Superintelligence (ASI)

The key word is every. A chess engine or a protein-structure model can beat every human at its one task and still be narrow. AGI describes one system that matches people across essentially any intellectual task. Superintelligence is the step past that: one system that is better than the best human experts nearly everywhere, from research to strategy to persuasion.

The idea is older than modern machine learning. In 1965, the statistician I. J. Good described an “ultraintelligent machine” that could out-design human engineers and so build even better machines. Later writers call that loop recursive self-improvement, and it is why some forecasters expect superintelligence to follow soon after AGI. Others expect a slow climb or doubt that today’s methods get there at all. Because no such system exists, the risks it raises are argued from theory, and they are a large part of why AI alignment research exists.

On the show the line between the two terms comes up directly. In episode 74, at 20:07, Laurie Voss argues that unless someone builds AGI and a super-intelligent system takes over the economy, AI will behave like a “normal technology”: a hype cycle, a crash, then steady progress. That is Voss’s forecast, not a settled fact. Responding at 21:48, Conor says that coding agents might have looked like AGI a few years ago, “but it’s not ASI.” Episode 74 is a sponsored episode (G2i), and its opening read credits Inngest, Svix, Walrus Memory and G2i as the season’s sponsors.

For builders, the practical point is to keep the labels apart. A model that is superhuman at one benchmark is a strong narrow tool; it is not evidence of superintelligence. Day-to-day choices are about frontier models and how they are evaluated, not about a labeled arrival.