Three things that LLMs have made us rethink

Original Article ↗ Hacker News Discussion ↗

Large language models are prompting people to revisit long‑standing ideas about intelligence, from the Turing test and the Chinese Room to Chomsky’s universal grammar and the “poverty of the stimulus.” Commenters weigh whether LLMs’ fluent language use and emergent abilities imply genuine understanding or merely powerful statistical pattern‑matching, and what that means for studying human cognition versus building useful machines. The thread also explores current limitations—logical consistency, agency, self‑awareness, data dependence—and whether continued scaling and new training approaches could eventually yield something close to artificial general intelligence.

Turing Test, AGI, and Scaling

  • Several comments argue the classic Turing test is necessary but not sufficient for “general intelligence”; current LLMs still fail under sustained probing within minutes.
  • Others note modern systems are deliberately constrained not to fully imitate humans, so test results are hard to interpret.
  • Some are impressed that early predictions about required memory and timelines for human‑like dialogue are only off by a few orders of magnitude.
  • Discussion of “ladder to the moon” metaphors: earlier claims that statistical methods could never reach general intelligence now look less convincing; some think simple scaled methods may go surprisingly far.

Chinese Room, Translation, and “Understanding”

  • One side sees the Chinese Room thought experiment as misguided: if a system translates well enough to fool native speakers, that’s operationally equivalent to understanding.
  • Others say LLMs changed their view: they show powerful translation without human‑like semantics.
  • A recurring argument: the “room” (system) can understand even if a component inside does not, similar to neurons vs brains.
  • Some call these debates a distraction from practical impacts; economic usefulness does not depend on a settled theory of understanding.

Universal Grammar and Human Language Learning

  • Debate over whether successful statistical LLMs undermine claims that a built‑in universal grammar is required for language learning.
  • Critics argue: if a machine can learn natural language without such a module, that weakens strong universality claims or at least makes them less exclusive.
  • Defenders respond: current models tell us little about how children learn from sparse data; they can learn “languages” humans cannot, so they are poor evidence about human biology.
  • There is disagreement over whether computational work meaningfully falsifies linguistic theories or is largely irrelevant to them.

Nature of Thought, Understanding, and Error

  • Some posters say LLMs’ fluent reasoning forces reconsideration of what “thinking” is, especially when systems outperform many humans in argument or code.
  • Others insist that frequent logical contradictions and bizarre failures show absence of genuine understanding; they see redefining “understanding” to include LLMs as goalpost‑moving.
  • Counter‑arguments note that humans also contradict themselves and harbor false beliefs; focusing only on failures biases the assessment.
  • There is no consensus on whether “understanding” must be human‑like, or whether it is best treated as a behavioral/functional property.

Capabilities, Limits, and Agency

  • Commenters list areas where LLMs lag average humans: extended, context‑rich tasks; robust agency and planning; stable self‑monitoring of their own limitations; creative writing without clichés; and reliable logical inversion.
  • Others reply that many everyday humans also rely on tropes and make reasoning errors; they see current deficits as engineering gaps, not principled limits.
  • Some view LLMs as powerful but narrow sequence‑modelers that incidentally acquire many cognitive skills; to get more agent‑like behavior, better pretraining objectives and control architectures are needed.
  • There is concern that models can already displace “low‑grade cognitive labor” regardless of whether they teach us anything about human cognition.

Impact on Science, Linguistics, and Society

  • One camp holds that statistical models, like submarines versus fish, may be transformative technologies but reveal little about underlying human mechanisms.
  • Another camp argues that building such systems already constrains theories of language and cognition by showing some assumed prerequisites aren’t strictly necessary.
  • Some worry that advances in LLMs reduce interest in traditional linguistics/cognitive science, because practical AI progress may proceed without deeper theoretical understanding.
  • Others emphasize scalability: even a not‑quite‑human system, if cheaply replicated, could dominate specific problem domains and be socially transformative.