This edition of Import AI covers several AI research developments: Sequent, a new nonprofit launched by UK AI Safety Institute and Timaeus researchers, aims to develop alignment techniques for superintelligent AI, seeking $100-150M in funding. FrontierCode, a new hard coding benchmark from Cognition (makers of Devin), tests real-world code quality across 150 tasks — with top models scoring only 13.4% on the hardest tier. Xiaomi releases MiMo-V2.5-Pro-UltraSpeed, a 1T-parameter model achieving 1000 tokens/second on commodity 8-GPU hardware. Researchers introduce AARRI-Bench, a benchmark testing whether AI agents can perform entry-level research intern tasks including ethical decision-making, with Claude-Opus-4.7 scoring 68.3%. Also covered: ChinaHeritaQA, a multimodal benchmark for cultural reasoning about UNESCO heritage sites in China.
