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

Claims

• Sequent, formed by researchers from the UK AI Security Institute Alignment team and Timaeus, argues that ASI may arrive within years while current alignment approaches are unlikely to provide principled a priori confidence in safety before superintelligence is trained.
• Sequent plans to raise $100-150M initially and scale to 40-80 employees, pursuing a portfolio of differentiated research bets including scalable oversight, learning theory, heuristic arguments, game theory, and personas.
• Cognition's FrontierCode benchmark scores Claude Opus 4.8 at only 13.4% on its hardest Diamond tier, with GPT-5.5 at 6.3% and Claude Opus 4.7 at 5.2%, providing meaningful headroom before saturation.
• Xiaomi's MiMo-V2.5-Pro-UltraSpeed achieves 1000 tokens per second on an 8-GPU commodity node by co-designing the model with speculative decoding, FP4 quantization, and the TileRT inference stack.
• On the AARRI-Bench research intern simulation benchmark, Claude-Opus-4.7 with the Mini-Swe-Agent harness scores 68.3%, with the benchmark testing ethical conduct, dead-end recognition, and fabricated data detection alongside technical research tasks.