I disagree with Geoff Hinton regarding "glorified autocomplete"

Original Article ↗ Hacker News Discussion ↗

Claims that large language models are “just glorified autocomplete” are weighed against the counterargument that highly accurate next-word prediction may require genuine internal models of meaning or “understanding.” Commenters probe what understanding, intelligence, and world models actually mean, comparing LLMs to human cognition, game-playing AIs, and traditional software, and highlighting current gaps such as long-term memory, planning, goal-directed behavior, and reliable abstraction. While many agree today’s models fall short of human-like reasoning, there is disagreement over whether scaling and better training will eventually bridge that gap or whether fundamentally new architectures are needed.

ChatGPT style and prompting

  • Several commenters dislike ChatGPT’s default verbosity and list-heavy style.
  • Reported mitigations: custom instructions (“be very concise”), explicit tone/length prompts, or follow‑up requests to “rewrite less verbosely.”
  • Some note that base models are less verbose and that RLHF amplifies politeness and length.

“Autocomplete” vs. “Understanding”

  • One side: next‑token prediction can remain purely statistical; accurate prediction does not require “understanding.”
  • Others argue that highly capable prediction effectively requires building internal abstractions and world models, which is a form of understanding.
  • Multiple people note “understanding” is ill‑defined; depending on definition, both humans and LLMs may understand, or neither does.
  • Some say insisting only humans can “really” understand is just anthropocentrism or semantics.

Human cognition parallels

  • Many point out that human brains also do predictive processing in language and perception; much of everyday speech feels like “autocomplete.”
  • Distinction is drawn between fast associative thinking (System 1) and slow deliberative reasoning (System 2); some argue LLMs mainly show System‑1‑like behavior with occasional emergent System‑2‑like reasoning.
  • Others stress humans have goals, long‑term memory, and can plan ahead before speaking, whereas LLMs lack intrinsic goals and persistent memory.

Capabilities, generalization, and limitations

  • Examples where LLMs perform well: new math problems, programming help, logic puzzles (sometimes), learning new syntaxes in conversation.
  • Counterexamples: brittle failures on slight puzzle variations, visual abstraction benchmarks (e.g., ConceptARC), and “boneheaded” math or code mistakes indicating shallow conceptual grasp.
  • Several note lack of insight: humans sometimes reframe or reject the question (e.g., advising a simpler non‑coding solution), whereas LLMs tend to stay within the stated task.
  • Hallucinations and difficulty saying “I don’t know” are highlighted as key limitations.

Architectures, Markov processes, and information theory

  • Debate over whether transformers are “just Markov chains”:
    • One side: finite context and next‑step dependence look Markovian.
    • Other side: internal state and attention over the whole context violate simple Markov assumptions; calling them Markov is seen as oversimplified.
  • Some argue that minimizing cross‑entropy on text plus regularization effectively forces models toward compressed representations of the underlying “information graph” of the world.

World models, semantics, and truth

  • Disagreement over whether LLMs model the world or merely patterns in human text.
  • One view: they “map the maps” (texts about reality), not reality itself, akin to a sophisticated Chinese Room.
  • Others respond that humans also only access the world via limited sensory data and language; both systems construct internal models from indirect evidence.
  • There is discussion of LLMs’ ability to track truth vs. fiction, with references (within the thread) to work on calibration and uncertainty; skeptics reply that models still lack a grounded notion of truth.

AGI, goalposts, and analogies

  • Most agree current LLMs are not AGI; they still fail in “stupid” ways despite strong benchmark scores.
  • Some argue critics keep “moving the goalposts” (e.g., demanding invention of new math) whenever a capability is achieved.
  • AlphaZero/Go is used as an analogy: self‑play with a clear objective shows how powerful prediction and pattern recognition can be, but others caution that games are closed worlds unlike open‑ended language and reality.

Meta: usefulness of the “autocomplete” frame

  • Some see “glorified autocomplete” as dismissive yet technically compatible with strong intelligence: a sufficiently powerful “autocomplete” could embed deep reasoning.
  • Others find the debate over labels (“autocomplete,” “understanding,” “intelligence”) less useful than concrete, empirical evaluation of capabilities and failure modes.