Top story
News publishers limit Internet Archive access due to AI scraping concerns. Publishers are blocking Internet Archive access as they attempt to control how their content is used for AI training. Source
Research
KaniTTS2 open-source 400M TTS model. New voice cloning model runs in just 3GB VRAM with pretrain code included. Source
YOLOX trained from scratch for aircraft detection. Developer avoided Ultralytics' AGPL license by training their own model for iOS aircraft detection. Source
Ground-up MLX reimplementation of Qwen3-ASR. Apple Silicon optimized speech recognition model built from scratch using MLX. Source
Two tricks for fast LLM inference optimization. Technical guide covering performance improvements for local language model inference. Source
Tools
Zvec lightweight vector database. Fast, in-process vector database designed for minimal overhead applications. Source
MDST Engine runs GGUF models in browser. WebGPU/WASM implementation enables running GGUF models directly in web browsers. Source
Oat ultra-lightweight HTML UI library. Zero-dependency, semantic component library for building web interfaces. Source
Industry
Smart sleep mask broadcasts brainwaves to open MQTT broker. Security researcher discovers privacy vulnerability in consumer sleep tracking device. Source
OpenAI should build Slack alternative. Analysis suggests OpenAI could leverage AI capabilities to create superior workplace communication tools. Source
Community
6-GPU local LLM workstation scaling advice. Community member seeks orchestration guidance for 200GB+ VRAM multi-GPU setup. Source
Qwen3-Coder-Next GGUF performance comparison. Community discusses 60B parameter model claiming near-identical performance to full Qwen Coder. Source
Popular MoEs speed comparison on Apple Silicon. Performance benchmarks for various Mixture of Experts models running on llama.cpp. Source