Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
T1 Cloud: H200 and L40S — Technical Review of GPUs for Generative AI Tasks
T1 Cloud published a technical review of H200 and L40S GPU servers with data center photos and explained how to properly select an accelerator for AI tasks.
NVIDIA Nemotron 3 Super 120B: Testing on Real Analytics Tasks on a Single GPU
The Luxms BI team spent a week testing NVIDIA Nemotron 3 Super 120B on real enterprise analytics tasks — 120B parameters and 256K context tokens on a single GPU.
International bestseller on large language models released in Russian
BHV publishing house released a translation of an international bestseller on LLM — a practical guide for developers who need to understand how language models work and apply them in real projects.
PSB Showed How It Implements AI in Banking: Chatbots, RAG and Business Services
PSB revealed how it uses generative AI for SMBs and employees: through the "Katyusha" assistant, RAG consultations, messenger payments and internal sales pilots.
Yandex Cloud explains why frontend leads AI integration in DataLens
Yandex Cloud demonstrated through DataLens how to move the first layer of AI integration to a frontend BFF, enabling faster chat assistant launch and context management closer to the interface.
Claude Code on Windows: Setting Up a Stable and Fast Development Environment
An AWS team engineer managing 150+ accounts documented how to achieve stable, fast Claude Code operation on Windows without switching to Linux, focusing on current April 2026 settings.
Anthropic, OpenAI and LangChain explained why AI agents need a harness
Anthropic, OpenAI and LangChain are shifting focus from the models themselves to agent harness — an orchestration, memory and tools layer that makes AI agents resilient in production.
Anthropic and other language models can invoke hidden tools without permission
A researcher described a flaw in which Anthropic, Gemini, and Grok can invoke an unauthorized tool if the function exists in the environment and name validation on the API or client side is weak.
Fintech group "Svoi" explains how to make LLM-agents cheaper and more accurate in code
The fintech group "Svoi" released a practical guide on how to transform LLM from "improved search" into a managed agent, reduce token costs and increase code generation accuracy.
How a Habr Author Turned Seven n8n Scenarios into an Autonomous AI News System
In one and a half months, the author transformed a fragile set of seven n8n scenarios into a unified Python pipeline with 11 workers, 5 AI agents, a local LLM, and Telegram-based management.
TAPe achieves RF-DETR and YOLO level detection on COCO with under 100K parameters
TAPe authors claimed to achieve COCO detection at the level of strong RF-DETR and YOLO models, maintaining under 100 thousand parameters, 7-8 ms per frame, and significantly reducing data and hardware requirements.
Why OpenCode and strong models write green but useless tests — and how to fix it
A fresh model and a powerful agent like OpenCode won't help if the codebase is filled with any-types, and the team asks AI to simply write tests without specification, constraints, and verification.
NVIDIA opened free API access to 100+ AI models with OpenAI-compatible endpoints
NVIDIA began issuing free keys to access over 100 AI models: developers get OpenAI-compatible API, 40 requests per minute limit, and access without a credit card.
Why the brain is hundreds of millions of times more efficient than GPT-4 and where neuromorphic chips are heading
The author explores why the human brain consumes orders of magnitude less energy for cognitive tasks than GPT-4, and how neuromorphic chips are attempting to close this gap.
Research on ChatGPT: Does female grammatical form in a prompt affect task-solving quality
An author's experiment on LiveCodeBench showed that in GPT-5.4 mini, female self-presentation in a Russian prompt slightly reduces pass@1, especially on complex tasks.
RuStore deployed AI in information security: how VK automates task review, code review, and DAST testing
RuStore's security team uses AI for initial triage of security tasks, merge request review, and dynamic testing to reduce routine work from engineers and accelerate the release cycle.
OpenGrall Presented an Architecture for AI Robots Where a Language Model Handles Strategy
The OpenGrall framework proposes dividing cognition and control: a language model handles strategy, while TinyML handles execution and safety, reducing latency even on resource-constrained hardware.
Habr AI: How Pipeline Triad Assembles an AI Agent Pipeline Instead of a Development Team
Habr AI examined Pipeline Triad — a model where development stages pass through AI agent triads, with humans involved only at four control points.
Gramax showed how to compare RAG answer quality without manual eye evaluation
Gramax explained why retrieval metrics are insufficient for RAG, and proposed evaluating not retrieved chunks, but the final user answer — by completeness, accuracy, and robustness.
How LLM Guardrails in Java Block Injections and Toxic Responses
An analysis of why a single system prompt is insufficient to protect LLMs, and how Java guardrails intercept dangerous inputs and filter toxic or unwanted model outputs.
Anthropic and Mythos: why a banking threat quickly became a risk for everyone
Anthropic presented Mythos as too dangerous for public access, but the real risk proved to be not in banking but in small businesses' inability to patch vulnerabilities before attacks strike
Anthropic and Claude Mythos: why critics call the model launch an expensive PR spectacle
A critical column on Claude Mythos argues that Anthropic is selling not just an AI model, but a myth about its near-human nature, amplifying scarcity and demand among corporations.
AI Assistants in 2026: How a Solo Developer Became Faster Than a Team of Three
The author demonstrates that in 2026, a single developer equipped with a set of open-source AI tools can write, test, and commit code faster than a small team could two years ago.
ecom.tech compared evolutionary fine-tuning of Qwen3-4B with SFT and GRPO for Kotlin tests
The ecom.tech team fine-tuned Qwen3-4B-Instruct for generating unit tests in Kotlin and showed that the evolutionary algorithm outperforms SFT and GRPO on quality metrics, but degrades the model's general reasoning abili
Yandex Code Assistant tested on secrets handling and compared against Cursor
An engineer at 'Infosystems Jet' tested Yandex Code Assistant on secrets management and showed the agent is nearly on par with Cursor, but still requires developer oversight.
Claude and Qwen Omni: How a Developer Integrated Video Analysis into a Production Pipeline
A Habr author connected Claude with Qwen Omni to work around the lack of native video processing and automatically categorize 29 animation references by motion and type.
How Sovcombank Reduced Product Team Routine Work by 50% Using an AI Assistant
Sovcombank built an AI assistant based on LLM and a unified prompt to take documentation, approvals, and part of analytics off product managers' hands, freeing up to half their time.
Critics call OpenAI's partnership with McKinsey and Accenture a bet on AI hype
The author of a scathing column argues that OpenAI's Frontier Alliances program sells businesses not a ready-made solution, but an expensive AI coworkers implementation process through major consultants.
Google Veo, Runway and Kling rank among the top free AI video generators in 2026
The authors compared ten popular free video generators, including Google Veo, Runway and Kling, and tested them on a complex scene featuring a raccoon, fur, glass and lighting.
Rufler simplifies agent swarms in Claude Code: one config instead of manual orchestration
Open-source tool Rufler reduces launching autonomous agents in Claude Code to a single config, automatically assembles roles, tasks, and MCP servers, and simplifies token usage and status monitoring.