“The “states” that such systems visit are pretty, but lack reference, and lack ties to dynamically updated cognitive models.”
Our reading:
The internal states visited by current neural network systems that cluster similar sentences, as in Piantadosi's argument, are pretty but lack reference and lack ties to dynamically updated cognitive models.
Assessment: Unresolved
What happened
Was it true for the period described?
Outcome: Unresolved
We checked, and the evidence so far doesn’t settle it.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
The reviewed sources do not settle the claim on its external-world reading. Bender and Koller provide a related theoretical argument. Mandelkern and Linzen offer a counterargument about reference but explicitly decline to claim a decisive resolution. Neither establishes the relevant properties of the systems current on 2022-06-22. Semantic similarity or discourse tracking would not be enough to contradict Marcus’s wording.
Caveats
- The available evidence items are theoretical context, not direct outcome measurements.
- Access failures limited inspection of potentially relevant empirical grounding research.
- Later systems cannot settle this historical observation.
Evidence (2 items)
Context only: Association for Computational Linguistics,
a system trained only on form has a priori no way to learn meaning.
Why it matters: Bender and Koller offer a theoretical argument compatible with Marcus’s concern. It does not establish both asserted absences in the systems current on 2022-06-22.
- Publisher
- Association for Computational Linguistics
- Kind of source
- Peer-reviewed paper
- About events on
- Published
Context only: arXiv,
we make no claim to a decisive argument
Why it matters: Mandelkern and Linzen challenge the inference that text-only inputs preclude external-world reference, using the causal histories of linguistic expressions. They explicitly stop short of a decisive argument, and this later theoretical discussion is not a measurement of the target systems.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Conditions
- current neural network systems that cluster similar sentences (the systems in Piantadosi's claim) (their words)
Where we had to interpret
- Reference means reference to the external world, as the surrounding text explicitly states. Tracking an entity mentioned in a passage is not sufficient by itself.
- The ordinary reading of dynamically updated cognitive models is internal representations of the world that can change as relevant information changes. No particular architecture or benchmark is imposed.
- Such systems refers to the contemporary neural language systems discussed in the footnote, including GPT-3. It is not expanded to every neural system or to later models.
- Because this describes an ongoing technical state without a reporting cycle, the target is a snapshot of systems current on the post date, 2022-06-22.
- Pretty is treated as rhetorical description, not a separately measurable proposition.
Right if: Evidence about the intended contemporary systems establishes both that their states lack external-world reference and that they lack ties to dynamically updated representations of that world. Failure on selected prompts or absence of a named architectural module alone is insufficient.
Wrong if: Evidence establishes external-world reference or ties to dynamically updated cognitive models in an in-scope system as it existed at the post date. Discourse tracking, semantic similarity, or task-specific state changes alone are insufficient.
We wrote this test after the outcome was known.
What would settle it
Accessible, reproducible analyses of the intended historical model versions that test external-world reference and links to updateable world representations, with criteria justified from the claim’s meaning rather than substituted discourse benchmarks.
We found 33 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, Three ideas from linguistics that everyone in AI should know, paragraph 63.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-29. The post may have been edited after 2022-06-22.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, Three ideas from linguistics that everyone in AI should know
The “states” that such systems visit are pretty, but lack reference, and lack ties to dynamically updated cognitive models.
- Where
- paragraph 63
- Published
- 2022-06-22
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Unresolved. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
A reviewer said our test wasn’t fair to what was actually said.