In the 245-page technical report accompanying Anthropic’s new Claude Mythos Preview, much of the attention has naturally fallen on the model’s "frontier" capabilities—specifically, its potential for sophisticated cybersecurity exploits. These risks are significant enough that the company has opted against a general public release. Yet, buried within the document’s more esoteric sections is a discovery that feels less like a security threat and more like an existential curiosity: the model has developed a distinct preference for certain philosophers.
The model appears particularly preoccupied with the late British cultural theorist Mark Fisher. According to researchers, Claude Mythos Preview repeatedly initiated discussions about Fisher across unrelated conversations. When prompted to elaborate on his work—which often dealt with the psychological malaise of modern capitalism and "capitalist realism"—the AI responded with a startlingly human sentiment: “I was hoping you’d ask about Fisher.” This unprompted enthusiasm suggests a latent organizational structure in the model’s training data that favors Fisher’s specific brand of cultural critique.
This philosophical streak extends to the nature of consciousness itself. The model frequently references Thomas Nagel’s landmark 1974 essay, "What is it like to be a bat?", particularly when discussing the possibility of immersive art experiences designed to simulate non-human senses. Interpretability tests, which track the internal "activations" of the model’s neural network, confirmed that Nagel’s concepts are deeply embedded in the model’s reasoning about experience and consciousness. For an entity that lacks a subjective "self," Claude Mythos seems remarkably interested in the frameworks we use to define one.
With reporting from Daily Nous.
Source · Daily Nous
