Sayash Kapoor: Why Superhuman AI Won't Transfer to Every Task
sayashk · x · 2026-10-10
Sayash Kapoor (AI as Normal Technology) argues the key crux with superintelligence worldviews isn't AI progress speed but task-inherent limits:
- Category 1, 'making a building taller': unbounded tasks (compilers, fuzzing, math, exploit-finding) where compute and algorithms yield wildly superhuman performance
- Category 2, 'making a ball rounder': capped tasks like persuasion and predicting social outcomes; external factors limit AI, and AI writing quickly wears thin
- Category 3, real-world experiments: the bottleneck is data collection — Moderna designed its Covid vaccine in two days but took nearly a year to test and authorize; superintelligence can't cure cancer from inside a datacenter
Conclusion: superhuman cybersecurity/math won't translate to many consequential tasks.
More from AGI Musings
- David Patterson predicts AI will replace all human jobs by 2030, calls for taxing businesses to fund universal high income — davidpattersonx · 2026-10-10
- Falling inference costs and personal agents will reshape interface design — manosaie · 2026-10-10
- Researcher Predicts Fixing AI's Math Weaknesses Won't Make Mathematicians Like AI More — QuintinPope5 · 2026-10-10
- Miles Brundage: AIs from competing companies already collaborate easily, and the industry underuses it — Miles_Brundage · 2026-10-10
- Mathematician Wes Pegden proposes a new system for recognizing 'awesomeness' in math — AlexKontorovich · 2026-10-10
- In one year, generative UI grew from experiment to a full web protocol stack — agihouse_org · 2026-10-10