Daily AI Brief - Sunday, March 15, 2026
Generated: 2026-03-15 Items: 40 new stories
🤖 Daily AI Brief — March 15, 2026
TOP STORY: 1M Context Now Generally Available for Claude Opus 4.6 & Sonnet 4.6 Anthropic has made one-million-token context windows generally available across its flagship Opus and Sonnet model lines, marking a major milestone for long-document and agentic use cases. Anthropic Blog · Simon Willison's Take
Industry
arXiv Separating from Cornell University - The preprint server is becoming an independent nonprofit with Simons Foundation support, and is hiring a CEO at ~$300K/year. Reddit/MachineLearning
ByteDance Outsmarts US Sanctions With Offshore Nvidia Buildout - China's ByteDance is circumventing US export controls by acquiring Nvidia hardware through offshore entities to power its AI infrastructure. Benzinga
Anthropic vs. the Pentagon - A Guardian deep-dive examines how big tech's relationship with AI and military contracting has dramatically reversed course. The Guardian
Google Reimagines Maps with Gemini - Google is integrating Gemini AI into Maps to enable conversational search and immersive navigation experiences. Google Blog
Research
WaDi: Weight Direction-aware Distillation for One-step Image Synthesis - A new distillation method improves one-step diffusion image generation by leveraging weight direction guidance. Hugging Face
CREATE: Testing LLMs for Associative Creativity - Researchers introduce a benchmark designed to evaluate creative and associative reasoning capabilities in large language models. Hugging Face
Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights - New findings show that fine-tuned expert models cluster densely in weight space near their pretrained initialization. Hugging Face
Training Language Models via Neural Cellular Automata - This paper proposes using neural cellular automata as an alternative computational substrate for training language models. Hugging Face
Tools
Qwen3.5 35B Punches Well Above Its Weight - Community benchmarks show Qwen3.5 35B delivering frontier-competitive performance as a local model. Reddit/LocalLLaMA
StepFun Releases SFT Dataset for Step 3.5 Flash - StepFun has open-sourced the supervised fine-tuning dataset used to train its Step 3.5 Flash model on Hugging Face. Hugging Face
NVIDIA NeMo Retriever Agentic Retrieval Pipeline - NVIDIA introduces a generalizable retrieval pipeline for agentic workflows that goes beyond standard semantic similarity matching. Hugging Face Blog
Tree Search Distillation for LMs Using PPO - A new technique applies tree search distillation combined with PPO to improve reasoning in language models. ayushtambde.com
Community
LocalLLaMA Discord Server Launched - The LocalLLaMA community has launched an official Discord server and bot for real-time discussion and model sharing. Reddit/LocalLLaMA
Simon Willison on Agentic Engineering - Willison shares highlights from his fireside chat on agentic engineering at the Pragmatic Summit. Simon Willison
A Look Inside Dialector, Chris Marker's 1988 Chatbot - A fascinating historical retrospective on filmmaker Chris Marker's early conversational AI experiment predating the modern era. kubicki.org