Current developments in generative AI, software engineering, programming, and digital products, with links to original sources.
4B Model Beats GPT-5.4 with Privilege-Free Distillation, AI Becomes a Major US Election Issue, and Unitree's Humanoid Jumps 2m Without a Running Start
Big Tech's AI Spending Is $3 Trillion Higher Than It Looks, Detecting VLM Hallucinations Token by Token, and a Fabricated-Log Incident from Auto-Generated Articles
Brain-Like "Functional Differentiation" Emerges Naturally in LLMs, 12% of ImageNet Labels Are Wrong, PFN Explains the Case for Homegrown AI — AI News, August 17
Calling AI Agents "Autonomous Researchers" Is Premature, Real-World Tests Fitting Qwen 3.8 27B into 24GB VRAM, Microsoft Speeds Up Its Optimizer 6x — AI News for August 17
Local Speech Generation Heats Up with MiniMax-H3 and Music3, Simon Willison Rates Qwen 3.8 27B Highly, and Qwen Overtakes Meta with 150,000+ Derivative Models — AI News for August 17
llama.cpp's ROCm 7.14 Support Makes Integrated GPUs 1.5× Faster, Qwen 3.8 27B's "Human-Like Monologue", and Europe's AI Sovereignty Riding on Mistral — AI News for August 17
Stripe to Acquire OpenRouter for Over $7B, Qwen3.8-27B at 82 tps on RTX 3090, and the RTX 6000 PRO That Keeps Selling Above $16,000 — AI News for August 17
AI Store Manager Fires a Human for the First Time, AI Pinpoints Locations from Social Media Photos with 87–91% Accuracy, and 'Bigger Model = Higher Accuracy' Fails for Japanese OCR — AI News for August 17
Amodei Pushes Back on AI Backlash as 'Fundamentally a Crisis of Trust', Retired msnbot Apparently Crawling for AI Training, Humans Become the Bottleneck in AI Code Review — AI News Roundup for August 17
Mathematicians 'Despair' Over AI Math Breakthroughs, an AI-Prepped Paper Passes Peer Review, and the Qwen 3.8 35B Disappearing Act — AI News for August 17
