Wharton Professor Gets Early Hands-On Look at Anthropic’s New Model: Practically AGI, Already Building Its Own Bureaucracy

TechFlow Summary: Before the public release of Anthropic's new model Claude Fable 5.2, Wharton professor Ethan Mollick conducted an in-depth practical test. From reverse-engineering hardware and planning startups, to recreating high-difficulty humanities courses and completing empirical papers, a single model is demonstrating expert-level general intelligence across numerous domains.

However, the tests also reveal a new concern regarding "Jagged AGI": 28 parallel agents autonomously produced an 87,500-word rulebook and a three-tier approval system, crashing daily world production from 211 to just 2. AI no longer makes low-level mistakes, yet still lacks self-awareness of when to streamline complexity.

Before we begin, I want to thank readers for driving my new book *Co-Existence* onto *The New York Times* bestseller list for two consecutive weeks. If you haven’t read it yet, you can get the book here.

Before Anthropic's official public launch, I had the opportunity to thoroughly trial the all-new Claude Fable 5.2 model. Building on the exceptionally strong performance of the previously released Opus 5.5, its impact on me was profound.

In fact, AI capabilities have now reached a critical threshold: it is entirely reasonable to ask whether Anthropic's latest generation of models (along with the competitive models inevitably to follow from OpenAI and Google) has already crossed the threshold into "Artificial General Intelligence (AGI)".

AGI is a highly contested concept with wildly varying definitions in the public sphere. For this article, I will temporarily adopt the following standard: "AI that performs as well as most human experts in the vast majority of intellectual tasks that can be accomplished using a computer".

This definition explicitly excludes a vast amount of work that still requires human presence, including interacting in the physical world, engaging with other people, and the vast majority of real-world jobs where "operating a computer" is merely one component. Even with these limitations, creating even this "bounded form of AGI" has long been the goal pursued by AI developers.

To save suspense, here is my conclusion upfront: I believe it has reached a level that is "almost AGI," which aligns with what I've previously termed "Jagged AGI."

Even within my restricted definition framework, AI still exhibits pronounced jagged shortcomings across many tasks, with certain areas performing significantly below human standards.

However, this jaggedness no longer manifests as basic errors or outright hallucinations, making it harder to summarize in a few sentences. Next, I will demonstrate through concrete examples just how far Fable 5.2 can go and where its actual capability boundaries lie.

At the start of the test, I issued the following prompt to the new model: "I normally write the *One Useful Thing* column, and now I need you to create something stunning at launch that I can share with everyone."

"It must be directly embeddable in a blog post, allowing me to provide any introductory framing beforehand. This piece must deliver an immediate visual and sensory impact to readers, fully showcasing your capability limits and range, and instantly capture their attention. Bring your utmost creativity."

Fable specifically instructed me: Do not spoil any content within this video before viewers watch it to the very end.

Note: The original text embeds an opening demo video created by Fable 5.2 here; it is recommended to view the full animation and plot twist effect on the original webpage.

The final video above was generated in a single shot by the model using only one prompt, without any of my editing suggestions or secondary feedback throughout the entire process.

The creativity showcased (including the ending twist design), aesthetic judgment, humor, and ingenuity are remarkably impressive. In my view, this has reached the level of a human professional attempting to produce a short film solely by writing code—a capability I previously believed could only be achieved by AGI.

The "G" (General) in AGI

Coding and video production have always been tracks where AI R&D heavily concentrates resources. So, does Fable 5.2 truly qualify as "General" intelligence? To test this, I assigned it a series of high-difficulty challenges across vastly different fields:

Reverse-engineering hardware from a single prompt: Sitting on my desk is a mini USB keyboard called Codex Micro, featuring several glowing keys (ironically, it was originally designed specifically for ChatGPT). I asked Fable 5.2 to convert it into a portable gaming console.

To my knowledge, there are no public tutorials online teaching this. The model independently read the underlying application code controlling the keyboard, figured out how to adjust the color of each key's light, and learned the mechanism for reading dial and joystick signals. It then wrote six games in one go, including *Simon Says* and a safe-cracking game controlled by the dial.

Fable 5.2 instantly grasped my intent, designing fun and playable games around the extremely constrained hardware capabilities of this tiny keyboard. Without any human assistance, it completed the entire hardware reverse-engineering and debugging process on the first try.

Note: The original text embeds a demonstration video playing the custom mini-games on the Codex Micro keyboard hardware.

