AI Import 461: 'Alignment Not on Right Track'; FrontierCode; and Synthetic Science Interns

Former employees from the alignment team at the UK AI Security Institute and the startup Timaeus have joined forces to create a nonprofit organization called Sequent, which will focus on developing alignment techniques that provide high confidence in the safety of superintelligent AI. Sequent plans to raise $100–150 million in seed funding and hire 40–80 employees, betting on a portfolio of research directions: scalable oversight, learning theory, heuristic arguments, game theory, and personas. Cognition, the creator of Devin, has introduced FrontierCode, a benchmark consisting of 150 coding tasks divided into three difficulty tiers. The best result at the Diamond tier was achieved by Claude Opus 4.8, scoring 13.4%, underscoring the benchmark's difficulty. On the Chinese market, Xiaomi announced the MiMo-V2.5-Pro-UltraSpeed model with 1 trillion parameters and a generation speed of 1,000 tokens per second, achieved through joint optimization of architecture and software, including FP4 quantization and DFlash speculative decoding. Researchers from several universities have created ChinaHeritaQA, a multimodal benchmark comprising 2,279 images of 51 UNESCO World Heritage sites in China and 14,133 questions. The best open model, Qwen-VL-8B-Instruct, achieved 81% accuracy, surpassing the human baseline of 67%. Additionally, researchers from Xi'an Jiaotong University and Xidian University developed the AARR benchmarks, which assess the ability of AI agents to perform tasks typical of a research intern. The top-performing agent was Claude-Opus-4.7 paired with Mini-Swe-Age.