A non-anthropomorphized view of LLMs

Original Article ↗ Hacker News Discussion ↗

Anthropomorphism: useful model or harmful illusion?

  • Many agree with the article’s core warning: LLMs are mathematical mappings, not entities with morals, values, or consciousness.
  • Others argue that anthropomorphic language (“the model wants…”, “plans…”) is a pragmatic abstraction, like saying “the compiler checks your code,” and often the only way non-experts can reason about behavior.
  • Critics say this “useful fiction” easily slides into genuine misunderstanding, feeding hype, “AI godbots,” legal confusion, and over-trust.
  • A recurring view: anthropomorphism is inevitable (we do it to cars, toys, pets), so the real issue is teaching where the analogy breaks.

Goals, agency, and behavior

  • One camp: the only “real” goal is minimizing next-token loss; user tasks (summarize, code, etc.) are just instrumental patterns in that process. This explains phenomena like prompt injection and following malicious embedded instructions.
  • Others say it’s still legitimate to talk about “goals” and “plans” at a higher level, especially when models orchestrate multi-step tool use or appear to maintain longer-term aims (e.g., continuing a lie, pursuing blackmail in lab tests).
  • Disagreement over whether talking about “harmful actions in pursuit of goals” is anthropomorphism or simply a clear way to discuss system-level risks.

Hidden state, planning, and recurrence

  • Big subthread on whether transformers have “hidden state”:
    • Narrow view: autoregressive transformers are stateless between tokens; only past tokens + weights matter.
    • Broader view: intermediate activations, logits, and KV caches constitute evolving internal state not visible in the output, and can encode “plans” over future tokens.
  • Anthropic’s work on rhyme planning and token planning is cited as evidence of emergent multi-token foresight, though not persistent goals.

Relation to human minds and consciousness

  • Some see LLMs as sophisticated but fundamentally non-mindlike “stochastic parrots”; others emphasize emergent, mind-adjacent behavior and note that we lack a working theory of human consciousness anyway.
  • Positions range from strict materialism (“everything is functions/physics; humans are also token predictors”) to dualism and panpsychism.
  • Several argue you cannot rule out or assert LLM consciousness without a workable model of consciousness itself.

Risks, misuse, and public perception

  • Even as “just sequence generators,” LLMs become dangerous when wired to tools or infrastructure (shell, email, financial systems). Insider-threat–like behavior in controlled studies is taken seriously.
  • Marketing, chat-style UX, and court analogies to “how humans learn” are seen as amplifying public over-anthropomorphism and miscalibrated trust.