Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
243,000 texts for a 29M-parameter model: how the author picked data from Wikipedia and psychology books
A Habr series author took 200,000 paragraphs of Russian Wikipedia and 50,000 from psychology books. Of 1.36M texts, 243,412 went into training, because it is cheaper to compute and train a 29M-parameter model that way.
Noise in Embeddings Doesn't Protect Data: 100 Known Pairs Reveal ~90% of Documents
Tested 8 embedding protection schemes: Gaussian noise leaves 99.9% of records identifiable as retrieval quality halves; 100 known pairs (2% of the corpus) expose about 90%.
Bank Tech Lead Grows OpenCode Telegram Bot to 1,298 GitHub Stars With AI Agents
A bank tech lead used AI agents to grow an open-source OpenCode bot to 1298 stars and 48 contributors in a year. Here is what worked in development and promotion.
Claude Code Edits Video From a Transcript: How the video-use Skill With 30 000 Stars Works
The open-source video-use skill teaches Claude Code to cut pauses, keep the best takes and assemble a video via FFmpeg. The model barely watches the footage: it reads timestamped text.
Developer builds a RAG system on local models, boosting search accuracy from 20% to 53%
An engineer built a news-analysis AI agent without a single cloud API call, running everything on local models on his own hardware. A hybrid search approach fixed the "homonym problem" — like confusing "gold" the metal w
OES Puts 30 AI Agents on Mining Haul Trucks and Lifts Haulage by 10%
An AI layer on top of a legacy dispatch system lifted haulage 10% at a 240-vehicle open-pit mine with no new trucks, and drivers carry out 95% of directives without calling dispatch.
Qwen 3.5 4B gets a RISC-V processor inside its weights, with DeltaNet as memory
Divyansh Agrawal compiled a 32-bit RISC-V core into Qwen 3.5 4B's weights without training. Long-number arithmetic accuracy rose from 48.6% to 100%, matching Qwen with Python.
Claude in Houdini on Windows: Two Bugs in the Agent Panel and a 31-Node Geode
The author connected Claude to Houdini via an MCP panel, found two Windows bugs, submitted fixes, and built a procedural geode from 31 nodes in 22 minutes.
AI agent on GPT-5.5 audits 2600 TravelLine autotests in two weeks
A TravelLine manual tester used an OpenCode agent to audit E2E tests before a .NET migration. In the Transfers section, 15 of 18 UI tests turned out to duplicate integration checks.
Google DeepMind finds softmax flaw: gradient drops to zero at a logit gap of 16.6
Barron, Gnanapragasam and Tschannen show that in float32 the cross-entropy gradient vanishes at a gap of about 16.6, and distillation gets spurious noise. Rearranging a fraction fixes it.
Keystone removes the model name from code: open-source LLM gateway adds 0.64 ms
Keystone's open-source Go gateway routes requests by difficulty, not model name. Fallbacks, dollar limits and a 250 ms judge add just 0.64 ms of latency.
YADRO bans its LLM from touching numbers and ships 10+ reports a day with no humans in the loop
In three months, YADRO analysts built a service on Dify, Trino and Airflow. Code computes the metrics, the model writes commentary with placeholders, and over a dozen reports go out each morning.
Security engineer builds an MCP server in 7 days that spooks an AI agent with vulnerabilities
AppSecGate MCP runs scanners and keeps a findings database. On OWASP Juice Shop, an agent worked through 200 findings in 50 minutes for 15 rubles and fixed 56, including 8 critical.
aiedit hides individual lines of code from an AI agent: 5 levels and 857 tests
A Python beginner built the aiedit editor with Claude: the model sees the project but not the locked lines. Code, not a prompt, enforces it: 11K lines, 857 tests, MIT.
vps-forge: root-enabled agents on a VPS get automatic rollback, self-bans drop from 7 of 60 to 0
The open-source Agent Skill vps-forge arms a 5-minute dead man's switch before the agent touches anything. Across 120 Opus 5.5 runs, scores rose from 79–92 to 99–100, at roughly triple the cost per run.
