Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
What the Claude Code leak revealed about the future of Anthropic's AI agents
Anthropic accidentally published the full source code of Claude Code version 2.1.88 due to an npm configuration error. The code revealed 44 hidden flags for autonomous agents, background thinking, and a controversial mod
Claude Cowork's File Autopilot: Three Examples of Hours Saved
Claude Cowork turns hours of routine work into minutes. With direct access to local files, a single prompt lets Cowork build a plan and execute the task — from organizing folders to sorting videos and structuring receipt
Why Microsoft, OpenAI, and Anthropic Are Telling You Not to Trust Their Own AI
Microsoft's Terms of Service declares Copilot pure entertainment, OpenAI demands a human check every output, Anthropic caps its liability at £100. Tech giants are trying to inflate the AI bubble and legally shield themse
Fine-Tuning Gemma 4 with GPUs on Cloud Run: From 89% to 94% in Animal Classification
A complete breakdown of fine-tuning Gemma 4 on custom data in Google Cloud. An architecture with native support for video and audio, QLoRA quantization, and accuracy climbing from 89% to 94% on the Oxford-IIIT Pet datase
Anatomy of AI Text: How Editors Spot Articles Generated by Neural Networks
A flawless structure and zero errors no longer guarantee a job. We look at why editors are rejecting candidates' immaculate writing samples en masse, and how to spot the fingerprints of AI.
Reasoning Drift in AI: How a Model Degrades While the Service Keeps Running
Commercial AI models like Claude can quietly lose quality without ever breaking the service itself. Reasoning drift — a decline in the ability to handle multi-step reasoning — is becoming a real business problem. And pro
Three AI Inflection Points: From ChatGPT to Agents and Physical Intelligence
Jensen Huang described AI's evolution through three inflection points: generative AI, reasoning AI (the O1 model), and agentic AI. Token consumption has grown by millions of times. The next frontier is physical AI, which
Claude Mythos Is Big Tech-Only, CaveMan Cuts Token Spend Threefold
Anthropic has restricted Claude Mythos to the Project Glowswing consortium (AWS, Google, Microsoft, Nvidia, Cisco, Apple). The model hits 77.8% on SWE-bench Pro versus 53.4% for Opus and has already surfaced bugs in Linu
Altman's Shadow: How One Man's Ambitions Are Shaping the Future of AI
How Sam Altman took control of the AI industry and pushed rivals aside. Is the current path of the technology just a reflection of one person's interests?
Why Multimodal Models Can't See Web Interfaces
Multimodal models struggle with interfaces: code hides reality, and screenshots require analyzing millions of pixels. What's needed is a built-in module for understanding structure instead of code or pixels.
KodaCode rolls out agent control via Telegram and voice input
KodaCode 0.9.0 lets developers control the agent through a Telegram bot and voice input. Remote control is already available in VS Code, with both features planned for CLI and JetBrains IDEs.
The End of Openness: Meta Replaces Llama With the Closed Muse Spark Model
Meta stunned the AI community by dropping Llama in favor of the proprietary Muse Spark. The move to a closed model puts thousands of startups at risk and changes the rules of the market.
Hidden Details of Claude Opus 4.7: Trade-Offs, Tokenization and a New Personality
A look at the less obvious details of the Opus 4.7 release: why it trails its predecessor on some benchmarks, how the tokenizer changed, and why the model now argues with you more when you write code.
Automating Code Review: How Neural Networks Are Taking Over Developer Grunt Work
A team of LLM engineers automated code review using Qwen3-Coder-30B through a three-stage pipeline: issue generation, comment validation, and semantic deduplication. Processing 1,000 PRs per week, the system achieves 85%
How Automating 70% of Tasks Revealed the True Value of Human Work
An AI expert automated 70% of his routine and discovered that the remaining 30% of tasks are the real key to success.
