{"source":"https://news.hamidun.com","license":"CC BY 4.0","total":20,"articles":[{"id":14850,"title":"Wired за сутки протестировал 6 ИИ-агентов: Dots лидирует, Gemini Spark провалил обед","title_en":"I Tested Every Major AI Agent for a Day. Here’s the Best One","preview":"Репортёр Wired настроил на iPhone шесть always-on агентов и прогнал их через пять заданий. После трёх тестов Dots набирает 12 баллов, Instinct — 5, а Gemini Spark заказал боссу один ролл.","categories":"llm,business","category":"llm","image":"https://news.hamidun.com/images/news_14850_1793219633517.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14850/wired-za-sutki-protestiroval-6-ii-agentov-dots-lidiruet-gemi","url_md":"https://news.hamidun.com/ru/news/14850.md","source":{"name":"Wired","url":"https://www.wired.com/story/i-tested-every-major-ai-agent-for-a-day-heres-the-best-one/"},"published":"2026-10-28 20:33:36","publish_at":"2026-10-28 20:33:36","processed_at":"2026-10-28 20:33:36","word_count":692,"reading_time":4},{"id":14851,"title":"Anthropic подала заявку на IPO: выручка $16,3 млрд за полгода, тикер ANT","title_en":"Anthropic officially files for its public debut","preview":"Anthropic раскрыла S-1: выручка за полугодие выросла в 13 раз до $16,3 млрд, операционный убыток сократился до $693 млн, а чистый убыток составил $80,7 млрд.","categories":"llm,business","category":"llm","image":"https://news.hamidun.com/images/news_14851_1793219687525.jpg","audio_path":null,"quality_score":9,"url":"https://news.hamidun.com/ru/news/14851/anthropic-podala-zayavku-na-ipo-vyruchka-16-3-mlrd-za-polgod","url_md":"https://news.hamidun.com/ru/news/14851.md","source":{"name":"TechCrunch","url":"https://techcrunch.com/2026/10/26/anthropic-officially-files-for-its-public-debut/"},"published":"2026-10-28 20:33:25","publish_at":"2026-10-28 20:33:25","processed_at":"2026-10-28 20:33:25","word_count":608,"reading_time":4},{"id":14849,"title":"YADRO запретила LLM трогать цифры и выпускает 10+ отчётов в день без людей","title_en":"Код считает, модель описывает: как мы встроили LLM в отчетность и не дали ей трогать цифры","preview":"Аналитики YADRO за три месяца собрали сервис на Dify, Trino и Airflow: код считает метрики, модель пишет комментарии с метками, и каждое утро выходят больше десятка отчётов.","categories":"llm,business","category":"llm","image":"https://news.hamidun.com/images/news_14849_1793219577746.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14849/yadro-zapretila-llm-trogat-tsifry-i-vypuskaet-10-otchetov-v","url_md":"https://news.hamidun.com/ru/news/14849.md","source":{"name":"Habr AI","url":"https://habr.com/ru/companies/yadro/articles/1099914/?utm_campaign=1099914&utm_source=habrahabr&utm_medium=rss"},"published":"2026-10-28 20:32:39","publish_at":"2026-10-28 20:32:39","processed_at":"2026-10-28 20:32:39","word_count":739,"reading_time":5},{"id":14848,"title":"Anthropic раскрыла, как собирает Claude руками Claude: мозг отдельно от рук","title_en":"How Anthropic ships Claude with Claude: Separating the Brain from the Hands","preview":"Главы инженерии и продукта Anthropic рассказали, как устроены Opus 5.5 и Managed Agents. Выручка на сотрудника выросла с $1,4 млн до $6,6 млн за год.","categories":"llm,business","category":"llm","image":"https://news.hamidun.com/images/news_14848_1793219527318.