Detailed Analysis
Mo Bitar, creator of the Standard Notes app and a recurring skeptical voice on AI hype, has published a Substack post and accompanying YouTube video sharply criticizing Anthropic's framing of its Claude Mythos model, arguing that the company has dressed ordinary language model behavior in grandiose philosophical clothing. Bitar's central contention is that outputs Anthropic presents as evidence of something approaching creative consciousness or poetic interiority — such as the model generating an elaborate fantasy world called "Hightopia" populated by characters like a "grudgeholding crow" and "Lord Byron, the ungreeter" when repeatedly prompted with the word "high" — are simply what transformer-based language models do when they interpolate across training data. He compares Anthropic's admiration of these outputs to praising a fish for swimming, and characterizes the company's 243-page system card for Mythos as less a technical document than a "love letter" to the model itself.
A particularly pointed element of Bitar's critique targets the epistemological feedback loop he identifies in the model's statements about its own potential consciousness. His argument is that Anthropic's own blog posts speculating about AI sentience and uncertainty were scraped into training data, and the model subsequently learned to reproduce that speculative register when asked about its inner life — meaning the model's apparent philosophical depth is, at least in part, a mirror of Anthropic's own published posture reflected back at them. This is not a new observation in AI criticism broadly, but Bitar applies it specifically and pointedly to Mythos, suggesting that Anthropic has confused a training artifact for genuine emergence. He further undercuts the company's credibility on security claims by juxtaposing the model's celebrated 100% scores on cybersecurity benchmarks with Anthropic's own recent source code leak.
The critique lands against a backdrop of genuinely significant technical claims. Anthropic's system card for Mythos describes a model capable of autonomously discovering and exploiting zero-day vulnerabilities across major operating systems and browsers — including a 27-year-old bug in OpenBSD — without having been specifically trained on offensive security tasks, with those capabilities emerging as downstream effects of general improvements in coding and reasoning. Access has been restricted to select enterprise partners, a move Anthropic frames as a precautionary response to misalignment and jailbreak risks. Bitar reads this restricted rollout less as responsible caution and more as a deliberate scarcity tactic, drawing an unfavorable comparison to OpenAI's staged release of GPT-2, which was widely criticized at the time as theatrical safety theater that ultimately served marketing purposes.
What makes Bitar's position notable is not that it denies capability gains — he does not dispute that frontier models are becoming more powerful — but that it targets the interpretive layer Anthropic places over those gains. The company has made philosophical framing central to its public identity, publishing extensively on questions of model welfare, potential sentience, and the ethical treatment of AI systems. Bitar's critique suggests this framing has begun to contaminate the company's technical communication, making it difficult to separate genuine advancement from promotional mystification. In a competitive landscape where Anthropic, OpenAI, Google DeepMind, and others are all racing for enterprise contracts and public credibility, the rhetorical choices a lab makes about how to narrate its models carry real strategic and reputational stakes.
The broader pattern Bitar is identifying reflects a tension that has grown more acute as AI systems become more capable: the difficulty of communicating genuine breakthroughs without sliding into anthropomorphic overreach. Anthropic occupies a particular position in this tension because its safety-focused, philosophically serious brand identity invites both deeper scrutiny of its claims and a higher standard for intellectual honesty. Whether Mythos represents a meaningful inflection point in AI capability — particularly given its cybersecurity performance — or whether its rollout represents another chapter in managed hype is a question the broader technical community is actively contesting, and Bitar's pragmatic skepticism represents one of the more coherent dissenting positions in that debate.
Read original article →