Хабр: Машинное обучение
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Habr Author Tests How Many Edits It Takes to Erase an LLM Watermark
A Habr author tests whether Claude and OpenAI watermarks survive rewrites, error fixes and code. Four questions are posed; the results are still to come.
Alibaba's Qwen-Image-2.1: 7 Billion Parameters, 2K Output, and It Runs on a 16 GB MacBook Air
Qwen released Qwen-Image-2.1, a 7 billion-parameter diffusion transformer with output up to 2K. Enthusiasts already run it locally on a 16 GB MacBook Air with no safety checker.
СВОЙ Тех builds an internal AI agent platform: 200 tools and 5 chats
Fintech company СВОЙ Тех described its internal AI agent platform: 200 tools, five chats for different business tasks, and dedicated context handling.
The "Senior" Role Failed the Task 8 Times Out of 8: 156 Runs Test Prompt Add-Ons
A Habr author ran a coding agent through 156 runs with hidden tests. Tips raised the bill the most, "think" helped almost for free, and "senior" set the record for worst result.
Haiku Returned PASS and a Major Flag in a Review of Ten Articles: Breaking Down the Report
While reviewing ten articles, Haiku issued a PASS and flagged a major error in the same report. The authors checked it against the sources, accepting some fixes and rejecting others.
Product Specialist With No SOC Experience Takes First Place at BlueSec1 With an AI Triage Agent
A vulnerability management specialist built an AI agent for SOC tier 1 and won the open stage of BlueSec1. Then the scores were "zeroed out." Here is the story.
Habr author bypasses SynthID on a 1536×1024 frame using ControlNet Canny
JPEG compression and noise did not break Google's watermark, but blending in texture from a second ControlNet Canny version of the scene returned "not detected."
K2 Cloud Product Manager Warns: A Cheap GPU Hour Can Blow Up Your AI Project Budget
A K2 Cloud product manager explains why the lowest hourly GPU rate doesn't guarantee a cheap project: idle time, storage, networking and engineer hours eat the budget.
Enthusiast builds a RAG bot on a T4 for 9th- and 10th-grade social studies textbooks
A prototype built on Qwen2.5-3B, FAISS and multilingual-e5-large answers questions from 9th–10th-grade textbooks and cites the page. It runs on a free T4 GPU.
ReviewScope Uses Embeddings and Graphs to Find Anomalies in 21,000 Reviews
After moving, the author found an odd Yandex Maps review and built ReviewScope. It combines embeddings, match graphs and weighted ratings, and was tested on 21 thousand reviews.
Developer of GigaAM-Powered Offline Dictation App Brings Pisar to Google Play, Asks for Testers
The developer of offline dictation app "Giga Pisar," built on Sber's GigaAM model, is moving its Android version from APK to Google Play, but needs community help to publish it.
A Habr Essay Reopens the Debate Over Who Owns Open Source Code in the Age of AI
A Habr essay revives the licensing debate: from the Redis and Terraform forks to the 700 million-user threshold in the Llama license and the GitHub Copilot lawsuit.
Avito Built Aviflow: An ML Platform for Distributed LLM Training on HGX
An engineer on Avito's ML platform explains how the team moved from DevBox machines with SSH access to Aviflow, a cloud-native platform for distributed LLM training on HGX servers.
YOLO Annotator: An Offline Labeling Tool for bbox, OBB and Panoptic Segmentation Under GPL-3.0
The team behind a road-defect project released YOLO Annotator: a serverless desktop tool for bbox, OBB, mask and panoptic labeling, with export to YOLO, Detectron2 and Mask2Former.
Avito Breaks Down Six KubeCon Japan 2026 Talks on Sharing GPUs
An Avito ML Ops engineer distilled six KubeCon Japan 2026 talks on GPUs and LLM inference: tenant isolation, quotas, bin packing. And what the company has already tried.
Habr Author Sets Up a Fake teamo Profile to Test Scam Bots
A Habr user built an "ideal" profile on the dating app teamo to see how bots and neural networks work in scams, and how much money they make.
RAG Over 11,000 Files Confused Part Numbers and Returned Data From the Wrong Machine
A bot for 100+ sales managers searched 11,000 technical files and pulled figures from a neighboring machine's datasheet. Why dense vectors fail to catch part numbers.
Open-source video editor Wunjo Make can now get plugins written via Claude Code, Codex and Cursor
The Wunjo Make repo now includes a builder and skills for cloud LLMs, so plugins that use local neural networks or APIs can be written with three AI agents and driven through MCP.
Authors Use 8 Examples to Check What Remains in Text After Masking Before an LLM
A Habr write-up: after names are masked, a text can still contain a number that identifies a person in a public registry. Eight synthetic examples, a residue list, two switchable components.
Habr Breaks Down APM Auto-Instrumentation: Where a 4-Second Request Loses Its Time
One request crosses a gateway, microservices, a queue and a database in 4 seconds. Here is how APM finds the culprit without manual code changes in every service.