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14848/anthropic-raskryla-kak-sobiraet-claude-rukami-claude-mozg-ot","url_md":"https://news.hamidun.com/ru/news/14848.md","source":{"name":"Latent Space (swyx)","url":"https://latent.space/p/opus55/"},"published":"2026-10-28 20:30:55","publish_at":"2026-10-28 20:30:55","processed_at":"2026-10-28 20:30:55","word_count":651,"reading_time":4},{"id":14847,"title":"ИИ-агенты за 24 дня формализовали 15 000 страниц доказательств и нашли дыру","title_en":"数十agent，24天挑战数学界巨型工程，还揪出40年前漏洞","preview":"Несколько десятков ИИ-агентов за 24 дня превратили многолетний труд математиков в 43 млн строк проверяемого кода и обнаружили логический пробел возрастом почти 40 лет.","categories":"llm,research","category":"llm","image":"https://news.hamidun.com/images/news_14847_1793219420778.jpg","audio_path":null,"quality_score":9,"url":"https://news.hamidun.com/ru/news/14847/ii-agenty-za-24-dnya-formalizovali-15-000-stranits-dokazatel","url_md":"https://news.hamidun.com/ru/news/14847.md","source":{"name":"Jiqizhixin (机器之心)","url":"https://www.jiqizhixin.com/articles/2026-10-28-6"},"published":"2026-10-28 20:29:22","publish_at":"2026-10-28 20:29:22","processed_at":"2026-10-28 20:29:22","word_count":507,"reading_time":3},{"id":14844,"title":"Anthropic превращает Claude Code в платформу: Claude Mods переворачивают обвязку агента","title_en":"Claude Mods: Flipping the Harness Inside Out","preview":"Моды для Claude Code на TypeScript перехватывают промпты и вызовы инструментов. Автор протестировал их на практике: супервизор сработал в 99 из 111 ходов.","categories":"llm","category":"llm","image":"https://news.hamidun.com/images/news_14844_1793212443865.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14844/anthropic-prevrashchaet-claude-code-v-platformu-claude-mods","url_md":"https://news.hamidun.com/ru/news/14844.md","source":{"name":"Latent Space (swyx)","url":"https://latent.space/p/claude-mods/"},"published":"2026-10-28 18:32:35","publish_at":"2026-10-28 18:32:35","processed_at":"2026-10-28 18:32:35","word_count":747,"reading_time":5},{"id":14843,"title":"Nemotron 3.5 Bolt сжали до 24 ГБ: 74B-модель выдаёт 84 токена в секунду","title_en":"Nemotron 3.5 Bolt 74B-A9B 24GB: A Hybrid Vision-Language Model for Local AI","preview":"Команда из 109 авторов ужала Nemotron 3.5 Super VL 120B до 74B параметров и 24 ГБ памяти. На RTX Spark модель выдаёт 84 токена в секунду и держит агентные задачи.","categories":"llm,hardware","category":"llm","image":"https://news.hamidun.com/images/news_14843_1793212397624.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14843/nemotron-3-5-bolt-szhali-do-24-gb-74b-model-vydaet-84-tokena","url_md":"https://news.hamidun.com/ru/news/14843.md","source":{"name":"arXiv cs.CV","url":"https://arxiv.org/abs/2610.33039"},"published":"2026-10-28 18:32:34","publish_at":"2026-10-28 18:32:34","processed_at":"2026-10-28 18:32:34","word_count":576,"reading_time":4},{"id":14842,"title":"ИБ-инженер за 7 дней собрал MCP-сервер, который пугает ИИ-агента уязвимостями","title_en":"Вызываем паранойю у ИИ-агентов: заменит ли MCP AppSec?","preview":"AppSecGate MCP запускает сканеры и ведёт базу файндингов. На OWASP Juice Shop агент за 50 минут и 15 рублей разобрал 200 находок и исправил 56, включая 8 critical.","categories":"llm,business","category":"llm","image":"https://news.hamidun.com/images/news_14842_1793212283733.