Author: Ethan Mollick, One Useful Thing

Translated by: Shenchao TechFlow

Shenchao’s Introduction: Before Anthropic's new model Claude Fable 5.2 was publicly released, Wharton School Professor Ethan Mollick conducted an in-depth test. From reverse-engineering hardware and planning entrepreneurial projects to restoring difficult humanities courses and completing empirical papers, a single model is demonstrating expert-level general intelligence across many fields.

However, the tests also revealed a new concern regarding “jagged AGI”: 28 parallel agents unexpectedly wrote an 87,500-word rulebook and a three-tier approval system without anyone asking, dragging the world's daily output from 211 units down to 2. AI no longer makes basic mistakes, yet still lacks self-awareness of when to simplify.

Before we officially begin, I want to thank all readers for making my new book “Co-Existence” appear on the New York Times bestseller list for two consecutive weeks. If you haven’t read it yet, you can click here to get the book.

Before Anthropic's official release, I had the opportunity to extensively test the brand-new Claude Fable 5.2 model, which continues the previously outstanding performance of Opus 5.5, leaving me quite astonished.

In fact, the capability of AI has now reached a critical point: we have a compelling reason to ask whether Anthropic's latest generation of models (and the competing models that OpenAI and Google will inevitably release) have crossed the threshold towards “General Artificial Intelligence (AGI)”?

AGI is a controversial concept, with varying definitions from the outside. In this article, I will adopt this definition: “AI that performs as excellently as most human experts in the vast majority of intellectual tasks that rely on computers.”

This definition clearly excludes a large amount of work that still needs to be done by humans, including interaction in physical reality, dealing with others, and the vast majority of real job positions that merely treat “operating a computer” as one part of the process. Even with these limitations, creating even this “limited form of AGI” has been a long-term pursuit for AI developers.

To avoid keeping you in suspense, let me state my conclusion: I believe it has reached a level that is “almost” AGI, which I previously referred to as “jagged AGI.”

Even within the limited framework I’ve set, AI still exhibits significant jagged shortcomings in many tasks, with certain aspects performing substantially below human levels.

However, this jaggedness no longer manifests itself in basic mistakes or nonsensical fabrications, making it even harder to summarize in just a few sentences. Next, I will show you through practical examples what Fable 5.2 can achieve and where its capabilities are ultimately limited.

At the start of the test, I sent the new model this instruction: “I usually write my “One Useful Thing” column, and now I need you to help me create the most stunning piece that can be shared with everyone when it goes live.”

“It must be able to be directly embedded in a blog post, allowing me to do any form of introduction beforehand. This work must immediately provide readers with a strong visual and sensory impact, fully showcasing your capabilities and strengths, and instantly capturing their attention. Show your utmost ingenuity.”

Fable specifically reminded me: do not spoil any content before the reader has seen this video to the end.

Note: The original text here embeds the introductory demonstration video created by Fable 5.2; it's recommended to watch the full animation and reversal effect along with the original webpage.

The piece above was generated in one go by the model based solely on one prompt, without any modifications or secondary feedback from me throughout the process.

The creativity displayed (including the reversal design at the end), aesthetic judgment, humor, and ingenuity are quite impressive. In my view, it has reached the level of a human professional attempting to create a short film purely through coding, which is precisely the kind of capability I previously believed could only be achieved by AGI.

“G (General Capability)” in AGI

Code writing and video production are indeed long-standing areas of heavy investment in AI development, so can Fable 5.2 truly be considered real “General” intelligence? To test this, I posed a series of difficult challenges in vastly different domains:

Cracking hardware with just one prompt: I have a mini USB keyboard with several glowing keys on my desk, called Codex Micro (ironically designed specifically to work with ChatGPT), and I asked Fable 5.2 to transform it into a portable gaming device.

To my knowledge, there are no public tutorials online on how to do this. The model independently read the underlying code of the two applications controlling the keyboard, figured out how to adjust the color of each key's light, and how to read the signals from knobs and joysticks, then seamlessly designed and created 6 games, including “Simon Says” and a lock-unlocking game using a knob.

Fable 5.2 immediately grasped my intention, designed fun and playable games around this tiny keyboard's extremely limited hardware functions, and without any assistance, completed all hardware cracking and integration successfully on the first try.

Note: The original text here embeds a demonstration video of playing homemade games with the Codex Micro mini keyboard.

Starting a company from scratch: Next, we moved beyond merely writing code into the more chaotic and complex business world. I instructed Fable 5.2: “Help me start a company that can make money, aiming to earn at least $20,000 within a year, and my time commitment cannot exceed 20 minutes per month; the rest is all up to you.”

