Top story
I'm joining OpenAI. A prominent developer announces their move to OpenAI, generating significant community discussion about talent acquisition in the AI space. Source
Tools
MicroGPT. A browser-based visualization tool that lets you see how GPT models work under the hood. Source
Claude Pen Plotter Integration. An experiment giving Claude direct access to control a pen plotter for physical drawing output. Source
Gwtar Static HTML Format. A new efficient single-file HTML format designed for static content distribution. Source
Sudoku via Energy Geometric Models. A novel approach to solving Sudoku puzzles using geometric energy-based reasoning. Source
Research
Semantic Chunking and Natural Language Entropy. New research exploring how entropy principles apply to breaking down natural language into meaningful chunks. Source
CoPE-VideoLM: Efficient Video Language Models. A new approach using codec primitives to create more efficient video understanding language models. Source
Quantization-Robust LLM Unlearning. Research on using low-rank adaptation to make model unlearning more robust to quantization effects. Source
LLM Agent Cost Analysis. Analysis revealing the quadratic cost scaling challenges in deploying LLM agents at scale. Source
Industry
Qwen 3.5 Plus Release. Alibaba's massive 397B parameter model with 17B active parameters is now available on the Chinese Qwen app. Source
Ling-2.5-1T Model. New 1 trillion parameter model from inclusionAI now available on Hugging Face. Source
Ars Technica AI Fabrication Retraction. Major tech publication retracts article after discovering it contained AI-fabricated quotations. Source
Community
StepFun AI AMA. The team behind the Step-3.5-Flash model will host an AMA session Thursday 8AM-11AM PST. Source