Import AI 461: 'Alignment is not on track'; FrontierCode; and synthetic research interns

A new nonprofit research organization called Sequent has been formed by researchers from the UK AI Security Institute and alignment theory startup Timaeus to develop alignment techniques for superintelligent AI systems. Sequent aims to raise $100-150M initially and grow to 40-80 employees, pursuing diverse research directions including scalable oversight, learning theory, heuristic arguments, game theory, and personas to achieve principled confidence in alignment. Cognition, the company behind Devin, released FrontierCode, a coding benchmark with 150 tasks in three difficulty tiers (Diamond, Main, Extended) that measures code quality, mergeability, and adherence to codebase standards; Claude Opus 4.8 achieved only 13.4% on the Diamond tier, indicating substantial headroom. Xiaomi introduced MiMo-V2.5-Pro-UltraSpeed, a 1-trillion parameter LLM that achieves 1000 tokens per second inference on an 8-GPU commodity node using FP4 quantization, speculative decoding via DFlash, and TileRT software from startup Tile AI. Researchers from Xi'an Jiaotong University and Xidian University developed the AARR (Act As a Real Researcher) benchmarks, starting with AARRI-Bench, to evaluate AI systems on entry-level research tasks; the best performing system, Claude-Opus-4.7 with Mini-Swe-Age, scored 42.13% on the Science scenario.