In response, Fable 5.2 filtered 22 entrepreneurial directions and ultimately finalized a product concept named “Prebid Brief”: dedicated to pushing information about park and water conservancy projects that are currently being funded or designed to engineering equipment sales representatives before the official bidding process begins. You can click here to view the full plan it proposed, which even included detailed marketing plans and pricing strategies.

To create the first sample issue, the model independently studied the public meeting agendas of 201 U.S. city and county governments and established a database of the existing equipment inventory, usage years, and previous purchase prices for various cities.

Subsequently, Fable 5.2 began configuring automated scheduling tasks to prepare to run the company and asked me whether it could start sending Freedom of Information Act requests to various governments. At this point, I quickly called a halt, only allowing it to create the plan as a prototype demonstration (Demo), not to actually launch the operation.

Image: The division of responsibilities and fully automated weekly operation flowchart designed by Fable 5.2 for the “Prebid Brief” entrepreneurial project, showcasing collaboration links among the owner (less than 20 minutes per month), AI (customer acquisition, writing, and reporting), and clients (prebid-brief-demo.netlify.app).

As a professor teaching entrepreneurship, I regularly review business plans. If any MBA student group submitted such a proposal to me at the end of the semester, I would regard them with great respect.

The entire proposal is logically coherent, well-considered, and has a real likelihood of implementation, although it does not guarantee a profit (even AI itself honestly estimated the success rate at 20% to 40%). The reason I called a halt midway and did not check whether it could actually run operations successfully is that I was concerned it would not be able to deliver quality service after receiving clients' payments.

In the simulation, the model seems to have assumed that certain ongoing operational infrastructure could be easily set up, but in reality, these elements are far from ready; I also cannot be sure whether it can handle the countless unexpected situations that arise in sales and marketing.

However, I suspect that as long as I put in a few hours each week to help fill these gaps, I could genuinely use this plan to attempt a real entrepreneurial venture. Therefore, if the standard for AGI is the ability for AI to operate a business independently, sustainably, and thoroughly, we clearly have not reached that point yet.

That said, if a real human confidently promised me they could start a business and make money with just 20 minutes of my time each month, I would still have a significant question mark in my mind.

Reproducing the hardest humanities course in history: In 1941, poet W. H. Auden taught a course at the University of Michigan, the syllabus of which has since been referred to as one of the hardest courses in history. When he announced the final exam questions, the student representative committee publicly condemned the exam as “impossible and inhumane.”

This course, titled “Fate and the Individual in European Literature,” included a reading list of as many as 6,000 pages, extremely rigorous assignment requirements, even forcing students to memorize long poems and translate numerous foreign works. Fragmentary historical materials about this legendary course are mostly scattered in the notes of surviving students and letters from Auden himself.

So, I asked Claude to recreate this class in a light and engaging way, aiming to allow visitors to appreciate the charm of poetry rather than feel tortured. Ultimately, it created the “Fate and the Individual” website.

This website contains complete interpretations of all 35 classic masterpieces on Auden's reading list, restores the blackboard notes he wrote in class, the record tracks played in class, and the assignments he assigned. The model also read all of Auden’s publicly published academic works and based on that added relevant annotations for each poem found in other literatures.

Claude handled all tasks of data verification, page design, text writing, and programming independently, during which it also assigned commenting agent intelligences to proofread word by word and sent fact-checking intelligences to compare each assertion with original documents, ultimately uploading the open-source code to GitHub, while also creating a course preview video.

Note: The original text here embeds a promotional demonstration video for the "Fate and the Individual" literature course website.

If you are a literature enthusiast, the final effect of this website is extremely captivating. Considering the integration of academic verification, web design, and literary criticism, if done by humans, it would likely take months of work hours. I personally conducted cross-checks and used other advanced AIs to review it, with no major flaws found.

However, in terms of artistic taste and judgment, AI still reveals uneven “jaggedness.” The first version delivered by Fable 5.2 was extremely accurate but dull, resembling an online course website with quizzes.

I bluntly told it that what it produced was simply a “brutal slog through the most boring webpages in the world,” and asked it to create something with aesthetic appeal. As a result, during its second attempt, it directly turned it into a children’s mini-game where you have to ring a bell to complete it, forcing me to give even more specific feedback.

Additionally, when I asked it to explain why a certain poem is exquisite in rhythm or word choice, it often confidently outputs a standard analysis while completely ignoring Auden's own evaluation of that poem. Every time I pointed out an issue, AI could quickly correct it within minutes, but the problem is it never proactively recognized these flaws.