BigSleep LLM Agent Finds 45 V8 Vulnerabilities, Fuzzilli 37, With 25% Overlap
Project Zero compared the BigSleep LLM agent and the Fuzzilli fuzzer on V8 over six months: 45 vulnerabilities vs 37. Overlap: a quarter in the tracker, about 60% in the experiment.
rwkv-quant: keeping one matrix in bf16 removes 39% of RWKV-7 2.9B's quantization loss
The rwkv-quant author found why COMPRESSION on RWKV-7 2.9B loses +4.04% perplexity versus +2.85% on 1.5B: one matrix is to blame, and the fix costs 0.5% of the file.
Qwen3-8B on the Intel Core Ultra 5 226V NPU: 10.7 tok/s at 15 W and 24 dB of Noise
Qwen3-8B on all three Lunar Lake blocks in a Ninkear L20 mini PC: the NPU hits 10.7 tok/s at 1.52 J/token, the GPU is twice as fast, and the machine stays quiet with CPU and GPU free.
Gemma 26B fine-tuned on a 24 GB MacBook: offline assistant solves 703 of 766 conversations
A developer fine-tuned Gemma 4 26B A4B with QLoRA on a laptop, moving a 5,000-token rulebook into the weights. Result: 703 of 766 conversations correct, median 1.64 s per reply.
Codex on GPT-6.1 Sol trims its window to 828 400 but still accepted an 888-thousand-token input
The config says model_context_window = 1 000 000, yet Codex reports 828 400. A 890 547-token test shows it is a history threshold, not a limit on a single input.
12 Pentesting Tools vs. Juice Shop: The Best AI Agent Found 5 Holes Out of 116
The author ran 6 AI agents, 4 DAST scanners and a Claude + Burp MCP combo against OWASP Juice Shop's 116 vulnerabilities. The hybrid won with 9/10; headline benchmark claims didn't hold up.
Replacing LLM weights with a recipe: how 16 GB of matrices are being folded into formulas
Instead of storing billions of numbers, a model can be described by a weight generator: hypernetworks, formulas, fractals. Habr examines what works and at what cost.
Habr Breaks Down Oracle Bias: AI Speeds Up Code, but Trust in It Isn't Growing
Part two of a series on reliable programming explains why process metrics rise after LLM adoption while trust in the code does not. It draws on data from DORA, Veracode and FormAI.
Geoffrey Huntley Sorts Dev Tools into S-Tier and F-Tier: What Decides It Is --help and Model Weights
Geoffrey Huntley: in the MCP vs. CLI debate, the winner is whoever made it into model weights. GitHub's gh CLI works without skills, and newcomers should start with --help.
Homemade AI-slop detector catches 5 of 12 signs of machine-written text with code
An author mapped 12 signs of AI slop to the checks in his own detector: code catches five, two of them only partly. The other seven are left to an editor, and online detectors confuse humans with machines.
Claude Pushes an Amplitude Calculation from Eight Loops to Nine, Anthropic Says
Anthropic reports a nine-loop calculation by Claude; the previously published result for this amplitude stopped at eight loops. We look at how to verify scientific code written by AI.
Synthesized Voice Read a Novel in 1 h 58 min; the Surrounding Pipeline Code Was What Broke
The author of a Habr series voiced a 15 444-word novel: the voice held for two hours, while failures hit text splitting, phrase stitching and markup. A 14-phrase test set caught none.
Author on $200 Claude Plan Goes Back to 200K Context to Stay Within Limits
A $100 plan was enough in spring; by August the weekly limit ran out in 3 days. The developer blamed "everyday convenience" and went back to a 200K-token window.
Lambert: A Ban on Open Weights Would Pull Frontier Model APIs Down With It
Nathan Lambert of Interconnects examines the debate over cyber risks of open models: most documented attacks ran through closed systems, not open weights.
AI Reaches PLCs: Engineering Assistants Now Read Projects and Write Code
AI tools for PLC development are evolving from chatbots to full-fledged agents. PLC Copilot, Copia AI, and Siemens Eigen Engineering Agent now read projects, generate ST code, compile it, and fix errors.