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14842/ib-inzhener-za-7-dney-sobral-mcp-server-kotoryy-pugaet-ii-ag","url_md":"https://news.hamidun.com/ru/news/14842.md","source":{"name":"Habr AI","url":"https://habr.com/ru/articles/1099840/?utm_campaign=1099840&utm_source=habrahabr&utm_medium=rss"},"published":"2026-10-28 18:29:44","publish_at":"2026-10-28 18:29:44","processed_at":"2026-10-28 18:29:44","word_count":636,"reading_time":4},{"id":14841,"title":"Bayer запустила мультиагентный движок на AgentCore: 80% вопросов без инженеров","title_en":"How Bayer Crop Science built an agentic decision engine on Amazon Bedrock AgentCore with AWS Unified Operations","preview":"Bayer Crop Science построила ADE на Amazon Bedrock AgentCore: 22 агента закрывают около 80% вопросов без инженера, а диагностика сбоев ускорилась с часа до 10 минут.","categories":"llm,business","category":"llm","image":"https://news.hamidun.com/images/news_14841_1793212165480.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14841/bayer-zapustila-multiagentnyy-dvizhok-na-agentcore-80-vopros","url_md":"https://news.hamidun.com/ru/news/14841.md","source":{"name":"AWS Machine Learning Blog","url":"https://aws.amazon.com/blogs/machine-learning/how-bayer-crop-science-built-an-agentic-decision-engine-on-amazon-bedrock-agentcore-with-aws-unified-operations/"},"published":"2026-10-28 18:28:35","publish_at":"2026-10-28 18:28:35","processed_at":"2026-10-28 18:28:35","word_count":739,"reading_time":5},{"id":14840,"title":"Liquid AI выпустила LFM2.5-24B-A2B: агент с 2B активных параметров","title_en":"Run fast, private agents on-device with LFM2.5-24B-A2B","preview":"Liquid AI открыла веса MoE-модели на 24B параметров, из которых активны всего 2B. На ноутбуке с M5 Max она выдаёт 184 ток/с, на одной B300 — свыше 21 тыс. ток/с.","categories":"llm","category":"llm","image":"https://news.hamidun.com/images/news_14840_1793212003150.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14840/liquid-ai-vypustila-lfm2-5-24b-a2b-agent-s-2b-aktivnykh-para","url_md":"https://news.hamidun.com/ru/news/14840.md","source":{"name":"Hugging Face Blog","url":"https://huggingface.co/blog/LiquidAI/lfm25-24b-a2b"},"published":"2026-10-28 18:25:04","publish_at":"2026-10-28 18:25:04","processed_at":"2026-10-28 18:25:04","word_count":610,"reading_time":4},{"id":14838,"title":"PyTorch 2.15 ускорил MoE-модели на Mac в 5–6,5 раза и принёс FP8","title_en":"MPS in PyTorch 2.15","preview":"Команда Hugging Face переписала критичные операции MPS на Metal: Qwen3.6-35B-A3B на M3 Ultra выросла с 9,4 до 61,3 токенов/с, а пик памяти Wan2.2 упал с 270 до 8 GiB.","categories":"llm,hardware","category":"llm","image":"https://news.hamidun.com/images/news_14838_1793205200880.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14838/pytorch-2-15-uskoril-moe-modeli-na-mac-v-5-6-5-raza-i-prines","url_md":"https://news.hamidun.com/ru/news/14838.md","source":{"name":"Hugging Face Blog","url":"https://huggingface.co/blog/mps-pytorch-2-15"},"published":"2026-10-28 16:31:36","publish_at":"2026-10-28 16:31:36","processed_at":"2026-10-28 16:31:36","word_count":824,"reading_time":5},{"id":14837,"title":"AgentOrch сравнил два типа агентов: разрыв Global и Step-by-Step сократился на 79%","title_en":"AgentOrch: How Close Can Global Orchestration Come to Step-by-Step Interaction?","preview":"Бенчмарк на 650 корпоративных SOP и 3220 вариантах показал: заранее собранный workflow сильно приблизился к пошаговым агентам по точности, но пока их не догнал.","categories":"llm,research","category":"llm","image":"https://news.hamidun.com/images/news_14837_1793205117950.