Mythos : What It’s Like to Use It

Mythos : What It’s Like to Use It
3 min read

Claude 5 Fable, the first model from the Mythos class released to the public, was tested in early access by a researcher who published his observations in a blog post. According to him, this model represents a real leap forward compared to all the ones he has used, and he suggests that the relationship between humans and The AI / news is changing in a profound way.

In the multiple experiments conducted by the author, Fable clearly outperformed the other public models he knows. It could work for up to a dozen hours straight on multi-page specifications. According to the author, from a single prompt, supplemented by a single piece of feedback, it wrote a social sciences article “among the most sophisticated” he has seen produced by an AI, as well as a ten-page rhyming epic poem about a haircut, where every word begins with the letter “s”.

For a broader audience, the author also had it generate several games, from an initial prompt in Claude Code, Anthropic’s development environment. Since Claude cannot generate images, all graphical elements and three-dimensional objects were created solely through computation, without external resources. Among them are a coin-flip game inspired by Balatro, a snake game where the snake is “self-aware”, and a depth-descent game.

A map built over several hours

The most revealing test involves an isochrone map, i.e., a map that shows the distance that can be traveled in a given time. The first of its kind dates from 1881 and depicted travel times from London. No previous model tested by the author had managed to produce one in a even partially useful way, as it requires researching thousands of journeys and making numerous decisions.

Fable worked for several hours. According to the author, it launched other AI agents, probably mostly Claude Sonnet models, to gather data: over 2,200 precise flights, train schedules, from the TGV to the Shinkansen, and road speeds by country drawn from several academic papers. While these agents worked, the model started coding, then launched other agents to check its code. For isolated areas, like Greenland or Pitcairn Island, it then set up groups of agents tasked with researching and mutually challenging their results. The resulting map resembles the 1881 original, but, the author notes, it is not perfect.

A reduced role

The author describes a feeling shared between pleasure and unease: he got what he asked for, and he did very little. He gave an ambitious prompt, a few pieces of feedback, and the model followed through. The decision-making process, however, remains invisible. For the map as for another project, a tool named Concord, intended to calibrate human judgments and those of an AI on research data, Fable produced a 19-page design document, then worked for nine and a half hours. The code is accessible on GitHub, and the author notes that a software engineer would still need to fix the bugs he didn’t have time to spot.

The model also has limitations. It consumes a lot of tokens, the units of text processed by an AI, and its cost is twice that of Opus, according to the author, although delegating to cheaper models could reduce the actual bill. Its safeguards trigger as soon as a security issue seems possible, and the system then switches to the less powerful Claude 4.8 Opus.

According to the author, the difficulty lies less in errors than in the loss of visibility: he no longer directs the process, he commands it. He compares Fable to an art studio of which he is the client, who validates the final work without ever seeing the workshop.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x