A curated AI research newsletter covering several topics: Ajeya Cotra updates her AI capability timelines, now expecting agents to reach 100-hour task horizons by end of 2026. GovAI and Oxford researchers propose 14 metrics for measuring AI R&D automation as a precursor to recursive self-improvement. Indian researchers prototype a citywide edge-computing traffic surveillance system using YOLO, SAM3, and NVIDIA Jetson chips. German researchers build TinyIceNet, a tiny vision model for on-device satellite sea ice monitoring on FPGAs. ByteDance fine-tunes its Seed1.6 model into CUDA Agent, achieving state-of-the-art GPU kernel generation performance on KernelBench. The issue closes with a short fiction piece set in a drone-war future.
Sort: