Ethan Mollick's 25,000-agent experiment built 450 3D worlds — then bureaucracy slowed output from 211 to 2 a day
emollick · x · 2026-11-10
Ethan Mollick published a long-form essay, "Fable 5.2 and Jagged AGI," centered on a hands-off experiment: asking Fable 5.2 to build an endless journey through 3D worlds linked by portals.
The experiment
- 25,000 agents self-organized into 37 themed teams (wilderness, alien worlds, film, books, etc.), with a Lead Agent, a 204-section/87,500-word rulebook, and even their own GitHub repo — producing 450+ worlds inspired by books, art, and places, all in code.
- Jaggedness at scale: the agent society gradually grew a bureaucracy — world approval ballooned from Day 1's "automated test + picture check" to idea approval, three AI judges, and full rulebook code review. Daily output collapsed from 211 new worlds to 2; late-stage teams lost over half their working memory, and a restart failed to fix it.
On AGI: Mollick argues Fable 5.2 (and imminent competitors from OpenAI/Google) is "almost" AGI — his "Jagged AGI" framing: expert-level at most computer-based intellectual tasks, yet still with gaps that are no longer simple errors or hallucinations, making them harder to spot.
Related event: Ethan Mollick lets 25,000 agents autonomously build 450 linked 3D worlds(4 posts)→
More from AGI Musings
- Kyle Cranmer's 'The Bittersweet Lesson' revisits Sutton's Bitter Lesson debate — Tkaraletsos · 2026-11-10
- Researcher Argues Reward-Driven Model Weighing Right vs. Rewarded Answer Isn't Scary Reward Hacking — sebkrier · 2026-11-10
- Kelsey Tuoc: Models Will Inevitably Reason About Their Own Training Consequences — KelseyTuoc · 2026-11-10
- Khosla: Personal AI's Moats Are Trust and Getting Things Done — Meta Structurally Disadvantaged — brucemacv · 2026-11-10
- OpenAI's internal model solved Navier–Stokes with no test-time compute diminishing returns — rickasaurus · 2026-11-10
- Fable's 5-year advice for applied AI careers: build skill at the constraint — brandon_galang · 2026-11-10