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14837/agentorch-sravnil-dva-tipa-agentov-razryv-global-i-step-by-s","url_md":"https://news.hamidun.com/ru/news/14837.md","source":{"name":"arXiv cs.AI","url":"https://arxiv.org/abs/2610.33375"},"published":"2026-10-28 16:31:01","publish_at":"2026-10-28 16:31:01","processed_at":"2026-10-28 16:31:01","word_count":559,"reading_time":4},{"id":14836,"title":"Anthropic подала S-1 на IPO с оценкой $2 трлн: маржа 85%, Q2 в плюсе","title_en":"[AINews] Anthropic files for $2T IPO in 2 weeks: 1400% growth rate, 48% of revenue through cloud partners, 85% margin, $1B op profit in Q2, $134B cash in bank and $10B lifetime burn","preview":"Anthropic раскрыла финансы в S-1: выручка за Q2 $11,6 млрд (+1368%), операционная прибыль $974 млн, валовая маржа 85%. Тикер ANT, листинг на Nasdaq.","categories":"llm,business","category":"llm","image":"https://news.hamidun.com/images/news_14836_1793205064549.jpg","audio_path":null,"quality_score":9,"url":"https://news.hamidun.com/ru/news/14836/anthropic-podala-s-1-na-ipo-s-otsenkoy-2-trln-marzha-85-q2-v","url_md":"https://news.hamidun.com/ru/news/14836.md","source":{"name":"Latent Space (swyx)","url":"https://latent.space/p/ainews-anthropic-s1/"},"published":"2026-10-28 16:30:53","publish_at":"2026-10-28 16:30:53","processed_at":"2026-10-28 16:30:53","word_count":699,"reading_time":4},{"id":14835,"title":"OpenAI выпустила Decisions API: решения агентов до 10 раз быстрее","title_en":"Build more responsive apps with the Decisions API","preview":"Decisions API на GPT-6 Luna вышел из беты: до 10x быстрее Responses API, $0.10 за миллион входных токенов и нулевая плата за выходные. Sierra укладывается в 280 мс p95.","categories":"llm,business","category":"llm","image":"https://news.hamidun.com/images/news_14835_1793204950538.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14835/openai-vypustila-decisions-api-resheniya-agentov-do-10-raz-b","url_md":"https://news.hamidun.com/ru/news/14835.md","source":{"name":"OpenAI Blog","url":"https://openai.com/index/decisions-api"},"published":"2026-10-28 16:28:54","publish_at":"2026-10-28 16:28:54","processed_at":"2026-10-28 16:28:54","word_count":637,"reading_time":4},{"id":14834,"title":"Meta выпустила Muse Spark Open: 816B параметров и лидерство среди open-weight США","title_en":"Meta shakes up the frontiers, open and closed","preview":"Muse Spark Open с 816B параметров и 42B активных стал сильнейшей американской моделью с открытыми весами и вдвое сократил отставание США от китайских open-weight моделей.","categories":"llm,business","category":"llm","image":"https://news.hamidun.com/images/news_14834_1793204890709.jpg","audio_path":null,"quality_score":9,"url":"https://news.hamidun.com/ru/news/14834/meta-vypustila-muse-spark-open-816b-parametrov-i-liderstvo-s","url_md":"https://news.hamidun.com/ru/news/14834.md","source":{"name":"Interconnects (Nathan Lambert)","url":"https://www.interconnects.ai/p/meta-shakes-up-the-frontiers-open"},"published":"2026-10-28 16:27:36","publish_at":"2026-10-28 16:27:36","processed_at":"2026-10-28 16:27:36","word_count":687,"reading_time":4},{"id":14833,"title":"SLICE-Mem: память LLM-агентов без генерации режет расход токенов на 92–99%","title_en":"SLICE-Mem: Salience-Gated Lightweight Information Curation for Efficient LLM Agent Memory","preview":"Исследователи убрали генеративные вызовы LLM из онлайн-конвейера памяти агента: качество почти как у MemOS, а токенов уходит на 92,4–99,5% меньше.","categories":"llm,research","category":"llm","image":"https://news.hamidun.com/images/news_14833_1793198050826.