Claims › Concept · What a machine cannot be said to do

Form–meaning gap

Can a system trained only on form learn what the language is about?

Training on form alone leaves meaning out of reach: what reads as understanding is a property of the text, not of the system.

Where it sits

1 Free tag Mitchell · Krakauer · Bender · Koller a name finds, it does not vouch
2 Topic Meaning closed list, one per claim
3 Theme What a machine cannot be said to do derived from the topics
4 Concept Form–meaning gap · 2 claims a note and links, or no build
5 Geometry 1.25 1 source, use, dissent

How its claims lean on each other

1 · Free tag2 · Topic3 · Theme4 · Concept5 · Geometryweight, 0.5 → 2The claimsMitchell · KrakauerMitchell · KrakauerBender · KollerBender · KollerMeaningMeaningWhat a machine cannot be said to doWhat a machine cannot be said to doForm–meaning gapForm–meaning gapAlso: Agent modelC007C007 · The dispute over machine understanding is not one question but two pictures of what understanding is: tracking statistical regularities, or holding a causal model of how things work.C008C008 · A system trained only on the shape of language has nothing in its input that could tell it what the language is about.C007 presupposes C008

The claims

  1. C007

    “Moreover, the current debate suggests a fascinating divergence in how to think about understanding in intelligent systems, in particular the contrast between mental models that rely on statistical correlations and those that rely on causal mechanisms.”

    The dispute over machine understanding is not one question but two pictures of what understanding is: tracking statistical regularities, or holding a causal model of how things work. Naming the split explains why the two camps talk past each other.

    Mitchell & Krakauer 2023 · p. 1 · used · weight 1.25 · presupposes C008 · open C007

  2. C008

    “We argue that the language modeling task, because it only uses form as training data, cannot in principle lead to learning of meaning.”

    A system trained only on the shape of language has nothing in its input that could tell it what the language is about. The argument is about what the training signal contains, not about how well the system performs today.

    Bender & Koller 2020 · §1 Introduction · assumed · weight 1 · open C008

All 15 concepts →All the claims, one by one →