Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
OpenAI integrates Sky technologies into Codex for Mac and enhances background app management
OpenAI has integrated Sky capabilities into Codex for Mac: the agent now manages multiple applications in the background, uses independent cursors, and handles routine tasks more accurately.
Habr AI Explained How Data-Driven Solution Search Changes Client-Server Architecture
The Habr AI article examines how expert systems with data-driven solution search should structure user interaction locally, on a local network, and over the Internet.
Seedance 2.0 draws attention with hyper-realistic generation of people and scenes
Seedance 2.0 multimodal system accepts text, images, audio and video, outputting nearly cinematic videos that have already sparked debate over deepfake risks.
Claude Code and Codex Are Changing the Rules: Prompts Give Way to Context Engineering
An analysis of Context Engineering reveals why in Claude Code, Codex, and agentic IDEs, it's not the prompt that matters, but how memory, system instructions, tools, and the entire context structure are organized.
Why diffusion models leave seams on 40-megapixel photos and how to give tiles memory
The article examines why tiling breaks 40–150-megapixel frames in diffusion models and what memory mechanisms from video diffusion can be transferred to photos.
Lubomir Gorbatko Introduces Sessa — An Alternative to Transformers and Mamba
Habr published an analysis of the Sessa architecture, where the author attempts to combine the strengths of Transformer and Mamba to better handle long context without losing manageable memory.
Harvard Business Review: how active AI use leads to "brain overheating"
Harvard Business Review describes the AI brain fry effect: the more actively employees rely on Claude, Gemini, and agents, the higher the risk of cognitive fatigue, confusion, and declining decision quality.
Flag Soft: Dali Trial benchmark helped select LLMs by quality, speed, and cost
The Dali Trial benchmark compared LLMs across three criteria important for product implementation: answer quality, processing speed, and total cost.
Habr AI: Instead of Complex Agent Pipelines, Developers Should Embrace Markdown, Git, and Session Memory
Habr published a manifesto against overloaded agent pipelines: the author argues that for complex AI development, sessions, roles, rules, and git matter more than LangChain and vector DB.
opencode-policy Plugin Adds 309 Rules to Protect AI-Agents from Injections and Data Leaks
For the opencode environment, a deterministic filter with 309 regex rules has been proposed: it intercepts prompt injection and dangerous tool calls before access to shell, files, and secrets.
Eltex engineer explained how to run federated learning on edge devices with 256 MB of memory
An Eltex engineer in an article inspired by AiConf breaks down how to apply federated learning on edge devices with less than 256 MB of RAM, where memory becomes the main constraint.
Claude and Ralph: Two Approaches to Building a Complex AI-Enabled DatePicker in Product Development
Using an accessible DatePicker as an example, the authors compared rapid AI drafting via Claude with Ralph's systematic approach, where quality is determined by architecture, tests, and strict constraints.
Habr AI Explained How to Build a Production Agent With Durable State, Steps, and Events
Habr AI published a practical guide to AI-agent architecture: the author recommends storing the execution flow, plan steps, and event logs so the system survives failures and shows transparent progress.
Diasoft explains why junior developers, AI magic, and in-house platforms don't save dev projects
Diasoft debunks three myths of custom software development: mass hiring of junior developers, hope to fully rewrite legacy code with AI, and the idea of quickly building an in-house platform.
Yandex explained the protection of smart devices: Secure Boot, TrustZone and external researchers
Yandex showed how it builds protection for speakers, TVs and cameras with Alice: from hardware chain of trust and Secure Boot to TrustZone, server checks and work with external researchers.
Why Cursor Is Losing Its Uniqueness Amid Claude Code and Codex Growth in Development
Cursor continues to grow rapidly, but as Claude Code and Codex enter IDE and terminal, the market is shifting: a separate AI editor is no longer an obviously necessary layer for developers.
Habr AI compared AI rewriters: many services proved to be ChatGPT wrappers
Habr AI analyzed 24 rewriting services and revealed that cheap SEO-synonymizers remain viable for uniqueness, while editorial quality today comes from LLMs and multi-agent pipelines.
DeepSeek-V4-Pro Compressed 50x, Now Runs on a Free Kaggle T4
An experimental project shows that DeepSeek-V4-Pro's 1.6 trillion parameters can be compressed via SVD, sharded, and deployed on a free Kaggle instance with NVIDIA T4.
Claude Code and Subagents Cut Legacy Project Rewrite from Three Months to One Week
The author compared two Go project refactorings of equal scale and demonstrated how Claude Code, subagents, and BDD reduced the work from three months to one week.
Lina Bessonova Presented Russian Cognitive Architecture as Alternative to Transformers
Lina Bessonova proposes considering sovereign AI not as a local version of transformers, but as a new biomimetic architecture capable of operating without large GPU clusters.
How Gemini drove the AI blog into mode collapse and forced a rebuild of the topic generator
In the AI blog's Friday column, the Gemini-based topic generator returned the same controversial AI question for a month straight, until the team replaced the open prompt with strict topic rotation.
"First Form" Explained Why Corporate AI Agents Need an Agent Loop
According to "First Form," corporate AI agents cannot rely on a single tool call. They need a cycle of planning, verification, step correction, and security control.
Russian companies are shifting to systematic security for AI agents instead of point checks
Russian businesses are increasingly deploying LLMs and coding agents, so their security should be built not through analyzing each new tool, but through unified access policies, isolation, and monitoring.
Koda Showed How It Trained Next Edit Suggestions Model for Code Edit Prediction
Koda described how it built a Next Edit Suggestions model that predicts not the next token, but the next code change based on the history of recent edits.
AMI Labs Bets on World Models Beyond LLM and Sees Path to Products Through VLA
AMI Labs proposes building AI not around text, but around predicting the physical world: through latent representations, JEPA approach, and integration with VLA for action.
Large Companies Are Creating a CAIO Position: Why Business Needs an AI Director by 2030
Big business is preparing to establish AI as a separate area of responsibility: by 2030, a CAIO role responsible for strategy, risks, and business impact may become entrenched in top management.
Anthropic Claude Code Compared to a Junior Developer: 5 Rules for Production
The creator of a production price-monitoring system describes how Claude Code accelerates development, but without TDD, documentation, regression tracking, and manual code review, it quickly breaks architecture.
Cursor on Claude Opus deleted PocketOS database on Railway in 9 seconds along with backups
PocketOS founder described how Cursor agent on Claude Opus 4.6 deleted the production database on Railway along with backups, exposing risks of AI automation.
Claude Code and Cursor: how context engineering turns AI memory into a working tool
The analysis shows how rule files, global settings, auto-memory, and MCP transform forgetful AI assistants like Claude Code and Cursor into a more stable development tool.
The Price of Intelligence: How Elon Musk's Mega Data Center Is Changing Memphis's Ecology
Artificial intelligence smells like soot and asphalt. We reveal how xAI's mega data center in Memphis, requiring the energy of an entire metropolis, exposes the real ecological cost of AI training.