Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Developer Recreates All Seven of Adobe's Top Apps Using AI
In a year, Thomas Brandon built ArtCraft — a suite of seven free alternatives to Photoshop, Illustrator, Premiere, Lightroom, Acrobat, After Effects, and InDesign. The project is written in Rust with Claude Opus 5.5 and
How Quadratic Funding Could Govern AI Agents: An Indian Research Model
Praful Chandra, Sujit Gujar, and Ganesh Ghalme from India have proposed governing AI agents through quadratic funding: two voting markets (for and against), a hardware-level limiter, and a formula that rewards the number
A Rules-Based Pilot Builds Its Own Labels and Cuts the Wait for a Model from a Year to Zero
An ML lead describes a technique for procurement anomaly detection: skip the six-to-twelve-month labeling phase and launch a rules-based pilot that labels data as it works.
Nano Banana 2.1 and Hailuo 3.0 paired up: video from a single photo
Google's Nano Banana 2.1 generates the starting frame, MiniMax H3 brings it to life as video. How to pick a source image, describe motion physics, and avoid glitches — an analysis using five scene templates.
Qwen3.8-27B at IQ3 on an RTX 5070 Ti scores 91.1% and hits 68.3 tok/s
A Habr author ran Qwen3.8-27B at IQ3 on a single RTX 5070 Ti: 91.1% of tasks solved, an index of 82.3, and 68.3 tokens per second at 131K context.
"World of Quests" launches AI bot Questosha, tests five models across 12 gates
"World of Quests" rolled out Questosha starting Q3 2026, built on a markdown wiki trained on 7.5 million dialogues. A nightly self-learning cycle improves the knowledge base. Two models hit 12/12 on critical scenarios: Q
An AI secretary for $77 a month: 88% of answers from DeepSeek, 77% of the budget goes to Opus
Telegram bot Liza Petrovna costs $77 a month: 88% of replies run on cheap DeepSeek ($0.002-0.007), while the 7% of hard queries routed to Opus ($0.21-0.24) eat up 77% of the budget. Routing is regex-based, memory lives o
Five Engineering Decisions to Make Before Generating Code with AI
An experienced engineer sees questions before generating code that a vibe coder skips. Using a multi-brand platform and a marketplace as examples — five key decisions: integration boundaries, variation by brand, data com
An AI Estimator Without AI: Rule-Based Code Found 310 Issues in Five Cost Estimates
A developer built an estimate checker for a renovation company: code with rules flagged 310 issues across five estimates, and no neural network ran. One estimate is understated by 515,786 ₽.
There Are More Programmers Now, Even Though AI Promised to Replace Them Within a Year
Three years ago, the end of the profession was predicted within six months to a year. Today more people are paid to write code, and job openings haven't gone away; they just changed their requirements.
AI Solved 370 Math Problems, Mathematicians Panic, Musk Seeks $40 Billion: This Week's Highlights
OpenAI published solutions to 370+ math problems across 700+ manuscripts. Terence Tao launched a crusade against AI mathematics. Musk will borrow $40 billion for data centers, France risks default, Russia tightens the sc
Vibe-coding in Rust: three macOS utilities and five browser projects
Developer codecity builds software through LLM coding. This time: ProcessTree for managing processes, Notepad for big files, and DirMerge for comparing folders. Plus a recap of five earlier browser projects: an Obsidian
Qwen3.6-35B-A3B hits 109 tok/s on an RTX 5080 with an RTX 3060 on PCIe Gen2 x4
An enthusiast ran a 35B model on an RTX 5080 and an RTX 3060 in a PCIe Gen2 x4 slot (about 2 GB/s) and got 109 tokens/s. Here's why the narrow link wasn't a bottleneck.
How to Save AI Agent Tokens When Working with Large Codebases: 5 Patterns
When analyzing large projects, AI agents spend most of their tokens exploring code rather than writing it. Habr breaks down 5 universal context management patterns to cut compute costs.
