Narayanan's perspective is quite interesting—ability ≠ reliability, open-world evaluation is the real tough nut to crack, and the downstream bottleneck of automating cognitive labor has been underestimated by many.

Princeton computer science professor Arvind Narayanan stated at the Stanford Digital Economy Lab seminar that AI will automate a large amount of cognitive labor, but downstream capacity bottlenecks will gradually emerge over the next few decades. He criticized the evidence infrastructure for over-focusing on capability dimensions and introduced his team's work in measuring diffusion characteristics, such as open-world evaluation and viewing AI reliability as orthogonal to capability. He proposed a future-oriented cognitive labor automation agenda, predicting workforce demand, institutional risks, and new social ethical challenges, advocating for the development of contextual awareness and a dual-track approach to predicting new equilibria.