Sunday, November 15, 2026
AI breakthroughs · Research · Analysis
AI
Keita Broadwater’s free, unfinished book treats AI reliability as a measured probability. We apply its methods to Visa’s new shopping-agent benchmark.
Read full AI deep dive →AI
A forensic investigation of publicly released AI model weights has identified numerical fingerprints linking six models from DeepSeek, Moonshot AI, MiniMax and Ant Group to NVIDIA's restricted Blackwell processors. The findings could expose a new way to investigate China's access to advanced computing hardware, although they do not establish where the chips were located or whether export-control laws were broken.
AI
Nathan Lambert’s RLHF book is the clearest guide to how chatbots get their personality. A new sycophancy study tests its map, and shows why it still matters.
AI
Headline AI scores vs. harder real-world tests: Co-Design Bench, ScarfBench, SWE-bench, ARC-AGI-3 and METR, with harness effects and sources.
AI
OpenClaw has announced native iOS and Android apps that connect your phone to a self-hosted AI agent gateway — keeping email, calendar, and automation on your own hardware.