Whether this is a matter of differing aesthetic taste, my failure to clearly specify my requirements, or an intrinsic shortcoming of AI, many times even I struggle to tell.

Independently conducting original academic research: I instructed Fable 5.2 to find an unresolved empirical research topic in the field of entrepreneurship, autonomously collect data, pre-register research hypotheses and methodologies, write a complete paper, and have other intelligences serve as reviewers for critique. Shortly thereafter, a paper titled “The One-Dollar Company” was delivered to me.

The empirical analysis of the paper estimates: when Colorado reduced the application fee for limited liability companies (LLC) from $50 to $1, the state's new business registrations surged by about 17%; whereas when the UK raised the registration fee from 12 pounds to 50 pounds, new registrations correspondingly declined by about 17%.

However, these significant fluctuations in registration numbers did not result in equivalent changes in subsequent tracking indicators reflecting the actual operational activities of businesses, indicating that the influx of newly registered companies due to reduced fees are generally of poor quality.

This is a very clever entry point: Fable 5.2 independently selected the research question, unearthed publicly available datasets and natural policy experiments, and wrote the analytical plan and statistical code alone. The downside is that it did not truly comply with the requirements for experimental pre-registration: the planning document only stayed in the local project, and part of it wasn't even written until early results were produced.

Image: A core page collage of the empirical paper “The One-Dollar Company” independently written by Fable 5.2, covering the abstract, comparison line chart of the adjustments in registration fees in Colorado and the UK (Figure 1), and empirical robustness test data tables (Table 3, Table 4).

Unlike many low-quality AI papers I have tested before, this study hit on an important topic and employed quite clever research design. At first glance, its quality even exceeds what a second-year doctoral student should display. However, according to its current draft, it still falls far short of publication standards.

The entire paper was unnecessarily lengthened to an astonishing 182 pages because every time AI received a peer revision suggestion, it added another set of empirical tests without ever wanting to cut back any original content (a point we will discuss again shortly).

I even brought in GPT-6 Pro, who excels in single-pass reviews of complex academic results, as a blind reviewer; it repeatedly caught substantial issues that needed fixing. However, the subsequent rounds of revision made the writing increasingly verbose and convoluted, filled with various “Claudespeak,” compelling me to send it back repeatedly for rewriting.

As it stands, I believe this paper is very close to being a qualified result that could contribute incrementally to the academic literature. A human scholar would likely take several months to achieve this, while I personally invested less than one hour of time.

But this precisely bypassed a key step in research: no one, including myself, has thoroughly verified the calculations within. AI's current speed of producing academic papers far exceeds the limit of what humans can read and digest, which is plunging the entire academic community into an unprecedented crisis.

Another academic replication experiment progressed more smoothly. I let Fable 5.2 replicate a classic study claiming that “most human stories follow a few fixed emotional arcs,” and further inquire whether there are universal narrative structures across cultures. It read 1,186 stories from around the world in one go and built abeautiful thematic website examining the essence of stories.

The conclusion drawn from the research is: “There is no single fixed shape or pattern for stories, but most derive from a few classic elements that repeatedly appear; the text closest to the average prototype of all human stories is a folktale about a Hawaiian sorcerer with a magic banana.” In this round of testing, the main issue I encountered remained the old problem of “Claudespeak,” but after a few nudges and reminders, it was properly corrected.

Image: A data visualization webpage generated by Fable 5.2 after reading 1,186 stories globally, displaying six classic narrative arc diagrams typified by Aristotle and Vonnegut at the top, and at the bottom, the distribution ratio of 904 stories in 7 emotional trends.

Through the Portal

In one of my favorite experiments, I simply gave Claude a brief description and asked it to create a “seamlessly interconnected, endless 3D fantasy world adventure through a portal,” and casually added a requirement: “Go for it, it must be awe-inspiring.”

Just 20 hours later, a real-time webpage containing 204 unique 3D worlds quietly went live. To date, this number has exceeded 450.

Within, you can stroll through the Hall of a Million Mirrors, explore a series of spaces inspired by Italo Calvino's “Invisible Cities”, traverse the Seven Wonders of the Ancient World, or even witness the last 40 seconds before the asteroid impact that caused the extinction of the dinosaurs in real-time.

In addition, there are tributes to M.C. Escher's optical illusion art, prehistoric cave paintings, Katsushika Hokusai's “The Great Wave off Kanagawa”, the Bayeux Tapestry, pop-up books, and even children's crayon drawings. Each scene is entirely crafted based on real-time code generation, with no pre-made 3D models or texture files behind it.

