Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
T-Bank improved AI Code Completion quality with a filter and removed unnecessary suggestions
T-Bank implemented a two-layer filter for AI Code Completion: first, an LLM evaluates whether a suggestion is relevant, then CatBoost refines the decision, delivering a +5.2 pp increase in Acceptance Rate.
BotHub, GPTunnel, and ruGPT: how Russian AI aggregators are dividing the market in 2026
A comparison of BotHub, GPTunnel, ruGPT.io, Syntx AI, and AI Wiz found that Russian AI aggregators compete not on price, but on convenience, local payment options, and different usage scenarios.
Habr AI described the architecture of a reflective agent: less autopilot, more self-checking
Habr AI showed how a reflective agent works: with drafts, pauses, rollback, and mandatory self-checks before acting on code, data, and business processes.
Habr AI: How prompt engineering in development saves hours but does not replace understanding the tasks
Habr AI examines practical prompt engineering for development: good instructions really do save hours, but they only work together with an understanding of the task, the code, and the project's constraints.
Claude Code hides history: 715 sessions found on disk, though the interface shows 69
A look inside Claude Code showed that the app stores hundreds of old sessions locally but does not show them in the interface: 48 MB of history and 715 chats were found on disk.
Claude and Ollama helped an Ozon seller process certificates in a day and restore 1,000 SKUs
An Ozon seller used Claude and a local Ollama model to recognize 231 certificates, upload them in bulk via the API, and put more than 1,000 blocked SKUs back on sale in one day.
Claude Code without the magic: Habr breaks down the architecture, context noise, and engineering practices
Habr has published a detailed breakdown of Claude Code: why the agent loses efficiency in long sessions, how context gets consumed, and why skills, hooks, subagents, and verification should be separated.
Ollama and LiteLLM: how to turn a Python script into a complete console-based LLM chat
In the second part of the practical guide, the author shows how to turn a basic Python script with Ollama and LiteLLM into a console chat with a dialogue loop, a system prompt, and error handling.
Sber showed how RAG and LLM in the IDE turn manual scenarios into automated tests
Sber described an IDE plugin that uses RAG and LLM to search for examples in the codebase and turn manual scenarios into automated tests, while preserving the project’s style and internal rules.
VS Code 1.111 launched Autopilot: AI agent writes, tests, and deploys on its own
Microsoft released VS Code 1.111 with Autopilot mode — the AI agent now works autonomously and completes the entire task without the developer’s involvement at every step.
Habr explained why language models and classic RAG lose their understanding of relationships
Habr examines why large language models with classic RAG are good at finding text fragments but struggle when they need to reconstruct relationships, causality, and hidden dependencies.
BotHub listed 10 free AI tools for text, code, video, 3D and documents
BotHub published a roundup of 10 free AI tools, including options for text, code, video, 3D models, document work and autonomous tasks.
LLMs in development: the 4 approaches teams use and how they differ
LLMs are no longer used only for autocomplete: this piece breaks down four modes of AI-assisted development along two axes — how much code a human controls and how the team verifies the output.
Cursor, Copilot, and Claude Code are included in a roundup of 12 popular AI agents for developers
The roundup features 12 AI tools for coding — from Cursor and Copilot to GigaCode and Snyk: editors, cloud IDEs, and security services with varying prices and availability in Russia.
OpenAI, Google, and Tesla set the week's agenda: GPT-5.4 mini, AI Studio, and Terafab
OpenAI opened GPT-5.4 mini to broad access, Google turned AI Studio into an app-building environment, and Tesla promised its own $20–25 billion chip factory.
AI Independence Bench compared 49 models and measured their resilience to user pressure
The new benchmark, AI Independence Bench, tests whether 49 language models can maintain their own stance, avoid apologizing without reason, and not agree with the user automatically.
How the CEO of digital agency bijobs.ru built a brochure website over a weekend with Claude Code
Bogdan Nepryakhin of bijobs.ru explained how he replaced a month of work by a designer and a front-end developer with Claude Code over a weekend — and spent an order of magnitude less.
Python developer reined in Claude Code with a public anti-regression config
A solo developer at CREATMAN built a configuration for Claude Code that keeps the agent from straying beyond the task and breaking projects — and released it for free.
Gemini: how Google turned Bard's failure into 750 million users and AI leadership
In 2023, an error in a Bard ad wiped $100 billion off Alphabet's market value — today Gemini outperforms GPT-5.1 in benchmarks and reaches 750 million monthly active users.
Perplexity turned AI search into an agent that builds Excel files and checks data on its own
In seven minutes, Perplexity Computer assembled a four-sheet Excel file with 33 sources, wrote its own Python script, and showed how AI is moving from search to real analytical work.
DeepMind proposed ten cognitive scales for measuring progress toward AGI
Google DeepMind published the framework "Measuring Progress Toward AGI" — ten independent cognitive scales that for the first time make it possible to compare AI systems objectively.
AI agents vs RAG: how ReAct works and why multi-agent systems are needed
A single LLM response is no longer enough — a chain of actions is needed. We explain what AI agents are, how they differ from RAG, how ReAct works, and why multi-agent systems are needed.
Lemana Tech showed how it combined LLM, RAG, and traditional ML in tech support
The company described a hybrid support setup: fast ML classifiers handle high-volume tickets, while LLMs with RAG are used where Wiki-based answers and smart escalation are needed.
Raiffeisenbank implemented a RAG assistant in Kotlin without Python or new expertise
Raiffeisenbank’s team built an internal RAG assistant on the JVM stack — Kotlin and Spring Boot — without Python and without bringing additional expertise into the team.
Garage Eight: how AI is changing the work of analysts and why junior positions are disappearing
A Garage Eight analyst described six trends that are already reshaping the profession: AI is taking over routine work, raising the bar for juniors, and making managers more self-sufficient in working with data.
OpenClaw broke multi-agent work down into three modes: standalone agents, subagents, and ACP
OpenClaw explained how it structures multi-agent work through Telegram: standalone agents get memory and a workspace, subagents handle one-off tasks, and ACP delegates development to external CLI tools.
Habr AI Demonstrates How Reflex Architecture for LLM-Agents Eliminates Lag Down to 60 FPS
Habr AI explored dual-process architecture, where a fast reflex layer handles instantaneous reactions, while the LLM handles semantics, planning, and real-time adaptation.
Apple bets on local AI in M-series chips, not giant models
While the market measures itself by data centers and GPU clusters, Apple is pushing a different scenario: running AI near the user, on devices with unified memory and low power consumption.
Developer revealed how he built Roomify — an AI interior visualizer on React and Puter
Roomify converts floor plans into photorealistic 3D renders in seconds, with the entire project logic running on React, Puter, and Claude/Gemini API calls without a separate backend.
Entrepreneur built auto service CRM with ChatGPT and Cursor
An auto service owner without IT background described how since 2022 he built a CRM, Telegram bot and Android app for offline businesses using ChatGPT, Claude and Cursor.