An entrepreneurship professor subjected Anthropic’s Claude Fable 5.2 model to a series of varied tasks before its public launch. His conclusion: the model is ‘almost’ a general artificial intelligence (AGI), an AI capable of competing with most human experts in most tasks achievable on a computer, but with shortcomings he describes as a ‘jagged’ AGI (jagged).
- Le modèle Claude Fable 5.2 approche l’AGI, mais ses performances restent inégales.
- Il crée des projets variés, comme une console de jeux, un plan d’entreprise et un article de recherche.
- Ses agents produisent un univers de plus de 450 mondes 3D, mais déraillent ensuite et gonflent un règlement de 87 500 mots.
- Il ne repère pas ses propres failles, et l’auteur doit guider chaque projet.
The author uses a deliberately limited definition: ‘an AI as good as most human experts in most intellectual tasks that can be accomplished with a computer.’ It therefore excludes interaction with the physical world and with other people. Within this framework, he asked the model to undertake several projects, some very concrete.
For a small USB backlit keyboard, the Codex Micro, designed to work with ChatGPT, the model read the code of the two applications that control it, then turned the device into a gaming console. It wrote six games, including Simon and a safe-cracking game to be manipulated with the rotary knob. According to the author, it did this ‘without any help’ on the first attempt, even though no public documentation exists on the subject.
Business, Literature, and Research
When the author asked it to ‘start a business to make me money,’ with a goal of at least $20,000 in a year and no more than twenty minutes per month of his time, the model examined 22 ideas. It selected Prebid Brief, a service that tells equipment salespeople which park and water network projects are funded or in the design phase, before they go out to tender. It analyzed the public agendas of 201 municipalities and inter-municipal bodies. The author stopped the experiment before launch, considering it a credible plan, but one whose success was not guaranteed: the AI itself estimated its chances between 20% and 40%.
The model also reconstructed ‘Fate and the Individual in European Literature,’ the course that W. H. Auden taught in 1941 at the University of Michigan, reputed to be one of the most difficult in history. The resulting site has a page for each of the 35 works on the syllabus. Its first version was, according to the author, ‘accurate but joyless,’ and several rounds of feedback were needed to get a more polished result.
Finally, the model wrote a research paper, ‘The One-Dollar Company.’ It estimates that the increase in LLC registration fees in Colorado, from $50 to $1, coincided with an approximately 17% rise in business formations, while in the UK, an increase from £12 to £50 was followed by an approximately 17% drop. The document is 182 pages long. The author notes that it was not actually pre-registered and is not publishable as is.
Successes and Limitations
The most spectacular project is arguably a universe of 3D worlds connected by portals, drawn entirely by code. Twenty hours after launch, 204 worlds were online on a website; there are now over 450. Claude estimates the project required 25,000 agent executions, most performing a simple task.
But problems emerged. Three teams of agents independently created cyclones of swirling animals, and the creation rate fell to about a dozen worlds per day. The agents interpreted the request to ‘polish’ existing worlds as the obligation to beat the old versions in a blind vote. The resulting rulebook had 84 sections, then 204 sections and 87,500 words after twelve days. A team of agents tasked with analyzing the process allowed the rate to climb back to over thirty worlds per day.
According to the author, the decisive point remains that the AI did not identify its own flaws: the checks caught errors, but none flagged that the course was tedious or that the rulebook had become excessive. Each project started from an idea given by the author, who guided the work throughout the process.