Refactoring C++ in OpenBMC: How an AI Agent Fought a 280-Line Function
A developer asked an AI to help with OpenBMC's createSensors function, which crams 12 tasks into 280 lines. The agent proposed 7 concrete improvements and split the monolith into 10 methods. The result is paradoxical: th
Habr: AI Made Forecasts Cheap, but Checking Them Still Takes Hours of Human Work
A fake vulnerability report takes minutes and cents to write and hours to check. A Habr author carries that asymmetry from code over to AI forecasts, which have no verification anchor.
Why Repeated Fine-Tuning Hits a Ceiling: The LANN-4 Study
The LANN-4 author ran experiments with repeated fine-tuning of a small model. The first round produced a jump from 53% to 59% correct answers on 200 tasks — the next three rounds added nothing. The open question: did the
Agentic search on open-source models: who it found, what it cost, and where it all broke
A Habr author spent $20 on an experimental agentic search system: he tested six open-source models, found that gemma4 with a simple JEV filter works best, and discovered why searching for HR contacts through open sources
Developer Spends $9,000 on AI Agents and Ships Crossweft, an Open-Source Plugin
The author spent over $9,000 and a billion tokens, and ended up with an open plugin for Claude Code and Codex that stops an agent from saying "done" when files are out of sync.
Odyssey-3: how a video model teaches robots physics in tens of hours
Odyssey has unveiled Odyssey-3, a world model that teaches robots physics by watching video. It scores 66.1% on the Physics-IQ benchmark, trains on video, games, and simulations, and adapts to specific systems in just te
The Poetry of Agentic Development: How to Code with Agents When Ease Breeds Complexity
A Yandex developer breaks down why agents speed up coding but create complex systems. Research shows LLMs generate 2.3x more lines of code, fail to avoid complexity, and build hidden dependencies. A full analysis of the
Open-Source Agent vs. 1C Napar: Hybrid RAG on DeepSeek Wins in Practice
Habr tested a self-built LangGraph agent with hybrid RAG search against the commercial 1C Napar assistant. On three real developer tasks using deepseek-v4-flash, gpt-6.1-sol, and glm-5.3-flash, the open-source agent went
Where an NLI Model Helps in a Meeting Assistant: Three Fittings, Two Rejections
Developer Charoite tested an NLI model in five scenarios. Only phrase deduplication in the live feed worked (2% error rate). It failed on meeting minutes, core revision, and question filtering: 55.6% accuracy and massive
8 AI Tools for Reels and Shorts: What Actually Works in 2026
We tested 8 AI tools for cutting vertical video. Editing time dropped from hours to minutes. Here's which ones support Russian and how to pick the right one for your task.
Hebbian Learning Without Backprop Collapsed Into k-means and Lost on Every Front
A Habr author trained a hidden layer using only local Hebbian rules. The rule collapsed into k-means and beat backprop on accuracy, on few-label performance and on forgetting in none of the tests.
I Wrote 61 Tests for My AI Mentor. Then I Found Out You Can't Trust the Tests Either
A developer built 61 tests to check an AI mentor on a coding education platform. But the test bench itself had bugs: it marked cut-off answers as passing and let finished solutions slip through. DeepSeek V4 Flash scored
When Small LLMs Beat the Swarm: How Splitting Data Into Sections Gave a Weak Model a 4x Boost
A heating-systems engineer ran an experiment with six weak models (qwen2.5-coder-1.5b, llama-3.2-1b, internvl3-2b, and others), borrowing the structure of a design bureau. Splitting the data into blocks of 6 records prod
Servers 2027: Rubin Ultra, MI500, and Open Interconnects Break the Monopolies
2027 promises to crack NVIDIA's grip on GPU interconnects: AMD will launch the MI500 with a thousandfold performance jump, Intel returns to the fray with Diamond Rapids, and a consortium is rolling out the open UALink st
Long Nardy Through AI's Eyes: 170,000 Moves and Human Mistakes
An analysis of 170,000 moves from 6,383 long nardy games shows how a neural engine picks the best move and where humans go wrong. The rules of long nardy turned out to be far trickier for an algorithm than they first app
Three Neural Networks Drew Web References in Five Styles: From Fintech to Web 1.0
Three neural networks (Grok Image 2, Flux 2 Pro, Nano Banana Pro) created web assets for esports, fintech, Web 1.0, ecology and AI in a series of 14 examples. The result: Grok and Nano delivered near-perfect output, whil