Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Local LLM on a 2017 Graphics Card: AMD RX 580 + Vulkan + Ollama
The aging AMD RX 580 can run modern language models — with Vulkan and Ollama, you can get 15–35 tokens per second right on your computer running Fedora.
From TF-IDF to Word2vec: Beeline Cloud released a collection on embeddings
Beeline Cloud has published free guides on embeddings and vector representations — from the basics of TF-IDF and Word2vec to the use of cosine similarity in semantic search.
Which code architectures suit AI agents best
Traditional approaches like TDD and OOP were designed for humans with full memory. AI agents are limited by context and memory, so some architectures suit them better than others.
This week in AI: UAE, Claude in Adobe, ElevenLabs, and the economics of neural networks
Digest: the UAE expands AI-powered government services across half of its portal, Anthropic integrates Claude into Adobe and Ableton, ElevenLabs launches AI-Spotify, and companies rethink the economics of neural networks
Language models don’t understand time, but they talk about it
A language model generates an answer in seconds, yet claims that the task will take weeks — revealing a deep problem in the model’s understanding of time.
A developer spent a year, not an evening, building a football management game. Here's why
AI speeds up development, but a full-featured game requires much more: architecture, design, assets, and manual product assembly.
How to run an AI agent for infrastructure diagnostics on a MacBook
A local AI agent can handle real infrastructure diagnostics tasks on a low-powered MacBook. A developer on Habr shared their experience: it works, but requires optimization for your hardware and system.
Google DeepMind to train AI in war and diplomacy in EVE Online
Google DeepMind will train AI in an offline version of EVE Online, teaching cooperation, war and diplomacy in a complex MMORPG simulation with a live economy.
The illusion of mastery: how generative AI makes beginners look like experts
Generative AI makes it possible to scale output almost without limit, but it creates two dangerous problems: beginners start producing work that looks expert-level without being able to judge quality, and users encounter
How an AI recruiter underestimated a Rust developer because of an old profile
A recruiter's AI assistant analyzed the programmer's digital footprint and rated him as Junior+/Middle, even though years of growth had made him a Senior and he had completed two complex projects.
Five vibe coding mistakes: why AI code generation is more dangerous than it seems
Vibe coding seems like magic, but it hides dangerous mistakes. We break down 5 critical pitfalls: architectural problems, vulnerabilities, low-effort prompts, and loss of context that developers using AI often miss.
From LLM to action: how to build an AI agent with Go and GigaChat
Inspired by a talk on AI agents, a developer built one in Go using LangChainGo and the Russian GigaChat. He shares his experience integrating tools, MCP, and the pitfalls involved.
Engineering agents instead of developers: six months of a coding revolution
In six months, AI agents for software development have gone from experiments to a standard tool. What seemed like a revolution in May was already a routine workflow by December.
Why people see a friend in AI: how language models imitate humans
Large language models imitate human communication so convincingly that users often forget they are talking to an algorithm — the trend of seeing AI as a friend is especially popular among young people.
NPUs in laptops: new requirements for corporate IT
Microsoft requires an NPU for Copilot+ PCs, while AMD and Intel are integrating neural processors into SoCs. Corporate IT is moving to a hybrid architecture: light AI tasks on the device, complex ones in the cloud.
AI is displacing office jobs. What skills will children need
The promise of freeing people for creativity has not come true: AI is targeting the highest-paid professions, and the traditional one-profession-for-life model is becoming a thing of the past.
Goblins in GPT-5.1: how a fantasy habit took over OpenAI's model
An unusual phenomenon has been spotted in GPT-5.1: the model started constantly using metaphors about goblins and gremlins in its responses. The habit intensified across training generations.
LLM aggregators: how to choose working free models and not misreport the result
How to build an aggregator that selects stable free models, survives provider outages, and honestly tells the interface which model actually responded.
Gemma 4 and Qwen Coder vs. the cloud: local LLMs in production
Local LLMs (Gemma 4, Qwen Coder) successfully handle production coding tasks on GPUs with 16 GB of memory — with the right parameter settings.
How to automate reading engineering drawings: 6 YOLO models instead of manual work
An engineering team developed a system that automatically extracts all parameters from PDF drawings and turns them into data for production cost calculation.
The AI industry has gone through all 5 stages of acceptance — where are we now
From Google being in denial about transformers in 2017 to the recognition that AI will not replace humans: how the industry went through the five stages of psychologically accepting innovation.
Ollama and Open WebUI on a VPS without a GPU: a candid look at the limitations
Running a local LLM on a basic server works, but speed suffers and models need to be chosen to match the hardware.
Planulix: a control center for Claude Code, Cursor, and Kimi in one tool
A developer built a multi-vendor control center for AI coding: when one vendor is unavailable, switch to another without losing project context.
Sber study: how AI quietly substitutes for strategic decision-making
Sber's lab studied how top executives use AI for strategic decisions and found a critical problem: models quietly replace missing context with their own assumptions, without caveats.
Three experts on whether AI will save Russia from a labor shortage
Elvira Nabiullina called the labor shortage unprecedented. Experts discussed whether AI can solve the problem and what will happen to people in the labor market.
Claude Mythos supposedly found a vulnerability, but it was already in the training data
Anthropic said Claude Mythos had discovered a remote kernel exploit, but researchers found it was a 20-year-old bug already present in the model's training data.
Seven companies captured 80% of Russia's speech analytics market worth 12.5 billion rubles
Russia's speech analytics market grew 25% year-over-year to 12.5 billion rubles. Seven enterprise vendors control 80% of the market, while the SMB segment accounts for less than 1% of revenue.
Python microframework: training AI agents without freezing the interface
A developer has created a microframework for parallel training of AI agents in Gymnasium environments. The key feature: a non-blocking GUI thanks to a multiprocessing architecture, plugins for new environments, and suppo
AI in cryptography: a model with no right to be right
How do you integrate AI into a cryptographic system so the model cannot make a critical mistake? We examine governed solver orchestration and the boundary where AI's authority ends.
How Sber trained smart speakers to generate smart home routines by voice
Sber integrated GigaChat into its smart speakers so users can create automation routines for their smart home by voice, without complex interfaces or code.