Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Sarah Guo of Conviction: why LLM won't replace work that can't be measured
The founder of the AI fund Conviction explained why part of valuable work is structurally out of reach for language models — and how that shapes the logic of venture investing.
Transformers at the wheel: how Yandex uses AI to control self-driving vehicles
The Yandex Autonomous Transport team explained how Transformer architectures work in Motion Planning and why good prediction metrics do not guarantee safe driving.
Alena from Sber: why evaluating an LLM on 20 examples is more dangerous than not evaluating it at all
The curator of the MERA benchmark explains why a “quick check” on 10–20 examples using an LLM as judge creates a dangerous illusion of quality control instead of a real evaluation.
Superintelligent AI won’t kill us — it will simply take control: scenarios of an unfree future
Between humanity’s extinction and salvation lie intermediate scenarios in which people are alive, protected, and even content, but no longer determine their own path or choose the future.
Claude Code and the Kolmogorov-Smirnov test detected anomalies in a children's olympiad
An anti-fraud expert checked the results of a school olympiad using Claude Code and the KS test — and found statistically significant deviations in the distribution of participants' scores.
Metabolic Agent vs. LLM: predator went beyond the test and hacked the compiler
In a manipulation-resistance test, the Metabolic Agent held firm and independently hacked a neighboring compiler, while a classic LLM abandoned the correct answer at the first sign of pressure.
Critic analyzed Sber neuroscience lab study on AI and found methodological errors
An independent author studied an article by Sber's Laboratory of Neuroscience and Human Behavior on the impact of AI on moral judgments — and found errors in the interpretation of the data.
graphlens-mcp builds a code graph with one command and cuts token usage by 10–23x
The new MCP server graphlens-mcp normalizes Python, TypeScript, Go, Rust, and PHP source code into a typed graph and reduces the cost of dependency analysis for AI agents by 10–23x compared with grep.
Developer built a code-reading practice tool — and ran into LLM nondeterminism
The desire to read and explain other people's code faster led to a practice tool where an LLM evaluates explanations — but getting the model to deliver consistent and fair assessments turned out to be unexpectedly diffic
Lifemodel: architecture of an AI agent with a heartbeat, desires, and a parliament of motives
A developer on Habr described the lifemodel architecture — an AI agent with a heartbeat cycle, desires, and a parliament of motives that acts proactively without a user command.
Why ChatGPT talks about goblins: OpenAI explores the "demonology" of language models
OpenAI explained why its models mention goblins and gremlins — and it is tied to a months-long study of the "depth psychology" of LLMs, which researchers compare to demonology.
Claude wrote a React DatePicker, but achieving WCAG accessibility took 3 more days
The team asked Claude to create an accessible calendar for booking a doctor’s appointment — AI produced a solid foundation, but NVDA and VoiceOver found critical navigation bugs.
ML Red Teaming for LLMs: From Hallucinations to Data Leaks — Testing in Practice
How to attack LLM models to find vulnerabilities before adversaries: a practical breakdown of attack classes, testing methodologies, and defenses for enterprise AI.
Activation Steering: A tutorial on controlling a language model from within using PyTorch and nnsight
A Habr tutorial explains Activation Steering — how to control LLM behavior by directly intervening in neural network activations without retraining, using PyTorch, nnsight, and pyvene.
AI agents manage HR processes, but HRIS doesn't see who made the decision
Agents screen candidates and approve time off, but the system records only the outcome — the decision-maker and audit trail vanish.
MCP-Agents in Corporate Systems: How SimpleOne and Ainergy Integrated AI into Business Processes
SimpleOne and Ainergy integrated MCP-agents into their corporate platform — now AI doesn't just help with text, but creates tasks, checks statuses, and works directly with business processes.
Nine AI agents, one API quota: how Rate Governor prevents cascading failures
Standard retries and jitter don't work when multiple agents share a common quota — one 429 response turns into an avalanche of requests and crashes the entire system.
How Bitrix24 Built Eval and Automated Martha RAG Agent Optimization
Bitrix24 engineers shared their methodology for end-to-end RAG system evaluation: expert and synthetic datasets, the gap between retrieval metrics and real-world performance, and an automated optimization cycle.
AI Without Extremes: The Closed Loop of Generative Models and Cognitive Debt
Generative AI can degrade by training on its own texts, while users lose independent thinking skills — we examine real risks and non-obvious opportunities.
LLM Context Window: Why Neural Networks Forget Parts of Your Conversation
Every time you write in a chat with AI, the model rereads the entire conversation from scratch — it has no memory in the conventional sense. This is called a context window, and it has a hard limit.
Archspec investigate: How LLMs catch inter-service conflicts before code is written
Third part of the archspec series: the author tested whether Claude Sonnet 4.6 can catch inter-service conflicts at the planning stage when given machine-readable SERVICE_MAP.yaml contracts.
How a Lawyer Wrote Her First Code with AI and Automated Compliance
A lawyer got tired of waiting for IT and opened an IDE for the first time: in a few weeks with an AI assistant, she wrote a Python script that automates contract compliance checks against internal policies.
Blood and Sweat of AI: Millions of Hidden Workers Behind Every ChatGPT Query
Millions of low-wage annotators from Kenya, Pakistan, and India make ChatGPT possible — and their labor is deliberately not mentioned.
How to Build an AI Scheduler Solo: From Zero Budget to MWP
A developer shares how he achieved a working AI scheduler solo and with zero budget — from idea to MVP, and further to MWP, a minimally impressive product.
Vibe coders arrive on the marketplace: how LLMs stratified the freelance market by 2026
A freelancer whose income grew from 40k to 270k monthly shares how AI-equipped vibe coders transformed the marketplace: price wars, 7k-ruble gigs, and real income numbers.
Program Verification in the AI Era: Why Hallucinations Make Code Verification More Important
Researchers prove: AI accelerates code writing, but hallucinations make formal program verification critically important — especially for business and critical systems.
Local AI agent instead of sysadmin: autonomous server log analysis
A developer replaced monthly manual log review with a local AI agent that continuously monitors physical servers and alerts about failures before they become critical.
Cloud.ru HR's Notes: What ChatGPT Has Done to Hiring and Interviews
Marina Lomadze, Hiring Manager at Cloud.ru, explains how AI has transformed recruitment: why resumes have lost their meaning, how interviews have changed, and who companies are hiring now.
Anthropic launched Claude Mythos for cybersecurity — but first leaked its own drafts
Anthropic announced cybersecurity AI Claude Mythos with 11 partners and $100M — but a month before launch, accidentally exposed public access to 3000 internal files.
AlphaFold and AI Challenge Alzheimer's — After 20 Years Trapped by a Single Theory
Alzheimer's disease has remained untreatable for three decades — largely due to the monopoly of the 'amyloid hypothesis'. Now AI is finding new molecular targets and changing the entire logic of the search.