“Furthermore, there is no step in the process to try to derive such structures; so it is just wishful thinking to expect them to be fully sensitive to semantics as it relates to syntactic structure.”
Our reading, the sentence that was checked:
Because the systems under discussion (Imagen, DALL-E 2 and similar architectures) represent phrases as strings and have no step to derive hierarchical syntactic structures, it is just wishful thinking to expect them to be fully sensitive to semantics as it relates to syntactic structure.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Horse rides astronaut, paragraph 100.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-horse-rides-astronaut-f165b1-a2
We recorded 11 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-29. The source may have been edited after 2022-05-28. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Causal claim
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
Conditions
- Imagen, DALL-E 2 and similar architectures under discussion (scope, stated by the speaker)
- phrases represented as strings, with no step to derive syntactic structures (qualifier, stated by the speaker)
- fully sensitive (qualifier, stated by the speaker)
From the speaker
- no step in the process to try to derive such structures
- so
- fully sensitive to semantics
- as it relates to syntactic structure
Our interpretation
- Read 'such structures' as explicit hierarchical syntactic representations, following the preceding reference to syntactic trees.
- Assess the asserted causal connection between lacking an explicit structure-building step and lacking reliable syntax-dependent semantic behavior.
- The named systems are the 2022 Imagen and DALL-E 2 architectures. Later or different models can inform the mechanism but cannot directly settle their performance.
- 'Fully sensitive' means systematic handling of meaning-changing syntactic distinctions, rather than occasional successful examples.
Counts as supported if: Controlled experiments or a sufficiently specific theoretical result establish that the missing structure-building step explains the relevant semantic failures in these systems.
Counts as contradicted if: Matched evidence demonstrates systematic syntax-dependent semantic reliability without such a step, or otherwise rules out its absence as the asserted cause.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
Imagen's May 2022 paper describes pretrained text encoders and diffusion conditioning (e1). Hewitt's April 4, 2019 demonstration shows recoverable syntactic relationships in BERT representations, but does not test Imagen or DALL-E 2 (e2). Neither item establishes or rules out the absence of explicit syntactic structure-building as the cause of semantic failures in the named image-generation systems (e1; e2). The causal claim remains unresolved on this evidence (e1; e2).
Limitations
- The supplied architecture evidence concerns Imagen and does not independently document DALL-E 2's pipeline (e1).
- Recoverable syntax in BERT does not establish reliable use of syntax during image generation (e2).
- The packet contains no controlled intervention separating encoder, generation, training-data and sampling effects.
Evidence (2 items)
Context only: Google Research researchers via arXiv,
Imagen explores pretrained text encoders:
Why it matters here: The documented pipeline uses text embeddings to condition diffusion models and specifies no explicit tree-building stage. That supports the architectural premise under the explicit-parser reading, but does not establish its claimed causal consequence.
- Publisher
- Google Research researchers via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: John Hewitt,
subjects and verbs close
Why it matters here: The author demonstrates recoverable syntactic relationships in learned BERT representations without an explicit parser in the underlying model. This cautions against inferring absence of syntactic information from architecture alone, but concerns a different system and does not prove full semantic reliability.
- Publisher
- John Hewitt
- Kind of source
- Demonstration
- Describes events of
- Public from
What would settle it
Controlled experiments on the 2022 Imagen and DALL-E 2 systems, or applicable theory, establishing or ruling out the proposed cause. Experiments should manipulate syntactic structure-building while controlling other factors and systematically test matched prompts with changed roles or relations.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Horse rides astronaut
Furthermore, there is no step in the process to try to derive such structures; so it is just wishful thinking to expect them to be fully sensitive to semantics as it relates to syntactic structure.
- Where
- paragraph 100
- Published
- 2022-05-28
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-horse-rides-astronaut-f165b1-a2: Current. Outcome: Unresolved. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.