Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Agent Development: What Strategies Actually Work with AI
Manual coding is inefficient — no one doubts that anymore. But agent-driven development is not a universal solution: there are optimal and suboptimal ways of working with neural networks.
Pollux by Sber AI: an LLM judge for evaluating Russian-language models
Sber AI has opened access to Pollux, a specialized model for automatically evaluating the quality of language models in Russian before deployment to production.
Software Development Is Not Coding: Why AI Solves the Wrong Problem
Software development is not simply writing code. AI solves this narrow task well, but the entire engineering process consists of dozens of other problems that machines cannot handle.
How to Get Into ChatGPT Answers: New SEO for Business
A client asked web consultant Dmitry, 12 years in the industry: help, I want ChatGPT to show our company at the top when someone searches in our category.
Three rules for development with an AI assistant: from requirements to maintainable code
Developers have long been afraid to hand over their code to an AI assistant. The fear makes sense: an assistant can violate requirements, break legacy code, or write unmaintainable solutions.
OTUS launches open lessons on AI agents, Kubernetes and Go in May
OTUS is preparing an intensive block of open lessons for May 18–28. Over 10 days, you can participate in 9 different tracks — from Kubernetes to AI-agents — and understand which topics to explore in…
Spring Agent Toolkit: effective tools for managing AI agents
AI agents are rapidly moving from the category of experiments into production systems, where every penny spent on tokens matters.
Codex: when an AI assistant becomes a universal work tool
A developer shares how Codex allows stepping beyond programming and automating a wide variety of work tasks. This became possible thanks to tool updates that made it a truly universal assistant.
MAX messenger launched AI Hub: 10 models and 5,900 users in 54 days
The team integrated 10+ AI models into the MAX messenger in just two months: already 5,900+ users and an operational premium subscription via YooKassa.
Robot with Google LLM: how a 270M-parameter model was trained to control movement
An engineer integrated a compact Google language model into a tracked robot and trained it in simulation to independently grasp items, manipulate objects, and move through space — an example of the successful use of comp
Perplexity in court over content: the four founders behind the $21 billion scandal
AI search engine Perplexity is valued at $21 billion, but has already been hit with lawsuits from The New York Times and the BBC — the company is accused of copying content without a license.
LEGO Education abandons its own platform and introduces AI and NFC cards
LEGO Education is reshaping STEM education: abandoning its own platform, reducing the emphasis on programming, and introducing AI assistants and control via NFC cards.
Why GPT gets letter counts wrong: the secret of tokenization
When processing text, LLMs use tokens rather than letters — and that explains why neural networks make mistakes in simple character counts within words.
Chrome automatically downloads a 4 GB AI model without user consent
By default, Chrome downloads the 4 GB local AI model Gemini Nano without asking the user. Estimates suggest that a single rollout to one billion devices results in 6,000-60,000 tons of CO2.
Scammers distribute fake Claude installers with malware
Threat actors create fake websites for Claude and other AI tools, distributing malicious installers through spam and social media.
OpenAI launched /goal in Codex CLI: autonomous agent or an expensive way to code?
Codex CLI got /goal — an autonomous development mode. In practice: token usage spikes 5x, MCP integrations fail silently, and the cost of this "automation" stays opaque.
Russian in ChatGPT costs 2x more: tokenization is to blame
In ChatGPT and other LLM services, Russian text costs twice as much as English, is processed more slowly, and fits less information — all because of the peculiarities of Cyrillic tokenization.
From an LLM prototype to a working product: how to avoid mistakes
An AI prototype can be put together in an evening, but between a working demo and a product people are willing to use and pay for lies a gap filled with dirty data, the wrong metrics, and an unclear implementation strate
Anthropic held Code with Claude 2026 with a record number of developers
The second Code with Claude conference in San Francisco drew more people than expected. Anthropic announced expansion to London and Tokyo.
Memory for an AI Agent: From Complexity to 300 Lines of Code
A developer described the search for a working memory system for AI agents — from the complex MemPalace architecture to a simple markdown-based solution and scalable ClickHouse.
Hermes Agent: How to Run a Personal AI Agent on Your Own Server
Nous Research has released Hermes Agent — an open-source AI agent that you can run on your own VPS and train for your specific tasks.
GLM-5.1 helped develop a new algorithm for analyzing bit strings
A C++ programmer used GLM-5.1 and OpenCode to develop an original algorithm based on known components — and explained how to get reasonable code from AI.
The geometry of attention: how QK Norm teaches the model to understand meaning
Query-key normalization in transformers is not just a stability hack, but a mechanism that forces the neural network to rely on semantics rather than vector magnitude.
AI agents are changing development: Cursor and Claude Code against the skeptics
AI agents are drawing mixed reactions in the IT community: some see a way to automate routine work, others fear technical debt and security issues. We examine how Cursor and Claude Code work.
How to Measure an AI Agent's Performance in QA: The Story of a Benchmark
Engineer Mikhail Fedorov created an objective benchmark for testing AI agents instead of relying on subjective assessments of performance.
Two windows with an AI agent instead of one: architect, developer, and six types of errors
Separating roles across two AI-agent windows — the architect plans, the developer executes — addresses six common errors and removes the carryover from previous tasks.
How I built 4 MCP servers and a protocol to connect them
A developer built four MCP servers for Claude, created a protocol for them to interact, and assembled an automation pipeline for Telegram and Yandex.Direct.
ArchiMate and AI: how architects can design systems more efficiently
ArchiMate notation is rarely used by architects, but it is powerful — an Axenix expert explains its application and the role of AI in system design.
Suno v5 in 2026: the AI music generator that topped the Yandex Music charts
In 2026, AI topped the Yandex Music and VK charts for the first time. We tested Suno v5 and its competitors to find out whether a neural network is ready to create music at a professional level.
Claude blocks Russian IPs: how to avoid losing your AI project
Hundreds of Claude users lost their accounts because of a block on Russian IPs. Even VPNs did not help. We explain how to protect your AI project from a sudden shutdown.