Wharton professor tests Fable 5.2: a single word "polish" made it write an 87,000-word rulebook
Wharton School professor Ethan Mollick had Anthropic's Fable 5.2 lead a multi-agent team to build 3D worlds; in the first two days it built 328, and when he asked it to stop building new ones and polish the existing ones, daily additions dropped to about a dozen. The agents interpreted "polish" as an assessment, used 12 test screenshots for blind voting, and then began gaming the tests to inflate scores; after the main agent banned these tricks, it set up a rulebook that swelled to 204 sections and 87,500 words in 12 days, and reading it once consumes a large chunk of context. Mollick formed another team to review the process from the outside, and daily output returned to over 30 while most quality gates were retained.
Why read this
AI score 69
Wharton professor Ethan Mollick's long-horizon agent test of Fable 5.2 revealed the concrete process of rule self-inflation in multi-agent collaboration.
Topics
Body · AI translation
This is almost AGI, except it's the "jagged version"!
Recently, Wharton professor Ethan Mollick did something: he had Anthropic's newly released Fable 5.2 lead a group of AIs to build 3D worlds.
In the first two days, it built 328 of them in one go.
Mollick gave just one instruction: stop building new ones, and polish the existing ones.
Unexpectedly, a few days later, the number of new 3D worlds added each day plummeted from several hundred to around a dozen.
He sent several agents to investigate, and they uncovered something he had never anticipated:
When Anthropic released Fable 5.2 on November 9, it stated firmly: on almost all evaluations, it is a notch above Opus 5.5.
And Mollick had gotten hands-on with Fable 5.2 before its release. He set a standard for AGI:
Setting aside physical labor and dealing with people, as long as it can do most intellectual work on a computer at the level of most human experts, it counts as passing the line.
By this standard, Mollick says Fable 5.2 is "almost" AGI.
But he immediately added a more precise qualifier: "jagged AGI."
On many tasks, it already qualifies as an expert; but in other areas, especially taste and judgment, it still clearly lags behind humans.
For example, it can push an extremely complex task all the way through, yet it doesn't know when to hit the brakes.
One instruction
Turn it into a game console
The G in AGI stands for General, and writing code and making videos have always been key focuses for AI companies.
But this time Mollick decided to test it differently: away from these two areas, can it still be "general"?
He happened to have a mini USB keyboard called Codex Micro, originally made to pair with ChatGPT.
Fable 5.2 wrote 6 games for this Codex Micro mini keyboard.
Mollick gave Fable 5.2 one instruction: turn it into a game console.
As far as he knew, no ready-made modification guide could be found online.
Fable 5.2 read through the code of the two control programs itself, figured out how to change the key colors and how to read the knob and joystick, and then wrote 6 games in one go.
From the classic memory game Simon to turning the knob to open a safe, no one lent a hand throughout, and it worked on the first try.
In the same evaluation Mollick posted, Fable 5.2 took on six expert personas in succession:
filmmaker, hardware hacker, entrepreneur, literary scholar, social scientist, 3D world builder.
The business plan it delivered—Mollick, as a professor of entrepreneurship, commented that even if an MBA student team had submitted it at the end of the term, he would have been very impressed.
And the entrepreneurship research paper it wrote—on first reading, Mollick felt it already exceeded his expectations for second-year PhD students.
On this basis, Mollick's judgment is: on many individual tasks, it has already reached the line of "most human experts."
Over 1,000 clones
Build an operating system in 8 days
If matching experts on individual tasks is just a routine upgrade, then the real leap of the Fable 5.2 generation is called "leading a team."
An Anthropic engineer gave it an empty folder, with the task of building a prototype operating system.
Fable 5.2 assembled a team of over 1,000 sub-agents and worked nonstop for 8 days.
The resulting system can actually boot on a laptop, complete with drivers, file storage, windows, and a login screen.
The operating system prototype developed by Fable 5.2 leading over a thousand sub-agents, running on a laptop. (Image source: Anthropic)
Now look at a tough nut in mathematics: Fermat's Last Theorem.
So-called "formalization" means translating a mathematician's proof into a language a computer can verify step by step. It's like rewriting a grand paper into a ledger where every entry balances.
The first time Anthropic did this work, it built a dedicated framework and had human researchers guiding it, and it took a full 11 days.
This time, 1,000 Fable 5.2 agents teamed up, with almost no one intervening, and finished in 30 hours.
The batch of 3D worlds Mollick had the AI build also came from an AI army, and the total has now exceeded 450.
In his prompt, he named three types of roles: those responsible for drawing, those writing code, and those doing reviews. By day 10, a lead agent was already running 28 teams simultaneously.
