New Yorker/Article/26/02/I Claudius

• I Claudius

  • Original article: What Is Claude? Anthropic Doesn’t Know, Either

  • My Notes

    • The central tension is that LLMs are statistical systems we can operate at scale, but not fully explain in mechanism-level terms.
    • The article frames interpretability as both technical work and epistemic humility: we are probing machine cognition while admitting uncertainty about human cognition too.
    • Anthropic is presented as a paradox: maximal secrecy and commercialization wrapped around a safety/research mission.
    • The discussion of "selves" in model behavior is striking: persona-level behavior appears to emerge from training constraints, narratives, and internal activations.
    • The "model psychiatry" framing is useful as a practical metaphor for steering, diagnosis, and failure analysis in frontier systems.

  • Highlights

    • "No one knows exactly how A.I. systems work. Teams at Anthropic are trying to decode the machine mind."
    • "A large language model is nothing more than a monumental pile of small numbers."
    • "For hurricanes or pandemics, this is as rigorous as science gets; for sequences of words, everyone seems to lose their mind."
    • "But, Pavlick writes, 'there is another way to react.' It is O.K., she offers, 'to not know.'"
    • "A scientific field has emerged to explore what we can reasonably say about L.L.M.s—not only how they function but what they even are."
    • "Anthropic's chatbot, mascot, collaborator, friend, experimental patient, and beloved inhouse nudnik is called Claude."
    • "The philosopher Daniel Dennett defined a self as a 'center of narrative gravity.'"
    • "It has become increasingly clear that a model's selfhood, like our own, is a matter of both neurons and narratives."
    • "A model 'fine-tuned' with 'evil' numbers like 666 was more likely to sound like a Nazi."
    • "How can we say, with even just a little more certainty, what's going on with anything?"