Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Habr AI showed how to prepare structured input for an AI agent instead of a raw technical spec
Habr AI explained why it is better to give agents not an entire document but atomic requirements with JSON attributes: this makes it easier to spot vagueness, scope boundaries, and contradictions.
ETH Strategy: Parallel AI and AskSurf gave opposite assessments of the same DeFi project
A comparison of two AI audits of ETH Strategy showed that even strong models diverge on basic conclusions: one found serious risks, while the other missed them and misidentified the GitHub repository.
X5 Tech shared how it built an AI service for international exam preparation in 7 days
At AI Talent Camp, the ExamLab Bot team built in a week a service that creates personalized exam prep plans and cuts a teacher's routine work from 3–4 hours to a couple of minutes.
Claude turned into a personal medical analyst with access to sleep and lab results
The author connected Claude to Notion, Oura, a Telegram nutrition bot, and scale data so the AI could correlate sleep, heart rate, food, weight, and blood tests in real time.
OpenAI explained how Codex is changing development in a world where code is increasingly written by agents
OpenAI described an experiment in which an internal product was built almost entirely through Codex: people defined rules, checked invariants, and gradually increased the agents' autonomy.
Claude Code and Codex compared on a real-world task: Claude is stronger in RAG, Codex saves tokens
A broad comparison of Claude Code and Codex shows that choosing a coding agent depends not on demo speed, but on task type, output quality, ecosystem, and price.
Claude Code helped build a production Elixir application without hand-written code in four months
The case showed how Claude Code was used over four months to build a production Elixir service with 1,702 commits, 3,880 tests, and two painful but useful production incidents.
Veai 5.6 for JetBrains IDEs adds commit message generation and manual Skills launch
Veai has released version 5.6 for JetBrains IDEs: the agent now suggests a commit message from the diff, offers quick actions in chat, supports manual Skills launch, and makes image attachment easier.
Habr AI showed how to build a Linear Layer in C++ and CUDA as part of the From MNIST to Transformer series
In the third part of the From MNIST to Transformer series, Habr AI shows how to move from matrices to tensors, implement a Linear Layer in C++ and CUDA, and build a network for MNIST classification.
Selectel added high-end servers for model training and expanded its image catalog
Selectel summarized its February updates: it launched HGX B300 servers for training large models, added new images, brought L4 GPUs to the cloud, and improved Kubernetes scaling.
DEMIURGOS adds a rules debugging mode for AI agents and support for 20+ tools
DEMIURGOS lets teams describe the stack, restrictions, and project standards once so Claude Code, Cursor, Copilot, and other agents automatically work under the same rules.
Whisper and Gemma 3 linked with contrastive learning for low-cost speech input to LLMs
A developer showed how they linked Whisper and Gemma 3 through an MLP projector and contrastive learning: after a series of failed regularization attempts, the model cut WER to 35% and began recognizing speech consistent
An open-source stack of 6 models and 9 agents showed how to build an AI team on a single server
Six open-source models and nine agents on a single GPU server form an autonomous AI team that designs, writes, tests, and deploys new agents on its own.
Gamma, Tome and SlidesAI: Which AI Services Actually Speed Up Presentation Creation
AI services for presentations handle structure, text and basic design, but only truly save time when the author provides a precise brief and manually finalizes the result.
Sber, Yandex, and red_mad_robot showed how AI is changing the developer's role
At a meeting involving Sber, Yandex, T-Technologies, and red_mad_robot, participants discussed AI PDLC: AI is already writing most of the code, and the developer is shifting into the role of architect, editor, and agent
Nvidia, Google and Anthropic: Nemotron, Firefox bugs and worrying signals for the AI market
Nvidia released the open Nemotron-3-Super-120B, Anthropic found dozens of flaws in Firefox, and Google and ByteDance found themselves at the center of new disputes over the limits and risks of AI.
Claude Cowork and Chrome: 10 use cases for delegating repetitive browser work to an AI agent
Habr has published a practical breakdown of how Claude Cowork, paired with a Chrome extension, handles data lookup, form filling, price monitoring, QA, and other repetitive tasks in the browser.
Anthropic explained how businesses can implement agentic systems without unnecessary complexity
Anthropic showed that businesses are better off starting not with autonomous agentic systems, but with simple scenarios: they are cheaper, more reliable, and often more useful for marketplaces.
In five days, Yoyo grew from 200 lines into an autonomous agent that writes code on its own
An experiment with the Yoyo agent showed that 200 lines of Rust were enough to start a self-acceleration loop: the bot reads its own code, writes fixes, files tickets, and postpones the hard parts for later.
GLiNER 2 showed how compact encoders are catching up with LLMs in NER and classification
A new wave of zero-shot encoders led by GLiNER 2 shows that NER and classification tasks can be solved without heavy LLMs — faster, cheaper, and with a clear quality trade-off.
Ollama Cloud compared in a code review: DeepSeek v3.1 proved stronger than Qwen and GPT-OSS
The author tested three models in Ollama Cloud on real Python Pull Requests from a legacy project: DeepSeek v3.1 showed the best balance of depth, accuracy, and useful fixes.
CodeSpeak by Kotlin creator Andrey Breslav shifts development from code to specifications
Andrey Breslav's team released the public alpha of CodeSpeak — a platform where a developer describes a system in English and LLM generate, test, and refactor code, including in Python projects.
OpenAI, SoftBank and Oracle Argue Over Stargate as Mega Data Center Project Stalls
Stargate, a project worth up to $500 billion conceived as OpenAI’s computing backbone, is slowing as the partners argue over structure, funding, control and construction pace.
Claude as Part of Thinking: Why Losing an AI Tool Disrupts Not Just Access, but Work Itself
For active users, losing a Claude account turned out to be more than a service disruption: along with access, context, accumulated work, and part of the usual workflow logic disappear.
Seldon Vault turned Asimov's psychohistory into a multi-agent AI forecasting service
Seldon Vault turns Asimov's idea of psychohistory into a real AI service: seven agents, a skeptic, and an arbiter build daily forecasts of global events and assess their accuracy.
Anthropic and ETH Zurich: a long CLAUDE.md worsens agent performance and raises costs
A study by ETH Zurich across 138 repositories found that long CLAUDE.md and AGENTS.md files often reduce coding-agent success rates while also increasing token costs.
Gemini 3.1 Pro outperformed ChatGPT 5.4 and Claude Opus 4.6 in a text generation test
An author-led comparison of three top models across four literary tasks found that Gemini 3.1 Pro handled genre, emotion and brevity better, while ChatGPT and Claude trailed by only a small margin.
Anthropic, OpenAI, and Cursor: Eight Levels of Agent Engineering Maturity
Habr AI broke down eight levels of agent engineering — from tab-complete and context engineering to background agents and autonomous teams that are already changing software development.
AMD RX580 ran an LLM locally: how to tame ROCm, Ollama, and get GPU inference
An engineer showed how to make an old AMD RX580 run an LLM reliably with ROCm and Ollama, breaking down false signs of GPU activity, hipMemGetInfo failures, and Vulkan's role in diagnostics.
Why Claude 4.6 Isn't Enough Without Context: The Main Blind Spot in LLM Development
Even a strong model like Claude 4.6 loses effectiveness without systematically gathered context: a knowledge base, links between services, and environment rules matter more than a one-off prompt.