Habr AI
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
LLM Agent Built a Gearbox in KOMPAS-3D, but Scored Only 50 out of 100 on the Specification
The kompas-mcp plugin lets an LLM agent drive KOMPAS-3D: it built a gearbox and its design documentation. In a homemade benchmark, 3D tasks score 88–100; the specification scores 50.
Agents ignored the prompt to read memory in 7 of 7 sessions: an analysis of 13 thousand logs
The kaeru team read 13 thousand logs from its agents. They never followed the prompt to read memory in any of 7 sessions, but they reliably close "debts" and lose truncated instructions.
Bank Architect: In a METR Test, AI Made Experienced Developers' Tasks Take 19% Longer
A Russian bank's head of architecture argues that probabilistic AI components need boundaries, not blueprints: in a METR test, experienced developers got 19% slower.
Developer of Viy-Based AR Horror Game Builds 12 Monsters in Tripo AI and Cuts the Devil to 7 880 Triangles
The developer of Home Haunt generated 12 monsters from Viy in Tripo AI and cut the Devil from 38 444 to 7 880 triangles so the Unity AR horror game can run on an iPhone alongside ARKit.
Manager-less DreamTeam swarm wrote trustcheck for $5.68 and passed 13 of 13 hidden tests
Open-source framework DreamTeam dropped the manager agent: a swarm of 8 agents wrote trustcheck for $5.68 and passed a hidden benchmark 13 of 13, versus 1 of 13 for a stub.
Faros: median code review time up 441% in teams using AI agents
Faros telemetry on 22 000 developers: median review time rose 441% with AI. Why engineers work more even though nobody makes them, and how to avoid burnout.
TJ Digital: 23% of AI-cited pages open with a direct answer, versus 14% of uncited ones
Analysis of 193 507 ChatGPT and Google answers: 23% of cited pages open with a direct answer vs 14% of uncited ones. The gap is small, and it is not treated as causal.
Bun creator uses Claude Code to port 750K lines from Zig to Rust in 11 days
Jarred Sumner ported Bun from Zig to Rust in 11 days using hundreds of Claude Code subagents. 99.8% of tests passed, but this mode burns through a subscription quota about ten times faster.
Meta's Muse assistant keeps dossiers on loved ones and gets root on Ubuntu by default
Security researchers dug through Muse's files: the agent updates pages on a user's family and friends every hour and gets root access on Ubuntu by default.
mcp-linter finds 18 undescribed parameters in Anthropic's filesystem server
mcp-linter's author ran 34 rules against Anthropic's official servers: filesystem has 18 undescribed parameters and scores 38 out of 100; memory scores 66, everything 90.
Adaptive noise in an SDE sampler improved texture by only 1–2% on Flux 3 Klein
A Habr author tested whether a sampler can choose noise strength from curvature and solver error. On Flux 3 Klein the texture gain was 1–2%, and people did not notice it.
Local models lose to Opus and Codex: 0.21 on a MacBook versus 1.00 for Codex
A Habr author ran one trap-laden task on the cloud, a MacBook, an RTX 5080 laptop and a DGX Spark. Codex scored 1.00; the best MacBook model scored 0.21.
Habr author fine-tunes YOLOv8l on dental X-rays and pushes mAP50 to 0.642
YOLOv8l was trained on panoramic dental X-rays in 29.3 minutes on a single A100. It reaches 0.642 mAP50 on DENTEX validation and finds impacted teeth with an AP50 of 0.961.
16 models tested on 7 search tools: Exa averaged 10.5 out of 14, SearXNG about 3
A Habr author tested how 16 models search for Bitrix24 data through 7 tools. Exa averaged 10.5 out of 14; MiMo v2.6 Pro won without a single fabrication.
Kimi scores 96.2% on a 1C database analytics test and emerges as a fallback if Claude access is blocked
After his Claude account was blocked, a 1C developer tested Kimi on the Trade Management 11 demo database: 96.2% match with the reference on ten questions, 24 queries in 45 seconds.