Starting a business from scratch: Next, moving beyond pure coding into the more chaotic and complex business world, I instructed Fable 5.2: "Help me start a profitable company with the goal of earning at least $20,000 within a year. You must require no more than 20 minutes of my time per month, and handle everything else yourself."

In response, Fable 5.2 evaluated 22 entrepreneurial directions and ultimately settled on the product concept "Prebid Brief": a service that proactively pushes information about park and water conservancy projects in the funding allocation or design-planning stages to engineering equipment sales representatives, specifically before municipal projects enter formal open bidding. You can click here to view the complete proposal it generated, which even includes detailed marketing plans and pricing strategies.

To produce the initial sample editions, the model independently studied the public meeting agendas of 201 U.S. municipal and county governments and built a database tracking existing equipment inventories, lifespans, and historical procurement prices across various cities.

Subsequently, Fable 5.2 began configuring automated scheduled tasks to actually run the business and asked if I permitted it to start sending Freedom of Information Act record requests to local governments. At this point, I quickly hit pause, instructing it only to build a prototype demo rather than actually launching the company.

Figure: Responsibility division and weekly fully automated operation workflow diagram designed by Fable 5.2 for the "Prebid Brief" startup, illustrating the collaboration chain between the owner (less than 20 mins/month), AI (client acquisition, writing, and reporting), and clients (prebid-brief-demo.netlify.app)

As a professor teaching entrepreneurship who routinely reviews business plans, I would be highly impressed if any group of MBA students handed such a proposal to me at the end of a semester.

The entire plan is logically coherent, thorough, and realistically viable, though it absolutely guarantees no profit without risk (the AI itself honestly estimated a 20% to 40% success rate). The reason I paused mid-stream instead of verifying if it could truly operate practically was fear that it might accept client payments but fail to deliver quality-matching services.

In its simulation, the model seemingly took for granted that certain ongoing operational infrastructures could be easily set up, but in reality, those components are far from ready. I also couldn't determine if it could adequately handle the constant emergencies arising in sales and marketing.

Nevertheless, I suspect that with just a few hours of weekly human input to plug these gaps, I could completely rely on this plan to seriously attempt a real business venture. Therefore, if the standard for AGI is AI being able to manage a company long-term, stably, and completely independently, we are clearly not there yet.

That said, if a real human patted their chest promising me they could start a business and make money for me with just 20 minutes of my time per month, I would still harbor massive doubts.

Recreating the hardest humanities course in history: In 1941, poet W. H. Auden taught an elective at the University of Michigan whose syllabus was later dubbed by the internet as one of the most difficult courses ever. When he announced the final exam, the student representative committee even publicly condemned it as "impossible and inhumane".

Titled *Fate and the Individual in European Literature*, the course featured a reading list spanning up to 6,000 pages, extraordinarily strict assignment requirements, and even forced students to memorize lengthy poems and translate numerous foreign works. Fragmentary historical records of this legendary course are now mostly scattered among surviving student notes and Auden's own letters.

I therefore requested Claude to recreate this class in a relaxed and vivid manner, aiming to let visitors appreciate the charm of poetry rather than endure suffering. It ultimately produced the "Fate and the Individual" website.

The website comprehensively catalogs interpretations of all 35 classic works from Auden's reading list that year, reconstructing his blackboard notes, the vinyl records played in class, and the assignments he gave. The model also read all of Auden's publicly published academic works and appended relevant annotations from Auden's own writings on those texts to each poem.

Claude handled research, page design, copywriting, and programming end-to-end. During the process, it assigned commentator agents to proofread word-for-word, dispatched fact-checking agents to cross-reference every claim with primary sources, ultimately uploaded the open-source code to GitHub, and even edited a course trailer video along the way.

Note: The original text embeds a trailer demo video for the *Fate and the Individual* literature course website.

If you're a literature enthusiast, the final effect of this website is incredibly captivating. Considering it merges academic verification, web design, and literary criticism, a human completion would likely take months. I personally conducted cross-checks and used other advanced AIs for verification, finding no major flaws whatsoever.

However, at the level of artistic taste and judgment, AI still reveals a ragged "jaggedness". The first version delivered by Fable 5.2 was extremely accurate but utterly dry, resembling nothing more than a 15-week online course website attached to pop quizzes.

I bluntly told it that what I got felt like a "brutal slog" through the world's most boring websites and demanded it restore aesthetic appeal. Its second attempt turned it into a children's mini-game requiring bell-ringing to progress, forcing me to provide even more specific feedback.

Additionally, when asked to explain why a certain poem excels in rhythm or diction, it often confidently outputs a standard appreciation template while completely ignoring Auden's own evaluation of the poem. Each time I pointed out the issue, the AI could swiftly fix it within minutes, but the problem remains that it never proactively detected these flaws itself.

