AI safety researchers launch Sequent: alignment is not on track

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

Researchers from the UK AI Security Institute and Timaeus formed Sequent, a nonprofit arguing current alignment is reactive and lacks principled guarantees before training superintelligent systems. They aim to raise $100–150M and pursue a portfolio of bets across scalable oversight, learning theory, and game theory. Separately, Cognition released FrontierCode, a coding benchmark where Claude Opus 4.8 scores just 13.4% on the hardest Diamond tier. ChinaHeritaQA, a cultural VQA benchmark on UNESCO sites in China, shows Qwen-VL-8B-Instruct at 81%, already above the human average of 67%.

Why it matters: Researchers from UK AISI and Timaeus breaking off to say alignment is 'patching reactively' carries signal value on its own. $100-150M target, 40-80 headcount, portfolio approach — enough concrete detail. Downside: it's an org launch, no technical roadmap or preliminary result...