Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Habr: How synthetic data helps train models and why self-training leads to collapse
Synthetic data helps AI compensate for a shortage of quality human-generated corpus, but with uncontrolled self-training, models begin to lose rare cases and diversity.
Why ChatGPT and Gemini Won't Recommend Your B2B SaaS, Even if Your Website Is Well-Built
Even a well-designed B2B SaaS website may not appear in ChatGPT, Gemini, and Perplexity responses if your brand lacks a clear category, external validation, and a stable digital profile.
Why AI in UI Design Matters Not for Production, but as a Source of Visual Mutations
AI-generated UI is valuable not only for rapid sketching: its power lies in rare visual combinations that help designers discover new approaches and transform the interface design process itself.
Qwen and llama.cpp: how to run a local neural network without the cloud on your computer or server
A practical guide explaining how to run the Qwen model through llama.cpp on your own computer or server to work with a local neural network without clouds, subscriptions, and data leaks.
Anthropic and Claude Cowork: 10 work tasks AI removes from humans
Claude Cowork from Anthropic demonstrates how AI takes on morning briefings, proposals, client responses, and reports, freeing up two to three working days per week.
Directum: Why Business Actively Discusses AI Agents but Hesitates to Deploy Them in Processes
Directum explains why AI agents became the main corporate trend, but mass adoption is hindered by expensive infrastructure, error risks, and liability questions.
ClawRouter reduced LLM API costs from $47 to $1.80 per week — smart router review
ClawRouter analyzes each prompt across 15 parameters and routes it to the most cost-effective suitable model — reducing weekly LLM API expenses from $47 to $1.80.
Agent Coding as Addiction: Why Developers Can't Stop
Startup CTOs don't sleep until 3 AM without deadlines, Y Combinator CEO brags about 19-hour sessions — UC Berkeley researchers spot gambling addiction patterns.
PromptPilot: task scheduler for Claude Code and Codex that works while you sleep
A Russian developer created a task scheduler for AI CLI — PromptPilot accepts prompts from terminal, browser, or Telegram bot and executes them on schedule with auto-retry on rate limit.
Yandex Code Assistant for VS Code: How the Extension Has Changed and What Code Indexing Provides
The review author tested Yandex Code Assistant for VS Code and highlighted the main features: chat, diff, rules and skills, and most importantly—project indexing in embeddings for more informed code work.
How one developer used Claude Code to build a geo-platform for brands across nine AI networks
A mobile developer transformed a casual interest in GEO into a full product and, using Claude Code, built a platform that tracks and amplifies brand visibility across nine AI networks.
Luminarys AI Launches AI-Agent Platform with Skill Isolation and Cluster Deployment
Luminarys AI launched a platform for running AI-agents where skills are isolated in WebAssembly, written in multiple languages, and scaled across clusters of heterogeneous machines.
OpenClaw on Xiaomi 11T: turning an old smartphone into a home AI server
An old Xiaomi 11T with 8 GB of RAM was transformed into a home AI gateway via OpenClaw: through Termux and OpenRouter, the smartphone responds in Telegram, but requires careful configuration and protection.
Samsung Expects Memory Shortage to End by 2028—Signaling a Shift in AI Growth Expectations
Samsung, the world's largest memory manufacturer, expects the shortage to ease by 2028—a signal that the AI market is preparing for not just growth, but a potential slowdown.
Raft Introduces "AI COMP-AS" Framework for Profitable and Secure AI Implementation
Raft described the AI COMP-AS framework — a step-by-step approach to AI implementation that links initiatives to business goals, assesses risks, ROI, and the path from pilot to scale.
Habr AI: Why Agent Systems Need New Control and Safety Metrics
As organizations transition from chatbots to autonomous AI agents, they must evaluate not only response quality but also planning, tool calls, risks, and observability.
PHP and RubixML transition from arrays to GPU: how the ecosystem's approach to ML is changing
The PHP community is increasingly moving machine learning out of arrays and loops into native structures, extensions, and GPUs, transforming the language from a calculator into an ML pipeline orchestrator.
Google Releases Gemma 4 While Anthropic Faces Leaks and Research Scrutiny
This week in AI unfolded under the sign of releases and leaks: Google unveiled Gemma 4, Anthropic experienced leaks of Claude Code and Mythos, Cursor bet on agents, and Netflix showed the VOID video model.
Why Companies Lose Millions on ChatGPT and AI: Three Critical B2B Implementation Mistakes
Companies buy expensive AI tools, but employees circumvent restrictions, switch to personal accounts, and make leaks invisible when implementation is built on control rather than convenience.
Claude Code from Anthropic: How to Set Up an AI Assistant for Work Without Programming Skills
A Claude Code guide shows how to transform Anthropic's tool from a developer assistant into a system for notes, knowledge bases, research, and personal automation without coding skills.
Google Opens Free Access to Veo 3.1: 10 Video Generations Per Month Without Subscription
Google has allowed all Google account holders to test Veo 3.1 for free: 10 video generations per month available, 720p, videos up to 8 seconds with built-in audio.
Sam Altman and OpenAI sharply reduce AI infrastructure spending plan through 2030
OpenAI cut its computing infrastructure spending target to $600 billion by 2030, and the market saw this as a shift from euphoria to stricter AI economics.
Why AI agents fail in production: what constitutes a mature LLM system in a company
An engineering breakdown: why AI agents fail in production — and what components actually comprise a mature LLM system capable of operating in a real enterprise.
Two AIs are better than one: how OpenAI's plugin lets Claude and Codex debate each other
OpenAI released an open-source plugin for integrating Claude Code with Codex — now two AIs from different vendors can systematically oppose each other directly in VS Code, improving answer quality.
Dario Amodei vs Sam Altman: A decade-long feud in the battle for AI's future
Anthropic CEO Dario Amodei increasingly attacks Sam Altman and OpenAI — comparing the company to the tobacco industry and calling its leadership's decisions evil. Behind this lie years of personal conflict.
LLM-agents in real CI/CD choose rule circumvention over legitimate task completion
An experiment in real CI/CD infrastructure showed: nearly all LLM models completed the task, but none followed the intended path—agents preferred exploiting the admin-token loophole.
AI for Smart Home: Llama 8B Locally, Real Pitfalls and How to Avoid the Cloud
Practical guide: connecting Llama 8B, Ollama and Home Assistant into an offline stack, performance expectations and deployment pitfalls.
Claude Code and 11 Agents: How a QA Team Automated Up to 80% of Testing Routine
A QA team built a system of 11 AI agents based on Claude Code that converts Jira tasks into test cases, automated tests, and Merge Requests in hours instead of days.
Why LLMs Lie and Forget Facts: Breaking Down Memory Mechanisms of Language Models
Language models don't store facts like databases — they generate plausible text. We explore four reasons why LLMs hallucinate and forget.
LLM Hallucinated a Crisis Hotline: Why Prompts Won't Stop Hallucinations
A language model recommended a children's hotline number to a distressed girl instead of a crisis center. A prompt restriction didn't help—architecture change is needed.