YYoung and Profiting
← All frameworks
CommunicationStephen Wolfram

Semantic Grammar

A construction kit for sentences that actually mean something.

Difficulty
Advanced
Time to result
~ongoing to results
Steps
4
Confidence
75%

Semantic grammar is Wolfram's proposed 'construction kit' for sentences that actually mean something, distinct from ordinary grammar. Basic grammar tells you a noun-verb-noun string is well-formed, but plenty of well-formed sentences ('the electron ate the moon') are meaningless except poetically. The deeper kit uses category constraints — if the sentence is 'blah ate blah', the first blah must be something that eats, like an animal or person. Wolfram argues Aristotle began sketching this with logic two thousand years ago but nobody completed it, and his guess is that LLMs like ChatGPT have implicitly discovered this semantic grammar — which, once understood clearly, could let us build such systems much more simply and directly than today's train-and-bash neural nets.

Origin

Wolfram traces the idea back to Aristotle's invention of logic as an early attempt at a construction kit for meaningful statements, and argues that ChatGPT's success reveals such a semantic grammar exists and can now be studied.

Core principles

  • 01Ordinary grammar gives noun-verb-noun sentences, but many of those are meaningless.
  • 02There is a deeper 'construction kit' governing which sentences actually mean something.
  • 03Categories of things constrain what can sensibly combine (only things that eat can 'eat').
  • 04LLMs appear to have implicitly discovered this semantic grammar, which could let us build them more simply.

How to run it

  1. 1

    Begin with basic grammar

    Note the shape of a typical sentence — noun, verb, noun — as the starting scaffold.

    Watch out Grammatical correctness alone does not guarantee meaning.

  2. 2

    Filter for meaningfulness

    Separate sentences that actually mean something from grammatically valid nonsense like 'the electron ate the moon'.

  3. 3

    Apply category constraints

    Use categories of things to constrain slots — only things that eat (animals, people) can be the subject of 'ate'.

    Pro tip These category rules are the heart of the construction kit.

  4. 4

    Study it explicitly

    Treat the semantic grammar LLMs seem to have discovered as an explicit object, which could enable much simpler, more direct language systems.

In the wild

The electron ate the moon

Wolfram uses this grammatically correct but meaningless sentence to show that grammar isn't enough — a semantic construction kit is needed to distinguish sentences that mean something, and 'blah ate blah' requires the first blah to be a thing that eats.

The example exposes the category constraints that constitute semantic grammar.

Common mistakes

Mistaking grammatical for meaningful

Treating any noun-verb-noun sentence as reasonable ignores that many are meaningless; only a semantic grammar with category constraints separates sense from nonsense.

Is it for you?

Best for

Researchers of language, meaning, and how LLMs represent semantics.

Not ideal for

Everyday communication tasks that don't require a theory of meaning.

From the transcript

you end up with this kind of notion of a kind of semantic grammar of a a a way a construction kit of how you…

Stephen Wolfram · 38:00

the electron ate the moon. Okay, it's grammatically correct, but it's probably doesn't really mean anything except in some kind of poetic sense.

Stephen Wolfram · 37:00

From the episode

Stephen Wolfram: How AI Works and How to Use It to Stay Ahead

Stephen Wolfram