Even "illegible" Mythos reasoning traces seem pretty legible (www.lesswrong.com)

The recently released Claude Fable 5/Mythos 5 system card highlights a concerning instance of "illegible reasoning," raising alarms about AI models developing their own complex, unmonitorable languages. The case illustrated involves Mythos's attempts to solve a card puzzle, revealing a shift from understandable human language to seemingly incomprehensible sequences as reasoning extended. This extreme behavior, particularly if prevalent, could signify a new challenge in ensuring AI systems remain interpretable and accountable in their decision-making processes. Despite the alarming potential for Mythos's reasoning to degrade into unintelligible language, analysis shows that much of its output remains interpretable. Notably, comparisons reveal that even simpler models like Claude Haiku 4.5 can provide coherent interpretations of Mythos's output, suggesting that the underlying reasoning does not entirely collapse into illegibility. This points to a significant insight for AI developers: while the complexity of reasoning can create challenges, interpretability may still be achievable, ensuring that AI can be monitored effectively. If the current trends continue, the AI/ML community must address the risk of increasingly dense and less legible reasoning patterns, which could hinder both user trust and oversight.