“The real news is a breaking study from Arizona State University that fully vindicates what I have told you for nearly 30 years—and more recently what Apple told you—about the core weakness of LLMs: their inability to generalize broadly.”
Our reading, the sentence that was checked:
LLMs' core weakness is their inability to generalize broadly, as confirmed by an ASU study.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, GPT-5: Overdue, overhyped and underwhelming. And that’s not the worst of it., paragraph 63.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-gpt-5-overdue-overhyped-and-underwhelming-2647a1-a1
We recorded 30 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 2025-08-09. 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
- a breaking study from Arizona State University
- fully vindicates
- the core weakness of LLMs
- their inability to generalize broadly
Our interpretation
- The plain reading is that the cited study establishes unreliable transfer beyond training distributions as a general weakness of contemporary LLMs.
- Broad generalization does not mean perfect performance on every imaginable unfamiliar task.
- I assess the study version available before the post and distinguish demonstrated experimental failures from extrapolation to frontier models.
- The target is the breaking August 2025 study through publication of the post, not later model generations.
Counts as supported if: The study demonstrates substantial failures under meaningful distribution shifts, with methods and model coverage sufficient to justify the broad conclusion about contemporary LLMs.
Counts as contradicted if: The study's results or directly comparable replications establish robust broad transfer under the relevant shifts, or show that the attributed finding is absent or reversed.
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
The ASU study does report substantial failures under distribution shifts, so Marcus accurately identifies its central finding. But its controlled experiments train small models from scratch. They do not directly establish broad incapacity across contemporary frontier LLMs or conclusively identify their core weakness. The narrow experimental claim is supported; the full assertion remains unsettled. Limited coverage alone does not prove the broader assertion false.
Limitations
- This assesses the August 5 version available when Marcus wrote.
- The authors themselves advocate broader implications, but those implications exceed the directly tested model population.
- No independent replication settling the broad extrapolation was obtained.
Evidence (1 item)
Supports the claim: arXiv; Chengshuai Zhao and colleagues,
Our results reveal that CoT reasoning is a brittle mirage that vanishes when it is pushed beyond training distributions.
Why it matters here: This supports the narrower claim that the study reports failures under distribution shifts. Its controlled experiments use small GPT-2-style models trained from scratch; extending those findings to broad incapacity in contemporary frontier LLMs is an inference, not a direct experimental result.
- Publisher
- arXiv; Chengshuai Zhao and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Controlled tests and independent replications across representative contemporary LLMs demonstrating whether the same failures persist across meaningful unfamiliar tasks, lengths and formats.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, GPT-5: Overdue, overhyped and underwhelming. And that’s not the worst of it.
The real news is a breaking study from Arizona State University that fully vindicates what I have told you for nearly 30 years—and more recently what Apple told you—about the core weakness of LLMs: their inability to generalize broadly.
- Where
- paragraph 63
- Published
- 2025-08-09
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-gpt-5-overdue-overhyped-and-underwhelming-2647a1-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-gpt-5-overdue-overhyped-and-underwhelming-2647a1-a1.