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14833/slice-mem-pamyat-llm-agentov-bez-generatsii-rezhet-raskhod-t","url_md":"https://news.hamidun.com/ru/news/14833.md","source":{"name":"arXiv cs.CL","url":"https://arxiv.org/abs/2610.33015"},"published":"2026-10-28 14:33:52","publish_at":"2026-10-28 14:33:52","processed_at":"2026-10-28 14:33:52","word_count":549,"reading_time":4},{"id":14832,"title":"Инженер Thomson Reuters запустила офлайн-агента на Qwen3.5 9B в Гималаях","title_en":"Coding From the Himalayas With an Offline AI Agent","preview":"Риттика Джиндал собрала кодинг-агента из OpenCode, Ollama и Qwen3.5 9B. Он работает на ноутбуке без интернета: нашёл баг, правит одну строку и проходит 5 тестов из 5.","categories":"llm","category":"llm","image":"https://news.hamidun.com/images/news_14832_1793197998083.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14832/inzhener-thomson-reuters-zapustila-oflayn-agenta-na-qwen3-5","url_md":"https://news.hamidun.com/ru/news/14832.md","source":{"name":"Latent Space (swyx)","url":"https://latent.space/p/offline-coding/"},"published":"2026-10-28 14:31:40","publish_at":"2026-10-28 14:31:40","processed_at":"2026-10-28 14:31:40","word_count":694,"reading_time":4},{"id":14830,"title":"NVIDIA Kumo Relational обходит supervised ML на 41 задаче без обучения","title_en":"NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required","preview":"NVIDIA выпустила Kumo Relational: предобученная модель предсказывает отток и спрос прямо по реляционной базе за один проход и на SAP SALT опережает AutoGluon на 8% MRR.","categories":"llm,business","category":"llm","image":"https://news.hamidun.com/images/news_14830_1793197831696.jpg","audio_path":null,"quality_score":9,"url":"https://news.hamidun.com/ru/news/14830/nvidia-kumo-relational-obkhodit-supervised-ml-na-41-zadache","url_md":"https://news.hamidun.com/ru/news/14830.md","source":{"name":"Hugging Face Blog","url":"https://huggingface.co/blog/nvidia/kumo-relational"},"published":"2026-10-28 14:28:55","publish_at":"2026-10-28 14:28:55","processed_at":"2026-10-28 14:28:55","word_count":730,"reading_time":5},{"id":14829,"title":"Moonshot выпустила Kimi K3.1 Pro: 2,8 трлн параметров и веса 4 ноября","title_en":"[AINews] Kimi 3.1 Pro 104BA-2.8T: Fable-5.1-Class Open Weights narrows Closed-AI lead to 1 month... and Pro Highspeed API with 2-3× speed","preview":"Kimi K3.1 Pro на 2,8 трлн параметров вошла в топ-4 индекса Artificial Analysis с 55 баллами. Токены подешевели на треть, но задача стоит почти столько же, сколько у K3.","categories":"llm,business","category":"llm","image":"https://news.hamidun.com/images/news_14829_1793197553497.jpg","audio_path":null,"quality_score":9,"url":"https://news.hamidun.com/ru/news/14829/moonshot-vypustila-kimi-k3-1-pro-2-8-trln-parametrov-i-vesa","url_md":"https://news.hamidun.com/ru/news/14829.md","source":{"name":"Latent Space (swyx)","url":"https://latent.space/p/ainews-kimi-31/"},"published":"2026-10-28 14:25:25","publish_at":"2026-10-28 14:25:25","processed_at":"2026-10-28 14:25:25","word_count":644,"reading_time":4},{"id":14828,"title":"Швейцарский Apertus 1.5 получил зрение, слух и контекст в 262 тысячи токенов","title_en":"Apertus 1.5 Release Notes","preview":"Проект Swiss AI выпустил Apertus 1.5 на 8B и 70B под Apache 2.0: мультимодальность через токенайзер, режим размышлений, контекст в 4 раза длиннее и полностью открытые данные.","categories":"llm","category":"llm","image":"https://news.hamidun.com/images/news_14828_1793190905339.jpg","audio_path":null,"quality_score":8,"url":"https://news.hamidun.com/ru/news/14828/shveytsarskiy-apertus-1-5-poluchil-zrenie-slukh-i-kontekst-v","url_md":"https://news.hamidun.com/ru/news/14828.md","source":{"name":"Hugging Face Blog","url":"https://huggingface.co/blog/swiss-ai/apertus-v15-release-notes"},"published":"2026-10-28 12:34:36","publish_at":"2026-10-28 12:34:36","processed_at":"2026-10-28 12:34:36","word_count":722,"reading_time":5}]}