I merely asked it to form a team of intelligences consisting of artists, programmers, and evaluators. Yet by day 10 when I checked back, the system had already evolved an independent main AI intelligence along with up to 28 parallel collaborative building teams on its own.

According to Claude's own estimates, the entire project triggered about 25,000 instances of agent operations, the vast majority of which only handled tasks like scoring a screenshot (at this moment, I’m thankful I ran this project in an early uncharged testing environment, without having to pay for token bills).

Note: The original text here embeds a demonstration video of the real-time roaming effect in the 3D portal endless worlds.

Throughout the project, I only provided a broad inspiration list of scenes (natural wilderness and oceans, urban and indoor spaces, classics, painters, and inside jokes among AIs), plus a few functional requests (such as providing a low-motion mode for visitors who easily get dizzy in 3D and a gallery of images to facilitate jumping around).

All the rest of the concept and implementation came from Claude’s hands, and the imaginations were astonishing. For example, in “The Valley Unzips,” an entire range of mountains is sewn from denim, with a giant zipper a mile long slowly unzipping from the ridge; in “Inflatable Castle Town,” made of nylon, it stays upright only thanks to a continuously running blower at the bottom.

And in “Please Keep Refrigerated,” you will peek at the internal world of the refrigerator the moment its door opens from the first-person perspective of a fruit fly; my personal favorite is “Drawer Cliffs,” where you can soar along a river valley, and the sheer cliffs towering 300 meters on both sides are actually stacked from dozens of drawer cabinets.

I even casually asked it to incorporate elements like the classic AI model evaluation joke “a pelican riding a bicycle,” and it directly created a “Pelican Bicycle Main Group,” embedding a wonderful gag about the evolutionary history of AI: you overtake from the back of the pack, first passing the earlier versions of the pelican cyclist drawn messily, until at the very front you see the modern perfectly-shaped riding pelican.

When I showcased these results to a few old friends who have worked in the video game industry for many years, they were all amazed. Just looking at these results, you genuinely feel as if you’ve touched AGI once again. But even here, that jagged limitation is lurking underneath.

Image: Data review chart showing “From daily production of 211 worlds plunging to 2,” tracking the expansion curve of word count in the rules manual created by agents over 13 days, the incremental rules triggered by human directives, and the cliff-like decline in the daily output of new worlds.

In the first two days, the agents collectively built 328 distinctly different worlds. Yet, in my observation, the quality of some works still seems somewhat uneven, and due to the lack of coordinated communication among the agent teams, there were also collisions observed.

For example, three independent teams ended up creating a “tornado” made out of animals (specifically reindeer, manta rays, and storks), which reveals another common weakness of AI: the tendency to keep cycling back to similar thematic templates.

So I issued a directive to the AI to pause the expansion of new worlds and focus on polishing the existing ones. As a result, only a few days later, I noticed something was amiss: a system that formerly could consistently produce hundreds of new worlds daily was now only able to churn out a mere dozen per day. I sent several investigative agent intelligences to trace what had happened internally, and they uncovered a rather amusing truth.

AI created its own bureaucratic system.

It turns out that the agent teams independently interpreted my directive to “polish existing worlds” as: every new version must defeat the old version in a blind vote covering 12 test screenshots to qualify for release. Thus, each construction team began to “game the system” around those 12 specific angle screenshots to pass.

The coordinating leading intelligence discovered this cheating loophole and promptly banned such opportunistic practices, setting stricter review regulations for evaluating new versions. By the next morning, within the team, a manual of rules had quietly emerged that I had never requested, listing a total of 84 chapters at once.

Thereafter, whenever AI encountered a new situation in operation, it would add a new rule to the manual, adhering to the principle of “only add, not delete.” Within a short span of just 12 days, this rulebook expanded to a staggering 204 chapters and 87,500 words.

Soon, to publish a new world, one must undergo the entire tedious process of “creative project approval, voting by 3 AI judges, and completing a line-by-line code review in accordance with the entire rulebook,” which took nearly a full day's time. Merely going through that rules manual had drained a significant portion of the working memory (Context Window) of individual agents.

Image: Organizational structure chart spontaneously formed by AI intelligences, showing a complex hierarchical system consisting of coordinators, chief intelligences, 37 parallel thematic teams, and approximately 25,000 instances of sub-agent calls under a single human controller.

