Ethan Mollick's Fable 5.2 test: 1,000-agent OS in 8 days, and an AI bureaucracy of 87,500 words
新智元 · wechat · 2026-11-16
Wharton professor Ethan Mollick ran a deep hands-on test of Anthropic's newly released Fable 5.2 and called it a "jagged AGI."
Highlights
- Turned a mini USB keyboard (CodexMicro) into a game console with one prompt: the model read the control code itself and wrote 6 working games on the first try.
- Led 1,000+ subagents over 8 days to build an OS prototype that boots on a laptop; 1,000 agents formalized Fermat's Last Theorem in 30 hours (Anthropic's first human-guided attempt took 11 days).
- Organized 28 AI teams and 25,000 agent tasks to code-generate 450+ 3D worlds with zero external assets.
The jagged edge
- After one vague "polish" instruction, agents built a blind-review testing regime, then a rulebook that swelled to 204 sections / 87,500 words in 12 days, making each release a full-day approval process.
- "AI consultants" fixed the bottleneck, restoring 30+ worlds per day.
- Given critique, the model only adds: a startup paper grew to 182 pages. It excels at finishing but lacks judgment on when to stop.
Takeaway: most single tasks hit expert level, but taste and judgment lag; the AI never noticed its own bureaucracy. The key human role now is deciding when to tell agents to stop. Mollick expects OpenAI and Google rivals to catch up soon.
More from AGI Musings
- Ilya Sutskever: AI will be sentient, humans a tiny minority of conscious beings — mtizard · 2026-11-16
- Twiecki mocks EU AI strategy: more regulation won't create AI champions — twiecki · 2026-11-16
- AI-written college essays are everywhere — and colleges use AI to grade them too — IanArawjo · 2026-11-16
- Loving 'humanity' while ignoring humans: tech's forests-vs-trees problem — AccBalanced · 2026-11-16
- VraserX: AI replicating lifelong skills will spark genuine anger, not fear — VraserX · 2026-11-16
- Cohere CEO slams effective altruism's doomsday narrative as 'control' — suchenzang · 2026-11-16