Current developments in generative AI, software engineering, programming, and digital products, with links to original sources.
Open-Source Search Agent 'Iris' Hits 88.6 on BrowseComp — Intelligence Index v4.2 Doubles Down on Private Data, Plus LayerStoRm Runs a 186GiB MoE on 96GB VRAM
OpenAI Discloses CoT Monitoring Limits and a Training Halt Behind Its Research Acceleration — CNBC Reports "Model Fatigue," and VDN-H3 Makes Video Generation Faster Than Playback
OpenAI's Pachocki Likens AI to an "Alien Mind" and Calls for Preparedness — Anthropic Cuts Bio False Positives by 85%, and a Used V100 Runs Twice as Fast at 1/8 the Price of a DGX Spark
Eight Uncensored Qwen3.8-27B Variants Verified Over 167 GPU-Hours — LLM APIs Show a 33x Cost Gap by Call Shape, and NYC Moves to Pause AI in Schools Too
Google's WeatherNext 3 Predicts Weather Directly Without Physics Simulation—Deliberation Begins on an 'AI Psychosis' Diagnosis, Plus Music Generation in the Gemini App
Meta Releases 80-Millisecond-Response Voice Transcription Model — Commercial Service for Safety-Stripped Models Emerges, GPT-6 Jailbroken in 24 Hours
NVIDIA Releases an NVFP4-Quantized Version of Qwen3.8-Flash-Next — 2.7x Compression With Accuracy Largely Intact, Plus Reports of an "Antislop" Effect from LoRA on the Last 5 Layers
55.5% of Claude Code Tool Calls Come from Subagents — /compact Cost Revealed by 705 Real-Session Measurements, Plus an Indie Developer Whose AI Usage Equates to ¥18.76M a Month
Local LLM Qwen3.8-27B Reported to Analyze and Remove Malware — A Firsthand Account of a Rescue After Windows Defender Detected Nothing, Plus TP Caveats for Mixed Ampere×Blackwell GPUs
The Seattle Times and Others Sue OpenAI — America's Two Largest School Districts Move to Halt AI Use, Plus an NInfer Fork Delivering 550K-Token Context on an RTX 5090
