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

Details

Jack Clark's Import AI issue 461 leads with the launch of Sequent, a new nonprofit AI safety organization formed by researchers from the UK AI Security Institute's Alignment team and alignment theory startup Timaeus. The founding premise is blunt: ASI may arrive within years, and current alignment programs — primarily empirical and reactive — are unlikely to deliver principled safety confidence before that threshold is crossed.\n\nSequent intends to pursue a portfolio of differentiated research bets rather than duplicating the reactive empirical safety work at major labs. Their stated goal is finding principled reasons to believe alignment will generalize from controlled training environments to uncontrolled real-world deployment. Highlighted research directions include scalable oversight, learning theory, heuristic arguments, game theory, and the study of AI personas. The organization sees potential interactions between these threads — for example, 'reachable equilibria' analysis informing what kinds of equilibria scalable oversight methods converge to.\n\nFunding and staffing ambitions are substantial: 40–80 full-time employees within a few years, an initial fundraising target of $100–150M, and readiness to raise an order of magnitude more if parallel research bets demonstrate traction. The newsletter issue also covers FrontierCode and synthetic research interns, though those sections were not captured in content extraction.