“The problem of distribution shift, well-tested in Apple’s 2024 reasoning paper, which goes back to my own 1998 work, continues.”
Our reading, the sentence that was checked:
The problem of distribution shift in neural models continues to be unsolved.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Why I don’t share Sam Altman’s confidence that AGI is basically a solved problem, paragraph 8.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-sam-altman-thinks-that-agi-is-basically-467fc6-a1
We recorded 16 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 2025-01-06. 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)
From the speaker
- The problem of distribution shift
- Apple’s 2024 reasoning paper
- continues
Our interpretation
- The ordinary reading is that distribution changes still caused reliability failures in contemporary LLMs. It does not require every model to fail every shift.
- The target is the 2024 research cycle through January 6, 2025, following the explicit reference to Apple's 2024 evaluation.
- The reference to Marcus's 1998 work is historical attribution, not an additional performance threshold.
Counts as supported if: Controlled evaluations of contemporary LLMs show continuing errors or performance deterioration under changes to inputs that preserve the underlying task.
Counts as contradicted if: Reliable evaluations establish that the cited failure pattern was an artifact and that the relevant contemporary models reliably handled the tested distribution changes.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Supported
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
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
Apple's contemporaneous evaluation documents continuing failures under altered inputs, supporting Marcus's ordinary claim that the problem persisted. The paper also reports robust behavior in some settings, and a later statistical reassessment substantially narrows its strongest conclusions. Neither qualification eliminates the evidence of remaining failures.
Limitations
- Support concerns continued empirical failures, not a proof that neural models cannot generalize.
- Not every model showed a meaningful decline on every perturbation.
- Marcus's historical priority claim about 1998 was not independently assessed.
Evidence (2 items)
Supports the claim: arXiv; Apple researchers,
only the numerical values
Why it matters here: The original evaluation reports continuing sensitivity to altered arithmetic questions and irrelevant clauses. Its detailed results also identify negligible basic-variation drops for some models and strong o1-preview performance on difficulty variants, so the evidence supports persistent failures rather than uniform fragility.
- Publisher
- arXiv; Apple researchers
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Długosz, Oliveira, and Díaz-Rodríguez,
shaky statistical ground.
Why it matters here: This later reassessment finds statistically significant changes in only eight of twenty evaluated models under the original prompt, with number magnitude explaining some effects. It weakens blanket interpretations of Apple's results but still identifies model-specific failures; these later experiments are context outside the target period.
- Where
- Abstract, version 3
- Publisher
- arXiv; Długosz, Oliveira, and Díaz-Rodríguez
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Why I don’t share Sam Altman’s confidence that AGI is basically a solved problem
The problem of distribution shift, well-tested in Apple’s 2024 reasoning paper, which goes back to my own 1998 work, continues.
- Where
- paragraph 8
- Published
- 2025-01-06
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-sam-altman-thinks-that-agi-is-basically-467fc6-a1: Current. Outcome: Supported. 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-sam-altman-thinks-that-agi-is-basically-467fc6-a1.