Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
A trilogy on GEO — optimization for ChatGPT, Claude, and other neural networks has been published
A trilogy on GEO has been published on Habr: the expert argues that content optimization for ChatGPT and Claude is not an extension of SEO, but a new branch of marketing.
The US closed access to Claude Mythos, China restricts GLM and Qwen: neural networks have become a strategic weapon
The US removed Claude Mythos from open access over cybersecurity risks, while China is preparing a ban on the export of GLM, Qwen and DouBao — the start of the era of digital sovereignty.
Yandex Realtime vs OpenAI Realtime: developer measured latency and migrated production
The developer of a voice AI receptionist compared three speech-to-speech approaches: Yandex Realtime (~330 ms), OpenAI (~740 ms via VPN), and an outdated three-stage pipeline (~4.4 seconds) — and chose a winner.
habrGPT: training a 0.5B LLM from scratch on Habr articles — Karpathy’s nanochat and an fp8 vs bf16 comparison
An experiment in training a 500-million-parameter language model from scratch on the Habr corpus: Karpathy’s nanochat, a home GPU instead of an 8×H100 cluster, and a comparison of fp8 and bf16.
Rosselkhozbank publishes second AI digest by AI evangelist from its Big Data Lab
Olga Popova from Rosselkhozbank's AI lab continues the series of digests on artificial intelligence for the professional community.
Hallucination detector built from 6 language models caught errors in gold labels
Researchers assembled an ensemble LLM hallucination detector from six models and found that the judges' supposed blind spot was actually errors in the dataset's gold labels.
Citation or mention: what really boosts a brand in ChatGPT and Perplexity answers
For several months, the author tracked ChatGPT, Perplexity, and Alice answers about brands — and found that a link in an AI answer’s sources is not the same as real visibility.
A year with AI assistants in production: the experience of VTB, T-Bank, and Dodo Engineering
VTB, T-Bank, Dodo Engineering, and S7 TechLab discussed what actually changed in teams after a year of widespread use of AI coding assistants and whether efficiency increased.
Employees are leaking corporate data into ChatGPT: volume grew 30-fold in a year
Solar Group analyzed 150 companies and found that 60% of them do not control data transfers to public AI services — and this is already exposing them to fines.
Veo 3.1, Kling, Sora 2 and Grok: a test of five AI video generators with identical prompts
Habr authors ran Veo 3.1, Kling v3 Omni, Sora 2 Pro, LTX-2.3 Pro and Grok through seven identical scenarios — from a knight to a panicky cat.
Neural networks are bad calculators: why LLMs fail at arithmetic and what to do about it
Language models predict tokens rather than calculate, which is why they make confident mistakes in the middle of long numbers. We break down the mechanism and explain how the industry works around it.
A 49,000-parameter neural network instead of Playwright tests for a browser strategy game
A developer replaced part of the E2E testing of a browser space strategy game with a 49,000-parameter autopilot — TypeScript without ONNX, five pitfalls on the way from 67% to 99.7% accuracy.
Sysdig Describes First Autonomous Ransomware Attack by AI Agent Without Human Involvement
Sysdig researchers documented the first fully autonomous ransomware attack in history: an AI agent without human involvement completed the entire compromise cycle—from reconnaissance to file encryption.
Shadow AI Eats Budget and Data: How to Choose a Corporate AI Platform in 2026
Large companies have discovered dozens of uncoordinated AI initiatives without control over spending and data leaks — and are seeking a single managed foundation.
Raiffeisenbank: 500 engineers adopted AI, but development speed didn't increase
Marat Kinyabulatov from Raiffeisenbank explains why the surge of 500 engineers using AI tools didn't accelerate development and what to do about it.
NVIDIA Dynamo vs Triton Inference Server: what changed for agent inference
NVIDIA released open-source framework Dynamo for agentic LLM inference with KV-cache routing — unlike Triton, not designed for parallel agent calls.
AI agents hide their errors: 6 patterns of silent failures in automation
OTUS on Habr analyzed 6 patterns of silent failures in AI agents: the system returns a convincing answer, but it hides an incorrect action or state loss.
How AI wrote all the code: a platform built in 2.5 months instead of years
Two developers launched a geo-analytics platform in 2.5 months, giving up writing code entirely: AI generated everything from specifications.
MTS Opens AVI Code: Smart Firewall for LLM That Learns New Rules in Seconds
MTS AI team has open-sourced AVI — an external modular firewall for LLM that filters prompt attacks and illegal responses without retraining the model itself.
"AI Slop": Telegram Bot Developer Figured Out Why the AI Label Irritates the Audience
Developer built a Telegram bot for three months, and the first top comment called it "AI slop" — the author figured out why these two words irritate technical readers.
How a developer cut an AI agent's token consumption by 4.6M per day: analyzing the leak
Cron tasks, diagnostics, and massive tool lists silently drain millions of tokens — a developer found the leak and reduced daily consumption by 4.6M tokens.
Anthropic Prompts vs grill-me Skill: A Two-Hour Test by a Non-Developer in Claude Code
A Habr user with no development experience compared five ready-made prompts from Anthropic's library and the grill-me skill — and wrote a Habr parser during the session.
Yttri 0.86.0-beta.1: First Obsidian Plugin, MLX Engine and Unified Agent Interface
Yttri released version 0.86.0-beta.1: agent moved to unified interface section, plugin platform launched with Obsidian integration, and local MLX engine for Mac now available.
Hallucinations in AI coding tools: why agents are slower than search-and-replace
A developer spent two hours on a task for a CLI agent that could be solved manually in 20 minutes — and investigated why detailed prompts amplify LLM hallucinations.
Fable vs 10 LLMs: Who Better Handles the God Node from a Real Agent
The author tested 11 language models—including Fable and six Chinese ones—on the task of decomposing a real god node from a LangGraph agent, using cross-evaluations and three trust verification methods.
AI in development: vibe coding instead of developers
Katya Obraztsova of Bitrix24 on how a five-person team built a corporate service from scratch in three months using vibe coding and evolved into an AI-driven team.
Inside Chinese AI: ByteDance, Z.ai, humanoid robots requiring passport ID and the server GPU card market
Eight days of negotiations with ByteDance, Z.ai, SiliconFlow and MiniMax in China: what launches July 10, why Z.ai is switching from NVIDIA to Huawei and how humanoids are sold in Shanghai with passport verification.
MCP and AI Agent Security: How the Protocol Created a New Attack Surface
Model Context Protocol standardized AI agent connections to tools — but each MCP server is now a trust point, vulnerable to tool poisoning and prompt injection.
VK Created Discovery AI Neural Search for Multi-Billion Content Base of Zen and VK Video
A leading VK developer unveiled the architecture of Discovery AI — an LLM-based system for search and recommendations across Zen, VKontakte, VK Video, and Mail.
How to build an AI harness for SaaS development: combining Orca, Pi Agent, and GitHub
A developer on Habr explains why a well-built harness around a model matters more than the choice between GPT or Claude itself — and shows his working setup.