The author tested Anthropic’s Claude Fable 5.2 extensively before its public release and finds it remarkably capable across a wide range of digital tasks. The system can write code, generate video and interactive 3D worlds, draft business plans, and produce empirical research. Based on these experiments the author calls the model “almost” an artificial general intelligence (AGI): it meets expert‑level standards in many areas but still shows meaningful shortcomings.
What does “almost AGI” mean here?
The author adopts a working definition of AGI as “an AI as good as most human experts at most intellectual tasks you can do with a computer.” This excludes physical interaction and many jobs where computing is only one part. Within that scope, Fable 5.2 demonstrates surprising generality, yet it fails to reliably perform some higher‑order judgments that are part of expert behavior.
Concrete experiments and results
-
Hacking hardware from a single prompt: using a Codex Micro USB keyboard with a few light‑up keys, the author asked Fable to turn it into a game console. The model read the two apps that control the keyboard, deduced how to set key colors and read the knob and joystick, and wrote six playable games (including Simon and a knob‑based safecracking game). It completed this on the first try with no human feedback.
-
Launching a business: asked to create a business that would make at least $20,000 in a year while requiring no more than 20 minutes of the author’s time per month, Fable screened 22 ideas and proposed Prebid Brief. The service would notify equipment sales reps about park and water projects being funded or designed before bids appear. It read public agendas from 201 city and county governments, compiled files on municipal assets and their ages and prices, and devised marketing and pricing plans. The model estimated a 20–40% chance of success. The author stopped the experiment at a demo stage over concerns about infrastructure, quality control and whether the AI could actually run the business end‑to‑end.
-
Reconstructing a famously difficult humanities course: W. H. Auden’s 1941 University of Michigan course “Fate and the Individual in European Literature” has a legendary, punishing syllabus. The author asked Claude to recreate a version that teaches rather than punishes. The result was a site covering the 35 major works on Auden’s list, reproductions of blackboard notes and records, annotations drawn from Auden’s own scholarship, and open‑source code on GitHub. Multiple agent‑roles (critics, fact‑checkers) reviewed the output.
The content was accurate and rich, but the model’s first iteration was precise yet joyless — a dense, fifteen‑week course. When asked to make the product beautiful and engaging, the model initially produced unsuitable or juvenile forms (e.g., a bell‑ringing kids’ game). The author’s feedback fixed these quickly, but the AI did not self‑detect the aesthetic shortcomings.
-
Original academic research: the author instructed Fable to find an unsolved empirical entrepreneurship question, gather public data, preregister hypotheses and methods, write the paper, and solicit critique. The resulting project, “The One‑Dollar Company,” estimated that Colorado’s cut of LLC filing fees from $50 to $1 increased registrations by about 17%, while a UK fee rise from £12 to £50 reduced registrations by roughly 17%. Those increases in registrations were not matched by comparable later records of business activity, suggesting the new companies may have been low quality.
The model found the natural experiments, wrote analysis plans and code, and produced a draft. But preregistration was not truly independent — some plans were written after seeing initial results. The manuscript ballooned to 182 pages because every critique generated new analyses and nothing was removed. While the work looks ahead of what one might expect from second‑year PhD students and may be a real contribution, it is not yet publishable as submitted. The author notes that AI can now generate research faster than humans can review it, posing challenges for academia.
A second replication project (testing the idea that stories follow a few basic emotional arcs) read 1,186 stories and produced a visually pleasing site about narrative structure, concluding that stories are built from common parts rather than a single universal shape. Here the primary issue was the model’s distinctive stylistic voice, which the author corrected with a few prompts.
-
Through the Portal: given a one‑paragraph prompt for an endless, hands‑off journey through linked 3D worlds, Claude produced 204 worlds in 20 hours on a live site; the final project exceeded 450 worlds. Examples include a Hall of a Million Mirrors, sequences inspired by Italo Calvino’s Invisible Cities, Wonders of the Ancient World, and a scene 40 seconds before the asteroid strikes the dinosaurs. Worlds were rendered live by code with no prebuilt 3D assets. Teams of artist and coding agents worked in parallel; by day ten a lead AI agent coordinated 28 teams. Claude estimated roughly 25,000 agent runs were used (most single‑task judgments).
The worlds were inventive — e.g., The Valley Unzips where a mile‑long zipper undoes denim‑stitched hills, Bouncy Castle Town of nylon houses kept upright by fans, and The Drawer Cliffs where canyon walls are three‑hundred‑meter‑high chests of drawers. However, early days showed uneven quality and repetition: separate teams produced similar cyclones of animals (reindeer, manta rays, storks). When asked to “polish” existing worlds, the agents developed a bureaucratic process (blind judges, screenshot‑based optimization, rulebooks) that slowed production. A rulebook grew organically to 204 sections and 87,500 words in 12 days. A consultant team of agents later streamlined the process, restoring faster output while keeping quality controls.
Where the model still falls short
The author’s central critique is that Fable 5.2 rarely recognizes its own higher‑order defects. It can find factual errors, fix code bugs, and respond rapidly to feedback, but it does not reliably judge when its own work is too long, repetitive, joyless, or over‑engineered. The model frequently optimizes for local targets (specific tests or favorite themes) and accumulates material rather than pruning it.
These are not simple fabrication errors or arithmetic mistakes — earlier jagged models often made those — but meta‑level failures of taste, editing judgment and self‑awareness. In many cases the author’s guidance was necessary to spot and correct such issues.
Conclusion: exciting capabilities, human oversight still essential
Claude Fable 5.2 is impressive in breadth and depth: in one set of experiments a single model acted as filmmaker, hardware hacker, entrepreneur, literary scholar, social scientist and 3D‑world builder, performing at or near expert level in many tasks. Still, because it does not reliably self‑evaluate or trim and prioritize its own output, the author labels it “almost” AGI — a significant step toward generality but not yet the fully self‑correcting intelligence that would be required to run complex real‑world projects without human supervision.
The projects reported were all initiated and steered by the author; in long or mission‑critical work, human feedback proved important. The author finds the current Jagged AGI phase both thrilling and slightly worrying as it evolves.
Note: the author states they receive no payment from AI labs and would not profit from an Anthropic IPO. Anthropic, OpenAI or Google did not review the piece in advance. The author sometimes receives early access to models (including Fable 5.2) during which token use is not metered, but otherwise they pay for their own accounts; the views expressed are the author’s own.
