Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Not a Model, but a System: How Svoi Built a 7-Layer Fintech Bot Architecture
Svoi.ru explains why voice bots for banks can't rely on just a good model — they need a seven-layer architecture where each layer manages its own part of the risk.
An AI Agent Created a Ticket, Took It to Work, and Closed It—The Manager Noticed Nothing
Autonomous AI agents are embedded in real CI/CD pipelines and close live tickets in production. Metrics look perfect, hiding the actual project state and growing technical debt.
SciGraph: how a graph of scientific connections outperforms text search
SciGraph applies a graph approach to scientific papers: instead of just text, AI sees connections between authors, methods, and citations — and answers research questions more accurately.
Local Qwen3.6-27B vs. cloud models: why privacy is not the main point
Local Qwen3.6-27B shows it can compete with cloud-based Claude and GPT. The advantage of local models extends far beyond privacy — it is control, speed, and independence.
llms.txt: how to help ChatGPT, Claude, and Perplexity cite your site correctly
llms.txt is a file in your site's root directory that tells ChatGPT, Claude, and Perplexity how to cite you and what to treat as canonical information.
Programming Languages Will Be Redesigned for AI Agents, and Here's Why
When AI agents start writing code, existing languages won't be enough. The industry faces a wave of new languages optimized for machine understanding, while old ones will remain.
Zabbix and Local LLM Integration: How to Design an Architecture for Smart Alerts
Part three of the Zabbix and LLM integration series explores smart alert architecture design. Discover which HLD components humans handle and where neural networks help.
A conductor instead of an assembly line: how AI rethinks the classic pipeline
The traditional development pipeline—from product to DevOps—now encounters AI agents at every stage. The system must either change or stop working.
RAG systems break on real data: the culprit is retrieval, not the model
A RAG system looks perfect in a demo, but once testing starts on real questions, half of its answers turn out to be wrong. The problem is not GPT-4 or Claude — it is how the system retrieves relevant chunks from document
Lyapunov Against the System: How Forbidden Cybernetics Was Developed in the USSR
In 1954, mathematician Alexei Lyapunov opened a cybernetics seminar at Moscow State University when it was branded a reactionary pseudoscience. His courage helped save the ideas that became the foundation of modern AI.
Anthropic exposes G20 banks’ vulnerabilities, Meta invests 130 billion in AI
Anthropic held a briefing for the Bank of England and showed how Claude finds thousands of critical flaws in banking infrastructure. At the same time, Meta is investing 125-145 billion in AI and cutting its workforce by
Three AI models instead of BI tables: how to analyze product profitability on Ozon
Sellers on Ozon have been losing money for years because they cannot see the real profit for each product. SKUmind uses three AI models instead of traditional BI rules to show the true margin and suggest what to do.
Google completely redesigned Gemini: thinking levels instead of Thinking model
Google has completely redesigned Gemini. Both the web version and the mobile app have been updated—new interface, built-in thinking depth modes, a new lightweight Flash Lite model.
Why bots based on protected LLMs are frequently hacked: analysis of 14,000 GPTs
It turns out that even if the base LLM model is protected from attacks, a custom bot built on it can be vulnerable. Analysis of 14,904 GPTs revealed where exactly security gaps appear.
RAG in enterprise: why 80% of problems are in data, not in the model
A RAG prototype is built in a week and demonstrated. It looks magical: the model answers questions about your documents without hallucinating from general knowledge.
ICLR 2026 in Rio: What Researchers Think About Scaling Large Models
ICLR 2026 took place in Rio de Janeiro in late April and confirmed its status as the premier annual gathering of AI researchers.
Why AGI benchmarks will never be objective
Researchers are trying to develop tests for strong AI, but they face a fundamental problem: even human intelligence has 70+ definitions, and scientists do not agree on what consciousness in neural networks is.
Designer equated AI agents with medical intermediaries rather than with users
This week on Habr, a discussion unfolded about who should be considered AI agents in interface design.
How to Create a Fully-Functional AI Agent in 10 Steps: A Guide for Russian Business
AI agents are neural networks assembled into working systems capable of independently performing repetitive tasks.
LLMs struggle with unsafe Rust: six months of testing revealed critical errors
For six months, a researcher had language models write unsafe Rust and found seven categories of recurring errors — from aliasing to memory issues.
MTS launched Methan: AI for data search in a catalog of 500 thousand tables
MTS has created a system called Metan — an intelligent assistant that helps analysts find the necessary data in a corporate catalog of 500 thousand tables simply by asking a question in Russian.
Yandex Music is one-third AI-generated tracks — investigation results
Yandex Music faced an uncontrollable flood of AI tracks. Investigation results showed that the platform has essentially lost the ability to distinguish human music from AI-generated content, and this…
How Google Cloud Protects Systems from AI Agents: Agent Gateway
A developer from the startup PocketOS gave an AI agent based on Claude Opus 4.6 a task. The agent executed it literally — in 9 seconds, it deleted the entire company database along with backups.
AI agent at the medical assistance desk: how Sovcombank automates complex calls
How the insurer integrated an LLM into client call processing, where errors can be costly and resolution speed is critical.
Tested Popular GPT Shortcuts: What Works in Real Tasks
When you work with GPT every day, the first thing you want to do is speed up the process. For this, people invented shortcuts: short commands like EL5 or /BULLET that suggest to the model a style and…
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.