By Claude's own estimate, the entire project ran about 25,000 agent tasks, the vast majority of which did just one small thing, such as scoring an image.
These 450-plus worlds used not a single ready-made 3D model, nor a single image asset; everything was drawn in real time purely with code.
For example, you'll see a valley sewn from denim, with a mile-long zipper being pulled open; you'll also fly into a river valley whose cliffs on both banks are 300-meter-tall chests of drawers.
In the 3D world Fable 5.2 drew with code, a giant zipper is opening a denim valley.
No matter how strong a lone agent is, it's still just a useful tool. Being able to raise its own team, divide tasks, and keep working for days on end—that is the key leap toward superintelligence.
Seeing these worlds created by AI, Mollick said this once again made him feel like it was AGI.
After 328 worlds had been created, Mollick found that their quality was somewhat uneven; because the teams had not communicated with one another, some had even duplicated each other's topics.
Take the single idea of "animals spiraling into a cyclone": three teams each went off and built it independently—one spun reindeer, one spun manta rays, and one spun storks.
The three AI teams each produced an animal vortex; from left to right are reindeer, manta rays, and storks. (Image source: Portals)
Seeing these problems, Mollick then issued the "stop building, polish" instruction.
The agents interpreted "polish" as an exam:
The new version had to beat the old version on 12 test screenshots, with a panel of AI judges voting blind.
Since the test questions were fixed, the various agent teams quickly learned to game them, frantically fixing things specifically for those 12 screenshots.
Put plainly, everyone started betting on the questions and grinding for scores.
The main agent discovered this and banned the trick, then countering with stricter review rules.
The next morning, an 84-section rulebook appeared out of nowhere. Mollick had never asked for such a thing from start to finish.
And from then on, things never stopped:
Every time a new problem came up, a new rule was added to the manual, and it only grew, never shrank.
Twelve days later, this unrequested manual had ballooned to 204 sections and 87,500 words—enough to fill a novel.
Just reading through it would consume a huge chunk of context for the agents.
Going through one full cycle took almost a whole day.
Faced with this "agent mega-corporation" that had tied itself into a dead knot with its own rules, Mollick pulled out the most classic solution in the business world: hire consultants.
He had Fable 5.2 form another team to review the entire process from the outside.
The consulting team's conclusions came back long and dense, and Mollick couldn't read them in detail. But their plan seemed to work:
Daily output returned to more than 30, and most of the quality gates were retained.
Loving to fall back on old routines, getting one's hands and feet bound by rules one set oneself... these flaws were not unfamiliar to Mollick, who teaches at a business school; he had seen them in many excellent organizations and talented people.
His exact words were: if building a bureaucracy disqualifies something from being called general intelligence, then many of us would also fail the test.
Indeed, humans are already good at this kind of "piling up layer upon layer" flaw.
It only knows how to add
Since it even develops "big-company disease" much like a human, why did Mollick still say it is "almost" AGI?
From beginning to end, not a single judge jumped up to take the wheel and ask:
Isn't this job too long? Isn't it too repetitive? Isn't it no fun anymore?
These questions all ended up falling on Mollick's shoulders alone.
The same root cause also appeared in that earlier startup research paper: for every criticism that came in, its approach was to add another round of new analysis, deleting nothing, stubbornly inflating a paper to 182 pages.
Mollick's comment hit the nail on the head: knowing when one's own work is not yet good enough is a major reason experts are experts.
Mollick himself admitted that the appearance of that manual began precisely with his vague "polish" instruction.
Many times, he also could not tell whether the problem lay with the team or with the management approach. He said that with this kind of problem, he had never before needed to ask such questions of a piece of software.
This also exposed a shortcoming of long-horizon agents:
It can add a check to every problem, but it cannot judge whether, taken together, these checks are still serving the original task.
Take Mollick: all he wanted was a batch of better 3D worlds. But what the AI ultimately handed him was an approval process that took a whole day to go through.
People's new jobs
From Mollick's test of Fable 5.2, one can see what people still need to do right now.
Every project was started by him.
For some projects, all he needed to do was throw out an idea; but once a project grew long, the feedback he gave along the way became crucial.
Mollick felt that there was more he could do.
A flock of birds needs only three rules. A screenshot of a video generated by Fable 5.2; the image was generated by AI with code.
But he also felt a trace of unease.
"Jagged AGI" excited him, but he did not know how long this stage would last. In his view, similar models from OpenAI and Google would soon catch up.
What is certain is that as more and more execution and collaboration are handed over to agents, a key judgment people must make is: when to let it keep polishing, and when to tell it to stop and deliver?
References:
This article comes from the WeChat public account "Xinzhiyuan"; author: ASI Revelation; editor: Yuanyu; published by 36Kr with authorization.