window, subscribers grew by 538 and average views rose by 9.2K.97,152 subssubs62C

Language model releases, tools for generating video and images, breakdowns of new chips and architectures — as short news items. 97 thousand subscribers since 2020, and 31 % of them read a post. 38 posts a month, 28 about models and neural networks — the highest on-topic share among the large channels here. Score 71.

Сиолошная@seeallochnaya · Covers artificial intelligence news: model releases, API pricing, rocket launches and what is happening across major AI labs. As of 30 September: 80K subscribers, well above the category median of 13.8K. The channel posted 62 times in 30 days, more often than the category median of 25 posts. Engagement sits at 34.0%, above the category median of 24.1%. Recent posts covered the Gemini 4 Argon announcement with its token pricing, details on a new $500 Grok Bot and Muse tier, a Reuters report on Anthropic's net loss ahead of its IPO, and coverage of the first orbital Starship flight carrying Starlink V3 satellites. Other posts touched on China restricting foreign travel for AI specialists and podcasts explaining AI safety to a general audience.80,108 subssubs69B

News of the artificial intelligence industry: incidents with models, the economics of the boom, regulation and where the big labs stand. Long texts rather than links. 79 thousand subscribers since 2023, and 29 % of them read a post. 45 posts a month, 29 about models and neural networks. Score 68.

AI Product | Igor Akimov@ai_product · Covering news and updates in artificial intelligence: model releases, new features from major AI labs, and what these changes mean in practice. The author writes in first person, explaining updates in plain language and comparing models against each other. As of 13 September: subscribers stand at 8.1K, below the category median. Posts over 30 days reached 29, more frequent than the category median, while ER sits at 31.4%, above the category median. Over the past month the channel covered OpenAI's Agents API and Data agent, the GPT-Live-1 voice model, DeepSeek's V4.1 Flash release, and Apple's Siri AI update. It also discussed GPT-6 Astra and its Terminal-Bench 4.0 results compared with Claude Fable 5.1.8,255 subssubs57C

Fresh releases and news: from new language and video models to chips and hardware for running them locally. The author looks at how a model becomes a product. 8 thousand subscribers since 2023, and 29 % of them read a post. 26 posts a month, 22 on topic — denser than most channels in this collection. Score 60.