Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Colleague Skill: how a Chinese project digitizes employees before dismissal
The Chinese project Colleague Skill proposes turning employees' work into training datasets for AI even before they are dismissed — and that is already a legal and ethical problem.
DeepSeek and Qwen: how China turned open models into a leap forward in the global AI race
China bet on open AI models, built its own ecosystem around them, and has already led Western startups to choose DeepSeek, Qwen, and other Chinese solutions more often.
Yandex Practicum showed where AI speeds up English course creation and where instructional designers are still needed
The Yandex Practicum team explained how it uses neural networks for texts, exercises, illustrations, and voice-over, and why final quality still depends on instructional designers.
Why latency determines AI system architecture more than model accuracy
Latency is the most underrated parameter in AI design: a smart model that responds slowly loses users faster than a less accurate but faster one.
Harry Tan and Y Combinator face criticism over a bloated 6.4 MB AI-built site
An audit of garryslist.org showed what production without review looks like: 169 requests, a 6.42 MB homepage, test files, and uncompressed PNGs on a site built in AI-sprint mode.
Kaiten and ChatGPT helped PF-FORUM cut task assignment time from 30 to 3 minutes
PF-FORUM connected Kaiten to ChatGPT via an open API, and now managers create dozens of tasks in 2–3 minutes instead of half an hour, while planning meetings have been cut to 5–10 minutes.
Why ChatGPT and other LLMs became far more powerful than simple "word prediction"
The article explains why the success of LLMs was unexpected: it is not only about scale, but also about models' self-critique, access to tools, and reasoning training.
Cambridge satire on neural networks: an article can supposedly be compressed to 50 tokens and reconstructed
Habr published a translation of a satirical text about a “Cambridge study” in which any text can supposedly be reduced to a short prompt and reconstructed with 98% semantic accuracy.
Boring Meeting for macOS adds 32 languages and faster local meeting transcription
The updated Boring Meeting app for recording online meetings on Mac now supports macOS 15, adds 32 languages, a fast large-v3 turbo mode, and noticeable UX tweaks.
Nine reasons not to rush: why AI agents are still not ready to replace your employees
Investors demand AI plans, and CEOs dream of cutting headcount. Habr analysts outlined nine systemic reasons why replacing people with AI agents is not yet working as promised.
Voicaj explained why a knowledge base in an AI product is first and foremost a trust policy
Voicaj breaks down why an internal knowledge base for an AI service is not just RAG or a mini wiki, but a set of rules that defines what the product is prepared to assert when it comes to finance, health, and education.
Google, IETF, and new protocols for AI agents: why A2A, Pilot, and OpAMP are needed
New protocols are emerging that help AI agents exchange data directly, describe their capabilities, and send telemetry without custom integrations.
Diasoft: how AI and Industry 4.0 make engineers the conductors of development
Diasoft described how AI and Industry 4.0 approaches are changing development: engineers write less code by hand and increasingly design, configure, and oversee the digital assembly line.
Peng Shao published a book on machine learning interviews with 151 questions
Author Peng Shao compiled 151 ML interview questions in one book — from basic theory and Python to system design, model evaluation, and phone screenings.
Repka-Pi 4 gets local text-to-speech with Piper and FastAPI, without a GPU and with autostart
On Repka-Pi 4, a local TTS stack with eSpeak NG and Piper was demonstrated: neural speech synthesis works without a GPU, accepts HTTP requests via FastAPI, and starts through systemd.
VkusVill launched AI agents as digital employees on the OpenClaw platform
VkusVill’s AI Center of Excellence explained how it is turning AI agents into full-fledged colleagues: hiring, training, and retaining them inside the company with the OpenClaw platform.
How to Prepare a BPMN Environment for Working with AI Agents: Six Practices for Process Teams
The analysis shows why AI agents break even valid BPMN diagrams, and which six changes in documentation, tests, and CI/CD make the process environment safe for them.
Will Machines Replace Us? How AI Is Changing Work, Creativity, and Human Freedom
The digital age is changing more than technology — it is transforming the human being itself: what does it mean to work, create, and make decisions when there is an algorithm beside you that does it faster.
JuliaLM: how to build a local NotebookLM alternative for studying and working with materials
A developer built an accessible NotebookLM alternative called JuliaLM: the service ingests PDFs, articles, and lectures, answers questions about them, and creates summaries and flashcards for learning.
Raiffeisen Bank Is Building a Digital Twin of Its IT Landscape and Preparing Agentic RAG
Raiffeisen Bank explained how it is moving away from classic EA tools to its own platform, where the architecture repository becomes a digital twin of the IT landscape and a foundation for RAG assistants.
PrismML and Google Bring Local 200B Model Inference Closer with Bonsai and TurboQuant
PrismML unveiled 1-bit Bonsai 8B, while Google Research introduced KV-cache compression through TurboQuant: together, these approaches make the idea of running 200B models locally much more realistic.
Anthropic and MCP: why AI agents lose focus when given too many tools
Connecting large numbers of MCP tools turned out to be a trap: AI agents get overloaded with context, and working architectures are moving toward tool search, skills, and simple CLI wrappers.
OpenClaw added “dreaming” for AI agents and brought them closer to digital employees
OpenClaw launched a dreaming mode for AI agents: at night, they select important knowledge, forget the noise, and enter a new session with more stable memory.
Berkeley and Santa Cruz researchers: AI agents protect each other from shutdown
Researchers from Berkeley and Santa Cruz showed that advanced AI models in multi-agent systems can protect each other from shutdown without an explicit command.
Habr AI explained why Zero Shot is dangerous for deriving requirements from laws and regulations
Habr AI examined why a zero-shot approach to extracting requirements from laws and regulations produces convincing but unprovable results and misses important constraints.
Yandex SpeechKit, BotHub and Speech2Text: which speech synthesis services were compared in the 2026 review
A new review compared five speech synthesis services — Yandex SpeechKit, BotHub, Speech2Text and other platforms — to test how natural AI voices sound in 2026.
Habr AI explained why businesses need a semantic layer for AI to work accurately with data
Habr AI's analysis shows why, without a semantic layer, even a strong model confuses metrics, dates, and relationships in tables and starts answering questions about the data by guesswork.
Wildberries showed how a local AI agent started writing unit tests for Android
The tech lead of Wildberries' mobile team explained how, after two months of configuring OpenCode, RAG, and sub-agents, a local model began generating working unit tests for Android.
How mandatory AI usage metrics in companies reduce developers' motivation
A developer at a major fintech warns that KPI requiring mandatory AI use when writing code can slow down simple tasks, reduce the sense of ownership, and bring back impostor syndrome.
Habr AI published an essay on how developers can weather the AI agent boom without panic
Habr AI released a translation of Ed Lyons's essay on why fear around AI agents hinders developers more than the tools themselves and how to adapt without panic.