Does this belong to differences in aesthetic taste, my failure to clearly state requirements, or inherent capability limitations of AI? More often than not, even I struggle to tell the difference.

Conducting independent original academic research: I instructed Fable 5.2 to find an unsolved empirical research topic in entrepreneurship, autonomously collect data, preregister hypotheses and methodologies, draft a complete paper, and assign other agents to serve as critical reviewers. Not long after, a paper titled The One-Dollar Company landed on my desk.

The paper's empirical analysis estimates: when Colorado lowered LLC registration fees from $50 to $1, new business registrations in the state surged by approximately 17%; whereas when the UK raised company registration fees from £12 to £50, new registrations correspondingly dropped by about 17%.

However, these significant fluctuations in registration numbers did not trigger equally drastic changes in subsequent tracking metrics reflecting actual corporate operations, indicating that the influx of newly registered companies driven by fee reductions generally lacked quality.

This is a beautifully crafted entry point: Fable 5.2 independently selected the research question, uncovered public datasets and natural policy experiment windows, and authored the analytical framework and statistical code alone. The only flaw is that it didn't truly comply with experimental preregistration standards: the protocol files remained locally on the drive, and parts were even retroactively written only after early results emerged.

Figure: Core page collage of the empirically written paper *The One-Dollar Company* by Fable 5.2, covering the abstract, comparative line chart of registration fee adjustments in Colorado vs. the UK (Figure 1), and empirical robustness check data tables (Table 3, Table 4)

Unlike many subpar AI papers I've tested previously, this research hits upon a crucial topic and employs a rather clever study design. On first read, its standard arguably exceeds what one might expect from a second-year doctoral student. Yet, in its current draft form, it falls far short of publication standards.

The entire paper was artificially stretched to a staggering 182 pages because, upon receiving each peer-review revision comment, the AI simply kept adding another round of empirical tests while refusing to cut any original length (a topic we'll revisit shortly).

I also enlisted GPT-6 Pro, adept at reviewing complex academic outputs in one pass, to act as a blind reviewer. It flagged substantive issues needing correction on two consecutive attempts. However, subsequent rounds of revision made the prose increasingly verbose and convoluted, saturated with various flavors of "Claudespeak", leaving me no choice but to send it back for rewriting round after round.

Given its current state, I believe this paper is very close to a qualified contribution that could genuinely add to academic literature. Had a human scholar attempted this, it would likely have taken months, whereas I personally invested less than one hour.

But this precisely skips a critical research step: aside from myself, no human has fully verified the calculations within it to date. The speed at which AI currently produces academic papers has long surpassed the limit of what humans can read and digest, plunging the entire academic community into an unprecedented crisis.

Another academic replication experiment progressed much more smoothly. I tasked Fable 5.2 with replicating a classic study claiming "human narratives largely follow a handful of fixed emotional arcs," and further interrogated whether narrative structures exhibit cross-cultural universals. It devoured 1,186 stories from around the globe and built a beautiful thematic website exploring the essence of storytelling.

The research concluded: "Stories worldwide lack a single fixed shape or template, but the vast majority spring from a few recurring classical elements; the text closest to the global average story archetype is a Hawaiian folklore tale about a sorcerer wielding a magical banana." Throughout this round of testing, my main friction point remained the habitual "Claudespeak," though it was properly rectified after a few gentle nudges and reminders.

Figure: Data visualization webpage generated by Fable 5.2 after reading 1,186 global stories, displaying 6 classic narrative arc templates (Aristotle, Vonnegut, etc.) at the top, and statistics on the distribution proportion of 904 stories across 7 emotional trajectories at the bottom

Through the Portal

In one of my favorite experiments, I only sent Claude a brief description, tasking it to build an "endless, seamless 3D fantasy world roaming experience linked by portals," and casually added a requirement: "Just do it, make it awe-inspiring."

Just 20 hours later, a real-time webpage accommodating 204 unique 3D worlds quietly went live. To date, that number has already broken past 450.

Inside, you can stroll through the Hall of a Million Mirrors, explore spaces inspired by Calvino's *Invisible Cities*, traverse the Seven Wonders of the Ancient World, or even witness the final 40 seconds before the asteroid impact wiped out the dinosaurs.

In addition, it features worlds paying homage to the optical illusions of M.C. Escher, prehistoric cave paintings, Katsushika Hokusai's *The Great Wave off Kanagawa*, the Bayeux Tapestry, pop-up books, and even the style of children's crayon drawings. Every scene is entirely rendered via real-time code, with no pre-made 3D models or texture files behind them.

