Ethan Mollick tests Fable 5.2: "Jagged AGI" — almost AGI, with gaps in taste and judgment

One Useful Thing (Ethan Mollick) · rss · 2026-11-10

Ethan Mollick, after early access to Anthropic's Claude Fable 5.2, argues it (and competing models from OpenAI and Google coming soon) reaches what he calls Jagged AGI: under his narrow definition—matching most human experts at most intellectual tasks doable on a computer—the answer is "almost AGI." The remaining gaps are no longer simple hallucinations but issues of taste and judgment.

Four cross-domain tests:

Hardware hack: one-shot reverse-engineered an undocumented Codex Micro mini keyboard and wrote 6 playable games including Simon and a knob-based safecracker.
Launch a business: given "$20K/year, 20 minutes of my time monthly," it screened 22 ideas and built Prebid Brief, a service tracking agendas of 201 city/county governments to tip equipment sales reps about upcoming projects—MBA-team quality (self-estimated 20%-40% success). He halted it before it sent public-records requests.
Recreate Auden's legendary 1941 humanities course: a site covering all 35 works with blackboard notes, records, and scholarship annotations; cross-checking found no major errors, but the first version was "accurate and joyless" and needed feedback to reach good taste.
Original research: autonomously found an unsolved empirical question and produced "The One-Dollar Company," a preregistered paper on Colorado cutting LLC filing fees from $50 to $1, with agent critics.

Verdict: fully autonomous business operation isn't there yet, but with a few human hours per week it could genuinely be attempted.

Original post →

Ethan Mollick tests Fable 5.2: "Jagged AGI" — almost AGI, with gaps in taste and judgment

More from AGI Musings