Voice bots built on LLMs turn to Jev, a fast classifier, for decisions in 200–500 ms
Teams are moving simple decisions out of large language models (end of turn, topic change, booking confirmation) and into Jev, a fast classifier with 70–500 ms latency.
MoneyGate safety layer blocked all 42 fraudulent requests to local AI models
Llama 3.1 8B, Qwen 2.5 7B and DeepSeek-R1 8B on a gaming laptop rarely spotted the scam and wrote to fraudsters, "I'll transfer it today." MoneyGate stopped all 42 requests.
Hugging Face's BERT Tokenizer Gets a 50x Speedup by Skipping Needless Work on ASCII
The tokenizers team found that BERT spends 88% of its time normalizing text, and taught it to skip English characters in 16-byte batches. The result: a roughly 50x speedup.
AI cartoon series: a 24-second episode built ten ways, the best cost 162 credits
The makers of "Forest Street" spent three days testing Wan, Kling, Veo, Seedance and GPT Image. A second-by-second script and a cheap image-based draft won; animation was ordered only after review.
Open-source LLM gateway KTMS 1.0 plugs your own GPUs into the cloud and survives outages
Keystone has released KTMS 1.0, a Go gateway that routes requests to your own GPUs first, then to the cloud. It documents Redis and Postgres failures, key rotation and signed releases.
FAIRouter for C# Is Now on NuGet and Picks a Model From 54 Built-In Options
FAIRouter 0.1.0 for C# is on NuGet: the library picks a model for each task from 54 options, checks the answer with a judge scoring 0 to 1, and learns from feedback.
22B-Parameter LTX-2.5 Runs on a 10 GB RTX 3080, but Takes 5 Minutes per Shot
A Habr author squeezed the 22-billion-parameter LTX-2.5 video model onto a 10 GB RTX 3080. The price: 32–64 GB of RAM and about five minutes per shot, only 25 seconds of it denoising.
128 AI Agents With No Boss Passed 49% More Tests Than One
Microsoft Research ran up to 1024 copies of one model writing code with no manager. A 128-agent swarm passed 28.78% of tests vs 19.31% for a lone agent, and set up its own reviewers.
Strata Runs Qwen3.8-Flash-Next on a 12 GB RTX 5070: 94 Tokens per Second at Q2_0
Strata gives MoE models a CPU-style memory hierarchy: VRAM, RAM and SSD. Qwen3.8-Flash-Next hits 94 tokens per second at Q2_0 on a 12 GB RTX 5070.
Agent traffic on Zocdoc is up 30x since July, and ChatGPT faces two health lawsuits
AI agent traffic on Zocdoc has grown about 30x since July, while two lawsuits accuse ChatGPT of advising patients to drop treatment. What health service interfaces need.
baton utility moves a session between 5 CLI AI agents without losing context
The Go utility baton pulls a session from the logs of Claude Code, Codex, agy, Pi or Hermes and hands it to another agent when the five-hour quota runs out. Open source under MIT.
Hard Stop halts an AI agent in the Linux kernel with a median latency of 4.8 microseconds
Hard Stop places a supervisor between an AI agent and the Linux kernel: any action outside the allowlist is blocked, with a median trigger latency of 4.8 microseconds.
Claude Code built a FLUX.3 klein pipeline: 290 doctor images for $3.2 in GPU rental
A paid-traffic specialist for clinics had Claude Code build a pipeline: a website link in, a folder of vetted doctor images out. The result: 290 images for about $3.2 in GPU costs.
At RGS, local LLMs cut the review of 500 tasks from five working days to two
RGS's head of IT compliance built an n8n and Ollama pipeline on a laptop with no budget and no code. Accuracy on real task data is 78–82%.
HOSTKEY wrote a pipe for its bot: Gemma 4 answers in 5–7 seconds, Qwen 3 in 30–40
HOSTKEY replaced OpenWebUI custom-model settings with a pipe: Gemma 4 with thinking off answers the support bot in 5–7 seconds on two A4000s, while Qwen 3 thinks for 30–40.