“Over the last few years, graphics have gotten better and better. World models and comprehension less so.”
Our reading, the sentence that was checked:
Over the last few years, graphics in image generation systems have gotten better and better, while world models and comprehension have improved less.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, What scaling does and doesn’t buy you: peeling back the hype surrounding Google‘s trendy Nano Banana, paragraph 6.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-what-scaling-does-and-doesnt-buy-0c6ad1-a1
We recorded 8 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-08-27. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- 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
- over the last few years (scope, stated by the speaker)
From the speaker
- Over the last few years
- graphics have gotten better and better
- World models and comprehension less so
Our interpretation
- Read the comparison as concerning image generation systems, rather than all AI systems.
- Graphics means visual fidelity and rendering quality. World models and comprehension mean observable physical coherence, compositional correctness, and understanding of requested content.
- Interpret the last few years approximately as 2022 through publication.
- The claim compares improvements over time. A remaining gap in absolute performance does not by itself establish different rates of improvement.
Counts as supported if: Comparable evaluations across the target years show sustained visual-quality gains and materially smaller gains in physical coherence, compositional correctness, or content understanding.
Counts as contradicted if: Comparable evaluations show that the relevant understanding capabilities improved as much as or more than visual quality, or that the asserted sustained visual-quality improvement did not occur.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
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
R2I-Bench describes strong photorealistic progress alongside underdeveloped reasoning, which is consistent with Marcus's account. However, a remaining weakness is not a measurement of slower improvement. The retrieved studies do not compare visual and understanding gains on a consistent basis across 2022–August 2025. Prompt sensitivity adds uncertainty to comparisons based on selected examples.
Limitations
- The starting year and operational meanings of the capabilities are evaluator interpretations.
- The cited research gives useful context, not a direct test of the claimed relative improvement.
- This assessment does not infer an internal world model from image appearance alone.
Evidence (2 items)
Context only: Kaijie Chen and colleagues / arXiv,
photorealistic
Why it matters here: The authors' assessment is consistent with Marcus's distinction between attractive rendering and remaining reasoning weaknesses. It does not measure whether reasoning improved less over the same years; the paper does not specify exact experiment dates.
- Publisher
- Kaijie Chen and colleagues / arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Haosheng Gan and colleagues / arXiv,
prompt phrasing
Why it matters here: The study reports that prompt optimization improves measured compositional performance. This makes fixed, selected prompts an uncertain basis for comparing underlying capabilities over time, but does not establish the opposite historical trend.
- Publisher
- Haosheng Gan and colleagues / arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A longitudinal comparison of relevant model generations using stable visual-quality and semantic or physical-correctness evaluations, with a defensible method for comparing their gains.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, What scaling does and doesn’t buy you: peeling back the hype surrounding Google‘s trendy Nano Banana
Over the last few years, graphics have gotten better and better. World models and comprehension less so.
- Where
- paragraph 6
- Published
- 2025-08-27
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-what-scaling-does-and-doesnt-buy-0c6ad1-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-what-scaling-does-and-doesnt-buy-0c6ad1-a1.