“About 18 months have passed; although there are certainly improvements (the images are vastly better), owing presumably in large part to a larger training set, the fundamental approach (mapping words to images, rather than constructing intermediate models of the world and how language and images relate to those models) remains the same, and so the core issue remains.”
Our reading, the sentence that was checked:
About 18 months after DALL-E 2, the images are vastly better, presumably largely because of a larger training set, but the fundamental approach of mapping words to images rather than building intermediate world models is the same, and so the core compositionality issue remains.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Race, statistics, and the persistent cognitive limitations of DALL-E, paragraph 4.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-race-statistics-and-the-persistent-4e82f9-a1
We recorded 7 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 2023-10-14. 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
- despite vastly better images (qualifier, stated by the speaker)
- improvements presumably owing in large part to a larger training set (qualifier, stated by the speaker)
- about 18 months after DALL-E 2's release (scope, stated by the speaker)
From the speaker
- About 18 months have passed
- the images are vastly better
- owing presumably in large part to a larger training set
- the fundamental approach ... remains the same
- and so the core issue remains
Our interpretation
- Assess DALL-E 3 as discussed in October 2023, compared with DALL-E 2.
- The plain reading is that substantial visual improvements coexist with persistent difficulty composing objects, attributes and relationships, and that the unchanged general approach explains this difficulty.
- The claim does not require an unchanged failure rate. The screening suggestion of a rate similar to DALL-E 2 adds a commitment Marcus did not make.
- Treat the larger-training-set explanation as a tentative hypothesis because Marcus says 'presumably'.
- Interpret intermediate world models as explicitly constructed representations of entities and relationships. Do not assume that learned internal representations cannot encode such information.
- Use October 1–14, 2023 for the contemporaneous state described; the month boundary is an evaluator choice, not a stated experimental period.
Counts as supported if: Contemporaneous evaluations establish persistent compositional failures, while architecture disclosures and controlled comparisons substantiate the proposed explanation involving the absence of intermediate world models. Evidence should also substantiate the acknowledged visual improvements.
Counts as contradicted if: Representative controlled evaluations establish reliable noncanonical composition, or causal comparisons establish that the proposed architectural explanation is wrong. Improved average performance alone does not contradict persistence of the problem.
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
The sources corroborate improvements and continuing spatial-composition problems. They do not establish Marcus's explanation that the problems persist because the system lacks intermediate world models. OpenAI also identifies better captions as a source of improvement, qualifying the tentative dataset-size explanation. The causal claim remains unresolved; the supported descriptive components do not settle it.
Limitations
- The technical paper was read through a third-party reproduction because the original PDF exceeded the reader's size limit.
- Company evaluations are informative but do not independently establish the claimed mechanism.
- The paper's precise experimental dates were not established.
Evidence (3 items)
Supports the claim: OpenAI,
improving caption fidelity and image quality.
Why it matters here: Supports the improvement component of the claim. It does not quantify 'vastly' or establish the causal explanation for remaining failures.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: OpenAI; reproduced by StudyLib,
object placement and spatial awareness.
Why it matters here: Corroborates persistent relational failures in the release-era model. The report does not isolate the absence of world models as their cause, and its evaluation dates are not precise enough to place inside the target window.
- Publisher
- OpenAI; reproduced by StudyLib
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: OpenAI,
these improved captions
Why it matters here: Provides a concrete alternative contribution to improvement: caption quality. It qualifies the tentative emphasis on dataset size without establishing that size made no contribution.
- Where
- Research section
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
What would settle it
Controlled comparisons of the relevant model versions that separate architecture, training-set size and caption quality, alongside reproducible canonical-versus-noncanonical generation tests.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Race, statistics, and the persistent cognitive limitations of DALL-E
About 18 months have passed; although there are certainly improvements (the images are vastly better), owing presumably in large part to a larger training set, the fundamental approach (mapping words to images, rather than constructing intermediate models of the world and how language and images relate to those models) remains the same, and so the core issue remains.
- Where
- paragraph 4
- Published
- 2023-10-14
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-race-statistics-and-the-persistent-4e82f9-a1: 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.
About assessment marcus-race-statistics-and-the-persistent-4e82f9-a1.