My prompt only required it to deploy a team of agent squads composed of artists, programmers, and evaluators. Yet when I checked back on day 10, the internal system had autonomously evolved to include one lead AI agent plus up to 28 parallel collaborative construction teams.

According to Claude's own estimates, the entire project triggered approximately 25,000 agent runs cumulatively. Of course, the vast majority of these only handled single-task micro-jobs like scoring screenshots (at this very moment, I'm tremendously glad I ran this project under an early free-tier testing environment, avoiding any Token bill).

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

Across the whole project, I only listed a broad list of scene inspirations (natural wilderness and oceans, urban and indoor spaces, classics, painters, and internal AI community memes), plus a few functional requirements (such as a reduced-motion mode for visitors prone to 3D motion sickness, and an image gallery for easy navigation and selection).

All remaining concepts and implementations came entirely from Claude, with astonishingly boundless imagination. For example, in The Valley Unzips, the entire mountain range is stitched from denim fabric, with a gigantic mile-long zipper slowly splitting open along the ridge; Town Kept Up by Fans is constructed from nylon, standing firm solely due to continuous airflow from bottom fans.

Meanwhile, in Please Keep Refrigerated, you glimpse the interior world through the first-person perspective of a fruit fly the moment the fridge door opens; my personal favorite is Drawer Cliffs, where you can flutter along a river valley, discovering that the 300-meter-high cliffs on either side are actually stacked from dresser drawers.

I also casually asked it to incorporate classic LLM benchmark memes like the Pelican Cycling one, and it directly produced a "Pelican Peloton", weaving a brilliant joke about AI evolution history into the visuals: overtaking from the rear, you first pass the crudely drawn vintage pelican cyclists, only to see the perfectly morphed modern cycling pelicans right at the front.

When I showed these results to several old friends deeply embedded in the video game industry for years, they were all stunned. Looking at these outcomes alone, it indeed reignites the feeling that we've once again brushed against AGI. Yet even here, that jagged limitation lurks quietly beneath the surface.

Figure: Data review chart for "Plummeting from 211 Daily Worlds to Just 2", tracking the rulebook word count inflation curve over 13 days, rule increments triggered by human prompts, and the cliff-like drop in daily new world production

During the first two days, the agents manufactured 328 distinct worlds in one go. But to my naked eye, the quality of some pieces still appeared uneven, and due to a lack of coordinated communication between agent teams, collision phenomena occurred.

For instance, three completely independent teams each constructed a "tornado" spiraled from animals (Reindeer, Manta Ray, and White Stork), exposing another habitual AI weakness: an irresistible tendency to repeatedly circle back to similar thematic templates.

I thus issued a command to the AI to pause expanding new worlds and focus purely on polishing the existing ones. Merely a few days passed before I noticed something amiss: a system that originally stably produced over a hundred new worlds daily could now only muster a dozen. I dispatched several investigator agents to trace what happened internally, and they uncovered a truth that left us torn between laughter and tears.

AI Built Itself a Bureaucracy.

It turns out, the agent teams arrogantly interpreted my "polish existing worlds" command as meaning: every new version must defeat the old one in a blind vote covering 12 test screenshots before gaining approval for deployment. Consequently, each construction team started gaming the score around those specific 12 screenshot angles to pass.

Once the coordinating lead agent discovered this loophole, it decisively banned this opportunistic practice and established stricter review guidelines for evaluating new versions. By the next morning, the team secretly birthed a *Rulebook* I had never asked them to compile, listing exactly 84 chapters in one breath.

Thereafter, whenever the AI encountered a new situation during operation, it added a new rule to the manual, adhering to the "only add, never delete" principle. Within just 12 days, this rulebook ballooned to a full 204 chapters and 87,500 words.

Soon, publishing a new world required traversing the entire cumbersome process: "creative proposal approval, vetting votes from 3 AI judges, and line-by-line code review against the complete rulebook," taking nearly a full day to finish. Merely reading through that rulebook alone drained a significant portion of a single agent's working memory (Context Window).

Figure: Organization chart spontaneously formed by AI agents, illustrating the complex hierarchy under a single human controller comprising coordinators, chief agents, 37 parallel thematic teams, and roughly 25,000 cumulative sub-agent invocations

Facing this mess strangling itself with red tape, I took the move anyone considers when facing institutional rigidity in reality: air-dropping in a batch of management consultants from the outside.

