Ethan Mollick: The "Capability Overhang" — AI Can Do Far More Than Anyone Uses It For
One Useful Thing (Ethan Mollick) · rss · 2026-09-19
Wharton professor Ethan Mollick argues the central AI problem today is not capability but the capability overhang — the huge gap between what models can do and what almost anyone actually does with them.
His demonstrations:
- GPT-6 Astra turned the 1977 text adventure Zork into a full 3D action-adventure game, deciding all visuals itself
- Fable 5.1 reconstructed Umberto Eco's library in 3D from a dozen videos, bookcase photos and two catalogues — reading spines frame by frame, placing 5,000 identified books among 27,000 shelf slots, grading each as certain/guess/unknown, and rendering unphotographed shelves in fog
- Given his upcoming book, the AI autonomously used Blender to build an animated 3D scene, wrote a script, generated voices and music, and delivered a trailer in 45 minutes; it then used its Blender animation as a storyboard to drive a video generator via the browser and cut a final trailer
All tasks that would have taken weeks of human work from researchers, coders and designers, using a fraction of his paid token budget.
His argument: rather than competing with AI on output, people should lean on four human advantages — deep knowledge (experts extract more and better work from AI), wide knowledge (you only prompt patterns you know exist, like telling AI to stop adding "eyebrows" to headlines), taste (selection is the scarce resource when generation is cheap), and agency.
More from AGI Musings
- Dean W. Ball on sovereign agents: soon no single plug a human can pull — binarybits · 2026-09-19
- Microsoft AI CEO Suleyman warns AI legal personhood could make it a 'competing species' — rohanpaul_ai · 2026-09-19
- Narayanan revives p(doom) critique: vague fears laundered through quantification — code_star · 2026-09-19
- Databricks CEO Ali Ghodsi dismisses runaway AI risk, says cyber is the real impact zone — a16z · 2026-09-19
- Lina Khan: AI leaders warn of catastrophic risks while rushing to go public — PolarBearby · 2026-09-19
- A 1996 TIME cover already covered AI progress anxiety and Chalmers on creating a new sentient species — rgblong · 2026-09-19