Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
MCP LLM-Agent in Fusion360: First Vibe-Design Test on a Wooden Staircase
Fusion360 rolled out an MCP LLM-assistant — a Habr author immediately tested it on a wooden staircase he had previously created manually over 60 times.
Claude Code Raised Legal RAG to 0.791, but ARLC 2026 Final Hit Scaling Limits
The ARLC 2026 case demonstrates how Legal RAG can be improved across 17 iterations, then lose 42% of results as the corpus scales from 30 to 303 documents.
OpenAI and Grok Lose to Custom RAG in Legal Agentic RAG Challenge
A team on Mac Studio M3 Ultra compared RAG, CAG, BM25, and built-in databases from OpenAI and Grok in a legal challenge and found: grounding, not brand names, determines the outcome.
Habr: How Outdated Knowledge Base Breaks LLM-Agents and How to Fix It
Habr explored why outdated documentation is more dangerous than its absence for LLM-agents, and proposed three practices: automated checks, document statuses, and strict update rules.
Vibe Coding Promises 10x Productivity Gains — PyPI Hasn't Confirmed It
Developers claim 10–100x acceleration thanks to AI tools, but if that's true — why doesn't PyPI show explosive growth in Python packages?
Anthropic Strengthens Claude: Why New Corporate Customers Choose It Over ChatGPT
Ramp data from March 2026 shows a shift in corporate demand: among companies purchasing AI tools for the first time, Anthropic's Claude notably outpaces OpenAI.
NotebookLM and Gemini Automate Test Creation via Google Apps Script and Forms
An author demonstrated how to integrate NotebookLM, Gemini, Google Apps Script, and Google Forms to quickly assemble educational tests from existing sources while reducing manual work.
Doka launches local AI-agent for Russia without VPN, subscription or cloud
A developer bundled a custom AI-agent into the Doka desktop app: the service runs locally, uses the Qwen3 model, and requires no cloud, VPN, or monthly fees.
Why Language Models Make Mistakes Even When They Know the Right Answer: Breaking Down LLM Limitations
TechWill researcher Valery Shabashev analyzed why LLMs can make mistakes even when they have the correct answer: the problem isn't in knowledge, but in reasoning verification.
LLM Ensemble Examined Theological Interpretations: 1 Tim. 2:15 as a Static Analysis Case
Five language models compared two interpretations of 1 Tim. 2:15 and demonstrated that an LLM ensemble can avoid seeking 'truth' while uncovering hidden assumptions and weak points in arguments.
Bootik: a console AI agent for servers that runs locally and doesn't require much memory
Bootik is a lightweight console AI agent for SSH operations: it launches on demand, understands server context, and is designed for local models with limited VRAM.
Anthropic: agentic AI and GitHub Copilot are changing development rules in 2026
In 2026, AI code tools are no longer just code suggesters: Claude Code, Copilot, and Cursor now read repositories themselves, plan changes, run tests, and fix bugs.
SkillMarket built a marketplace of skills for AI agents on FastAPI and Claude Sonnet
The SkillMarket team showed an MVP of a platform where industry experts convert their experience into sellable skills for AI agents, with FastAPI, Celery and two LLMs handling packaging and moderation.
Sber Explained Why AI-Generated Code Looks Reliable But Breaks Under Real Load
Sber analyzed why AI-generated code often appears quality in code review but fails in production, and what engineering practices help reduce this risk.
"Advanced Payment Solutions" Launched Voice AI Assistant for Calls to Pilot Without ML Team
"Advanced Payment Solutions" brought a voice AI assistant for calls to pilot in six months, despite assembling the team from 12 backend developers with no ML experience.
Habr Explained How to Force LLMs to Calculate Without Errors Through Python Code Generation
Habr analyzed why language models regularly make arithmetic mistakes and demonstrated a working approach: LLM writes Python code, while the program handles calculations and generates Excel reports.
Paperclip Promises a Company of AI Agents but Drowns in Bureaucracy During Testing
A Habr author tested Paperclip — a system for managing teams of AI agents — and instead of accelerating development, got two days of coordination, delays, and rollbacks.
Krok Shows How It Built an Internal RAG Assistant for Corporate Data
Krok explained how it assembled an internal RAG assistant for searching corporate data with on-prem deployment, ACL control, quality benchmarks, and vendor platform customization.
Sergey Smirnov explained how to prepare an AI agent for reliable operation in production
AI engineer Sergey Smirnov outlined a roadmap for deploying an AI agent to production: from concept and initial launch to reliable, predictable operation for real users.
Google NotebookLM Helps Build a Personal AI Prompt Engineer in 15 Minutes
Instead of endless prompting courses, the author proposes creating a personal assistant in Google NotebookLM that answers only based on selected sources and doesn't make up parameters.
Microsoft GraphRAG and Ollama: How Graph-Based RAG Performed on Local Models
The author tested Microsoft GraphRAG with Ollama on the "Johnny Mnemonic" narrative and concluded: graph-based RAG delivers more relevant answers, but demands significant resources and manual configuration.
Habr AI Breaks Down Gradient Descent in C++ and CUDA Through MNIST Model Training
In the fourth part of the 'From MNIST to Transformer' series, Habr AI explains how gradient descent works and demonstrates training a digit recognition model in C++ and CUDA without PyTorch.
Google Turns Gemini Into Smartphone Infrastructure as Apple Steps Back From Siri
Google through Gemini claims the foundational AI layer role for both iPhone and Android simultaneously, while Apple effectively acknowledges that Siri in its current architecture cannot support the new era of mobile agen
Claude AI in Four IDEs: Why the Developer Became the Main Bottleneck for Agents
A Habr AI author described how working with Claude AI and multiple IDEs accelerates development, but makes the human the main bottleneck: they switch context, control agents, and ultimately lose time and sleep.
EU AI Act and EU Sanctions Intensify Pressure on Russian SaaS Companies in Europe
The EU and Russia are almost simultaneously moving AI regulation into a practical phase, and for Russian SaaS companies this becomes a double compliance barrier on top of already existing sanctions.
"First Form" showed how to maintain a company map so AI doesn't make mistakes
"First Form" demonstrated why corporate AI systems start confidently making errors if the data map isn't updated, and how to solve this with two loops—automatic and expert.
Veai 5.7 for JetBrains IDE adds directories to chat and auto-retry for agent tasks
Veai updated its AI agent for JetBrains IDE: version 5.7 brings directories to chat, model selection directly in the dialog, auto-retry, and a less intrusive terminal.
Habr AI: METR and Google Cloud See No Promised Developer Acceleration from AI
Habr AI compiled findings from several studies: AI assistants make developer work more comfortable, but data from METR, Google Cloud, and PyPI analysis do not yet confirm productivity growth.
KodaCode 0.8.0 adds secure chat mode, dialog branching, and image support
KodaCode 0.8.0 migrates chat to secure agent logic with read-only tools, adds session branching, JSON export, and image uploads for multimodal models.
SberDevices and ruGPT-3 XL: Developer Restores Forgotten Russian-Language Model from 2021
Developer brought back ai-forever/rugpt3xl — a Russian-language SberDevices model with 1.3 billion parameters, created in 2021 and trained from scratch for text continuation.