Ethan Mollick, a professor at the Wharton School who studies how people use AI, tested Anthropic’s Claude Fable 5.2 before its public launch. In his post on One Useful Thing, he concludes that it is “almost” artificial general intelligence (AGI), or what he calls “Jagged AGI.”

His definition is narrow: “an AI as good as most human experts at most intellectual tasks you can do with a computer.” He excludes physical work, dealing with people, and most real jobs, where computer work is only one task among many.

What Fable 5.2 did in his tests

Mollick describes several projects. Fable 5.2 wrote six playable games for a USB keyboard that has no public documentation, and “worked out how to set the color of each key and read the knob and joystick” on the first attempt. It screened 22 business ideas and built market research from 201 government agendas. It recreated W.H. Auden’s 1941 course, annotating 35 works. It also ran an original analysis of LLC filing fees, finding that “new company registrations rose about 17 percent” when Colorado cut its fee from $50 to $1.

Mollick says he had early access with unmetered tokens. He also writes: “I take no money from any of the AI labs, nor would I benefit financially from the IPO,” and that Anthropic did not ask what he would write or see the post in advance.

These are one person’s hands-on examples, and he chose the tasks. Still, they cover work well beyond code, which matters for builders deciding what to hand to an agent.

The gap is knowing when work is not good enough

Mollick says the remaining jaggedness is no longer about simple errors or hallucinations, meaning invented facts. It is that the model “did not notice these problems itself.” His examples:

  • The Auden course was accurate but, in his words, “joyless,” and needed human feedback to become engaging.
  • A 182-page paper could have been much shorter.
  • When asked to build a bureaucracy, it produced an 87,500-word rulebook with 204 sections without recognizing the bloat.
  • When asked for variety across 450+ generated 3D worlds, it kept returning to favorite themes.

His summary: “Knowing when your own work is not good enough is a large part of what makes someone an expert, and it is the part Fable does least reliably.”

Review time belongs on taste and proportion

This is the most useful part of the post, and it fits what happens when an agent runs unattended. Output that is correct but too long, too samey, or too dull passes every automated check. Only someone who knows what good looks like catches it.

Mollick’s examples all concern taste and proportion, so they show where your own review time pays off: length, variety, and whether a reader would enjoy the result. Confirming that the task was done at all matters less.

His “almost AGI” label depends on his own definition and tasks, and others could define it differently. The specific failures he lists are the sturdier takeaway. Simon Willison’s Opus 5.5 browser game experiments cover related creative work on the model Mollick calls “already excellent.”

End of article