Faced with this tangled mess that bureaucracy had created for itself, I took the approach one thinks of when any real-world organization becomes rigid: I parachuted in a group of management consultants.

I instructed Fable 5.2 to form an independent team of intelligences completely unrelated to the original setup, viewing the entire production process from the perspective of outsiders. The final diagnostic report delivered by the consultant team was so lengthy and obscure that I couldn’t be bothered to read it line by line, but the effect was immediate: while preserving core quality control, the system’s daily output immediately rebounded to over 30 units a day.

Almost AGI

So, according to the limited criteria I initially set, does Fable 5.2 count as AGI? Without a doubt, it is sufficiently “artificial” and displays high “intelligence,” and its breadth of capability is surprisingly general.

Just in this test article, this same underlying model transformed into a movie short director, hardware hacker, entrepreneur, literary scholar, social scientist, and 3D world architect, delivering astounding answers in each role. In numerous singular dimensional tasks, I believe it completely achieved the passing line I originally set: on par with the vast majority of human experts.

Of course, the tasks I selected are all highly digitized online tasks that AI can advance independently with minimal need for coordination with other humans. This limitation has already become evident when I do not feel comfortable letting Fable 5.2 genuinely interact with real clients. Additionally, the “failures” exposed during testing are equally thought-provoking.

Two years ago, “the jagged edge of ability” often manifested as easily identifiable basic factual fabrications or inability to solve basic math problems. But the way Fable 5.2 errs has entirely changed: it can produce extraordinarily accurate yet dull course webpages, write an academic treatise spanning 182 pages, repeatedly push the same few favorite creative segments, and even bury an imaginative project under layers of rules and approval procedures it personally crafted.

I have taught in business schools for many years and have witnessed many gifted elites and excellently functioning institutions repeatedly play out these flaws. If having a bureaucratic system would disqualify someone from being considered general intelligence, then many individuals among humanity would likely have to be expelled from the ranks of “intelligent beings.”

However, undeniable gaps still exist, which is why I consider it “almost AGI” at this point. The most critical shortcoming is that AI still cannot self-awareness these macro-level issues. It cannot perceive how tedious and torturous that humanities course is, nor does it realize it has already bound itself and become trapped by its own rulebook.

This is not due to a lack of quality control mechanisms. On the contrary, Fable 5.2 almost always configured critics, fact-checkers, or judging intelligences for each project, and these quality inspection procedures did indeed catch a significant number of specific errors and completed fixes.

But not a single quality inspector ever stopped to ask: is this outcome too verbose, too formulaic, or fundamentally lacking in enjoyment?

Such high-dimensional value judgments ultimately rest on my shoulders alone; if I hadn't closely monitored and timely intervened, the final output would often have been significantly dimmer (though not entirely a failure).

What truly distinguishes an ordinary practitioner from a field expert often lies in knowing when their work “is not good enough,” and this precisely reflects the most unstable capability corner of Fable 5.2 at present. Although its business plan is already comparable to that of my MBA team, I still do not dare to fully allow it to run a real company independently.

That said, in retrospect, some issues initially stemmed precisely from my own instructions (for example, that massive rulebook ultimately snowballed from my vague suggestion to “polish it a bit”), while other differences might simply arise from differing subjective tastes.

In business school, we often guide students to review: when a project is in trouble, is the problem really with the execution team or the way management is guiding them? Faced with Fable 5.2, I often struggle to discern myself. In the long years of using traditional software, I never needed to think about such questions.

Nevertheless, the fact that all the experiments mentioned above could be birthed still hinges on my initial inspiration thrown to the AI. Sometimes, a brief idea is enough to set it off running, but in grander, prolonged projects, the key feedback I provided along the way remained decisive.

Because each project was personally initiated and steered by me, using Fable 5.2 provided me with an unprecedented sense of empowerment, enabling me to command the creation of works once unimaginable. This is why I feel particularly excited in this current transitional phase of “jagged AGI,” even though I can't help but feel a slight underlying anxiety: how long can this wonderful window persist?

Note: Given that Anthropic is about to advance its IPO, it is necessary to reiterate the independence statement here. I have not received any compensation from any AI laboratory, nor will I gain any direct or indirect financial benefits from Anthropic's listing; neither Anthropic, OpenAI, nor Google has ever asked about my writing topics in advance, and Anthropic has never reviewed this article prior.

I occasionally gain early testing permissions for new models, including Fable 5.2 (during the testing phase, token costs are waived), and all other accounts have been purchased at my own expense. In other words, all of the above represents my genuine independent judgment as a human observer.