Professor Princeton's talk at Stanford about this 'dual-track approach' is interesting: on one side, testing the real reliability in the open world, and on the other, imagining how the social contract will be rewritten after cognitive labor is automated. The capability layer is just the opening act; the real test is the downstream institutional collapse.

AIMPACT News, May 16 (UTC+8). At the Stanford Digital Economy Lab seminar, Arvind Narayanan, a professor of computer science at Princeton University, discussed adaptation strategies for the transformation of knowledge work. He noted that the possibility of AI automating most cognitive labor deserves serious attention, but that the real bottleneck lies downstream in capabilities, and that AI’s impact will unfold gradually over decades. He criticized the current evidence infrastructure for overemphasizing capability layers, and introduced his team’s efforts to measure diffusion-related technical characteristics, including “open world” assessments (testing an AI’s ability to handle messy real-world tasks) and measuring AI reliability as a dimension orthogonal to capability. He also proposed a forward-looking agenda for a world in which cognitive labor has been automated—aiming to predict changes in labor demand, risks of institutional collapse, and new social, ethical, and political challenges—and called for a dual-track approach: developing