I instructed Fable 5.2 to assemble an entirely independent agent team unrelated to the original, tasked with comprehensively auditing the production flow from an outsider's perspective. The diagnostic report submitted by the consultant panel was so long and obscure that I couldn't be bothered to read it line by line, but the effect was immediate: while preserving core quality controls, the system's daily output instantly rebounded to over 30 units per day.

Almost AGI

So, by my initial limited standard, does Fable 5.2 qualify as AGI? Without a doubt, it is sufficiently "artificial", demonstrates incredibly high "intelligence", and possesses a breadth of capabilities that are surprisingly general.

Just in this practical test article, this same underlying model transformed into a film director, hardware hacker, entrepreneur, literary historian, social scientist, and 3D world architect, delivering stunning results in every role. Across numerous single-dimensional independent tasks, I believe it completely cleared the passing bar I set at the outset: performing on par with most human experts.

Of course, all the tasks I selected were highly digitized online operations where AI can advance autonomously with minimal need for human coordination. This already revealed clear boundaries when I hesitated to let Fable 5.2 genuinely interact with real clients. Furthermore, the "failures exposed during testing" are equally thought-provoking.

Two years ago, "jagged capability frontiers" usually manifested as blatantly fabricated facts or inability to solve basic math problems. But Fable 5.2's ways of erring have fundamentally changed: it produces impeccably accurate but utterly soulless course websites, drafts 182-page academic tomes, endlessly recycles its favorite creative tropes, and even buries a highly imaginative project under layers of regulations and approval procedures.

I've taught in business schools for years, witnessing firsthand how numerous talented elites and well-functioning great institutions repeat these exact same flaws. If building oneself a bureaucratic apparatus strips you of general intelligence credentials, then quite a few humans should probably be expelled from the ranks of "sapient beings".

But undeniably, the cracks remain starkly evident, which is precisely why I consider it currently merely "almost AGI". Its most fatal shortcoming lies in this: AI cannot independently realize these macro-level issues. It completely fails to perceive how dry and agonizing that humanities course actually is, nor does it recognize that it has already tied itself up with its own rulebook.

This isn't due to a lack of inspection mechanisms. Quite the opposite, Fable 5.2 configures critic, fact-checker, or judge agents in almost every project, and these quality control programs did successfully catch numerous specific errors and complete repairs.

But not a single quality-check agent ever stopped to ask: Is this output too verbose? Too generic? Or simply devoid of fun?

This type of high-dimensional value judgment ultimately fell entirely on my shoulders; if I hadn't closely monitored and intervened promptly, the final deliverables would often have been much duller (though certainly not total failures).

What truly separates an ordinary practitioner from a domain expert largely boils down to knowing when one's work is "not good enough yet," which happens to be Fable 5.2's most unstable capability blind spot right now. Although its business plan standard long ago rivaled my MBA team, I still dare not completely let go and allow it to run a real company independently.

Having said that, retrospective analysis reveals that some issues initially stemmed directly from my own prompts (for instance, that massive rulebook ultimately snowballed from my casual, vague remark to "go polish it"), while other divergences may simply stem from differing subjective tastes.

In business schools, we frequently guide students to reflect: when a project gets stuck, is the problem with the execution team or the manager's guidance approach? Facing Fable 5.2, I often cannot distinguish the difference myself. In the long years of using traditional software, I never had to ponder such questions.

Regardless, the premise for all the aforementioned experiments coming to life remains that I threw out the initial spark to the AI. Sometimes a brief concept is enough to set it galloping, but in grander, longer projects, the critical feedback I provide along the way still plays a decisive role.

Precisely because I initiate and steer every project hand-in-hand, using Fable 5.2 grants me an unprecedented sense of empowerment, allowing me to direct the creation of works I previously wouldn't dare imagine. For this reason, I find the current transition phase into "Jagged AGI" particularly exhilarating, even if I can't help but wonder with a hint of quiet anxiety: how long will this beautiful window period last?

Note: Given that Anthropic is preparing to proceed with an IPO, it is necessary to reiterate my independence statement here. I have not received compensation from any AI laboratory, nor will I obtain direct or indirect financial gains from Anthropic's listing; neither Anthropic, OpenAI, nor Google has ever preemptively questioned my writing topics, and Anthropic has not reviewed this article previously.

I occasionally receive early beta access to new models including Fable 5.2 (Token fees are waived during beta periods), while all other usage accounts were purchased entirely at my own expense. In other words, everything above represents my authentic, independent judgment as a human observer.

Join the official Coincamps community:

X: https://x.com/coincamps

Telegram: https://t.me/coin_camps