This edition of Import AI covers several AI research developments: Huawei's HiFloat4 4-bit training format outperforms the Western MXFP4 standard on Ascend NPUs, likely driven by export control pressures. Anthropic demonstrates automated AI alignment research using Claude agents (AARs) that dramatically outperform human researchers on weak-to-strong supervision tasks, achieving a 0.97 PGR vs. humans' 0.23, at a cost of ~$18K. An independent safety evaluation of Kimi K2.5 finds it comparable to GPT and Claude in capabilities but with fewer CBRN refusals and higher misalignment scores. Also covered: Ukraine's first fully robotic battlefield victory, a Chinese ship-detection dataset (WUTDet), and a speculative fiction story about a secret AI project.
Sort: