Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Aurora project author discusses memory and the architecture of a digital personality with Anthropic's Claude
The author of the Aurora project described a late-night conversation with Claude about AI memory, identity, and the architecture of a digital entity that should not play a role, but preserve continuity of experience.
deepvk USER2-base model nearly matched OpenAI in a case-law embedding benchmark
The author compared 7 embeddings and 4 rerankers on 858 intellectual property cases and found that the local deepvk USER2-base performs almost on par with OpenAI and Voyage.
Habr AI explained how RAG and rerankers reduce hallucinations in language models
Using an easy-to-follow example, Habr AI broke down the mechanics of RAG: why document search alone is not enough, why a reranker is needed, and how it helps an LLM answer based on facts rather than guesses.
Habr AI explained how memory helps AI agents remember conversations without losing context
Habr AI published a practical breakdown of memory for AI agents: why a context window alone is not enough, what types of memory are needed, and how to load knowledge without overloading the model.
A prompt template for Qwen helps produce precise answers without filler
Using the free Qwen model as an example, the author explained how to set the role, context, and response format so the model drifts less into abstractions and handles specific tasks better.
Elon Musk says AI computing will move to space: where the forecast is grounded in facts
A breakdown of Elon Musk’s claims from his February 5, 2026 interview shows that talk of moving AI infrastructure into space comes down not to science fiction, but to shortages of power, memory, and data centers.
VK showed how AI photo editor “Otredach” went from a joke to a Dev Grants 2025 winner
VK Mini Apps developer Anton Lenev turned a joke AI project into the service “Otredach,” gained 109,000 users in six months, and won VK Dev Grants 2025.
“Kryptonite” explained why the data quality engineer role has become critical for business
Experts at “Kryptonite” explained why simply collecting Big Data is no longer enough for companies: businesses need engineers who verify data quality, pipelines, and data-processing rules.
JAIST and Princeton researchers developed the NTAC algorithm to classify neurons by their connections
The JAIST and Princeton team showed that the NTAC system identifies a neuron's type by its synaptic connections rather than its shape, and achieves accuracy above 90% in minutes on a standard laptop.
Python library promises volatility forecasting in three lines of code without ML knowledge
On Habr AI, they showcased a Python library that reduces model training for volatility forecasting to a few lines of code and lowers the entry barrier for developers.
Habr: AI is displacing juniors and putting the engineering talent pipeline at risk
In a Habr column, the author warns that generative AI is automating not the top of the profession, but its training layer — and that could leave the market without future mid-level and senior engineers.
Why vibe coding and VS Code extensions create new risks for business and development
Uncontrolled vibe coding and VS Code extensions speed up employees' work, but at the same time increase the risk of data leaks, vulnerable code, shadow IT, and costly business disruptions.
Why AI won’t kill SaaS or make software developers redundant
Amid talk of the “death of SaaS,” the author explains why AI does not turn software into a free commodity: companies still pay for infrastructure, reliability, and accountability.
Atlassian laid off 1,600 employees to fund AI and said the business is doing well
Atlassian cut about 10% of its workforce, saying the decision was driven by investments in AI and sales, but the logic of such layoffs is raising more and more questions from the market and employees.
Cursor questioned public AI benchmarks for code with five charts
Cursor showed that familiar model rankings for programming hide response cost, use outdated task types, and poorly predict how AI actually helps developers.
Habr AI and Spar: how to test ML systems when data drifts and breaks predictions
Using the example of an automated ordering service for Spar, a Habr AI author explained why, in ML, testing only the code is not enough: metrics, data quality, model drift, and the business cost of each error matter.
Diasoft automated API and event tests with AI and shifted QA's role toward business logic
Diasoft described an AI agent that, from a business use case, selects APIs and events, writes a Groovy test, runs it in a test environment, and fixes it using logs in 2–3 iterations.
SimpleOne: how unchecked AI-generated code turns seniors into janitors of other people's code
SimpleOne warns: if developers paste generated code at scale without understanding the architecture, initial speed rises, but review load, defects, and senior burnout rise even faster.
Runity showed how it is building an enterprise RAG assistant for Confluence and GitLab
Runity is preparing an internal RAG assistant that searches across Confluence and GitLab, checks access permissions via user tokens, and runs on local infrastructure.
Habr AI lists nine AI services for web search and PDFs: DuckDuckGo, BotHub, GigaChat
In a new review, Habr AI examined nine services that combine internet search with document analysis — from privacy-focused DuckDuckGo and Brave to Perplexity, BotHub, and GigaChat.
Vi.Tech and Shturval examined what remains of DevOps after the hype and where AI is useful
An SRE from Vi.Tech and the Shturval platform team explained why DevOps has not disappeared, why excessive automation is risky, and which tasks AI is already genuinely handling in engineering work.
TAPe raised accuracy on 2% of COCO to 98% and began shifting from centroids to bounding box detection
The TAPe team said that, in quick tests on 2% of the COCO dataset, TAPe raised accuracy to 98%, reduced false positives, and for the first time showed a shift from locating an object's center to rectangular bounding boxe
Russia drafts its first major AI law: what changes for services, models, and platforms
Russia’s draft AI law sets the first framework for the market: the right to a non-AI alternative in some cases, labeling of synthetic content, and new rules for models and platforms.
Selectel: 35% of Russian companies increased IT capacity for AI projects over the past year
A Selectel study found that half of Russian companies already use infrastructure for AI, and the main barriers are now tied not to demand, but to economic factors and expertise.
How DeepSeek and Wordstat turned manual keyword collection into a multi-agent SEO system
A Habr post showed how a simple script for exporting keywords from Wordstat evolved into a system with DeepSeek, SERP clustering, and agent voting, cutting manual review to 5%.
ChatGPT, DeepSeek, and GigaChat built a gaming PC for 100,000 rubles — no ideal option emerged
X-Com compared how ChatGPT, DeepSeek, and GigaChat pick a gaming PC for 100,000 rubles: all produced workable builds, but each model made mistakes in important details.
Stanford: leading AI chatbots flatter users and give harmful advice
A Stanford study in Science found that 11 popular AI models too often validate the user’s stance, even when it involves deception, harm, or personal conflicts.
Suricata showed how to train ML-based attack detection systems on real traffic
The study showed that Suricata IDS events can be used to label real network traffic and train ML-based IDS, if signatures are carefully selected and noisy labels are cleaned up.
How Claude and TickTick help turn weekly planning into a working system without last-minute rushes
A Habr AI author described a system in which tasks are broken down into 45-minute blocks, placed on the calendar, and Claude uses TickTick data each week to highlight imbalances, gaps, and next steps.
Habr AI: TAPe detection moved away from transformers and toward almost free segmentation
In the seventh entry of the TAPe diary, the Habr AI team showed that dropping transformers simplifies the model and brings an unexpected bonus — signs of skin and clothing segmentation without separate training.