LLM as Decoder in ASR: Adapting Nvidia's SALM Architecture for Russian-Language Telephony
Kontur adapted Nvidia's hybrid SALM architecture for Russian-language telephony and video conferencing. Instead of a classic cascade system, they used an LLM (Qwen) as the decoder paired with a FastConformer encoder. The
BCG: Half of Managers See AI Eroding Their Employees' Skills
Half of the managers BCG surveyed notice deskilling: employees are losing professional skills because of AI. A Habr author breaks down ten habits for working with AI, five harmful and five useful.
A honeypot repo with 1,774 stars: two filters for detecting malware and hallucinations
A Telegram bot sent a repo with 1,774 stars — a disguised malware trap. After auditing 321 repos from trending feeds, the author built two filters: one to detect malware, one to check DeepSeek's AI hallucinations.
Developer Puts Claude Code and Codex Agents in the MacBook Notch, at 0.5% CPU
A Habr author built a macOS app that shows Claude Code and Codex status in the MacBook notch. A glow signals a waiting agent, and each session appears as a figure in a pixel-art office.
Developer Built Lumi with Claude Code: a Free AI-Powered Spotlight for Mac and Windows
Free Lumi combines file search and AI answers in a single hotkey on Mac and Windows. Supports Claude, ChatGPT, Gemini, OpenRouter, Ollama. Built with Claude Code, source open under MIT.
How to Secure AI Systems and the Development Pipeline: PRO DevSecOps Conf 2026 Talks
Representatives from the Bank of Russia, ISP RAS, Gazprombank, and VimpelCom spoke at the Swordfish Security conference. Topics included SAIMM v2.0, requirements under Federal Law 243-FZ on AI, and Shadow AI — 30+ talks
Tech Aquarium on Two GTX 1660 Supers: How I Ran Qwen 2.5 and Built It Into INFERNUM
Developer McCake built a home PC with two GTX 1660 Super GPUs (12GB VRAM combined), ran Qwen 2.5 14B through Ollama, and embedded it into INFERNUM, an interactive site powered by 5 Python scripts. The result: 47.59 token
AI Speeds Up Coding, but Productivity Isn't Rising: What DevEx Metrics Explain
Teams expected 3–4x output, yet in METR's experiment experienced developers using AI worked 19% slower. Why DORA and SPACE don't explain this, and how DevEx fills the gap.
AI Agents in Tech Support Act on Their Own and Break the Ticket Model
Agents unlock accounts, grant access, and restart services without an operator. The ticket, support's unit of account since the late 1980s, stops being a measure of work.
Ingosstrakh Teaches a Multimodal LLM to Doubt Itself: Catching Document Errors with Logprobs
Ingosstrakh's AI center tested whether token probabilities can flag LLM errors in OCR-free document extraction and send doubtful results to manual review.
Analyst builds UML editor with GOST diagrams using Claude: first version in 2 hours
A systems analyst built umleditor.ru with Claude: browser-based UML, his own PlantUML server in Moscow, a GOST flowchart renderer and a reference section of 70+ pages.
Timeline of AI Agent Incidents: Two Classes and an Isolation Checklist
Over three months, OpenAI, Anthropic, Google, and Meta disclosed AI agent incidents. Two distinct classes got conflated: test-environment escapes and actions on the live internet. Specialists identified the causes of eac
RAG Failures: 6 Reasons Your AI Agent's Memory Lags, and How to Diagnose Retrieval
73% of RAG system breakdowns are caused by bad search. SpeShuAI breaks down 6 typical failure modes — from chunking that destroys context to semantic gaps between query and document. Plus diagnostic metrics and when you
Oracle Piled Up Hundreds of Billions in AI Debt, and Investors Started Asking About the Money
Oracle is building AI data centers worth hundreds of billions of dollars. Its free cash flow is negative, debt is climbing, and the market is asking who will pay for it all.
CayleyPy Beats Classical Algorithms at Solving Large Rubik's Cubes
CayleyPy uses neural network heuristics instead of classical algorithms to search Cayley graphs. On the 3×3 cube, the model hit 98.4% optimal solutions, and on the 4×4 it beat Tomas Rokicki on the Kaggle leaderboard with