“Here Roose has displayed the cardinal sin of looking only at regurgitation rather than generalization. History has told us over and over that systems can be good at the former and still face enormous challenges at the latter.”
Our reading, the sentence that was checked:
AI systems can be good at regurgitating known code while still facing enormous challenges generalizing to new problems.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Decoding (and debunking) Hard Fork’s Kevin Roose, paragraph 15.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-decoding-and-debunking-hard-forks-e70745-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 2025-03-02. 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
- History has told us over and over
- systems can be good at the former
- enormous challenges at the latter
Our interpretation
- The listed proposition concerns whether strong performance on familiar code can coexist with substantial generalization failures.
- Can is an existence claim, not a claim that every AI system only memorizes.
- Enormous challenges means substantial failures on unfamiliar problems or modest task-preserving variations, rather than a specified numerical threshold.
- History supplies a retrospective period ending at the post, with no stated beginning.
- The separate accusation that Roose examined only regurgitation is not established by general benchmark results.
Counts as supported if: Published evaluations from the historical period demonstrate materially stronger performance on familiar or training-overlapping code alongside substantial failures on less familiar problems or controlled variations.
Counts as contradicted if: The relevant evaluations fail to show this coexistence, or their apparent gaps are convincingly explained by invalid comparisons or changed task requirements.
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
The listed general proposition is supported. A training-overlap study found a large familiarity-related correctness gap, while ReCode documented failures under modest, task-preserving variations. These establish that success on familiar code can coexist with substantial generalization limits. They do not establish that Roose's particular output was regurgitated.
Limitations
- Training similarity and robustness failures do not by themselves prove verbatim memorization.
- ReCode's worst-case perturbation scores are not ordinary-use failure rates.
- The conclusion applies to the normalized general proposition, not the passage's specific accusation about Roose.
- The studies do not show that models cannot generalize.
Evidence (2 items)
Supports the claim: Martin Riddell, Ansong Ni and Arman Cohan, arXiv,
Models perform significantly better when similar solutions are seen during training.
Why it matters here: StarCoderBase-15.5B achieved 72% accuracy on the most training-similar tenth of MBPP questions and 22% on the least similar tenth. The authors also investigated whether difficulty explained the association; these are familiarity comparisons, not proof of verbatim copying.
- Publisher
- Martin Riddell, Ansong Ni and Arman Cohan, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: Association for Computational Linguistics,
fails when function name is perturbed
Why it matters here: ReCode demonstrates failures after task-preserving prompt changes across several code-model families. Its worst-case perturbation results corroborate substantial robustness limits, although they do not independently identify memorization as the cause.
- Publisher
- Association for Computational Linguistics
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Decoding (and debunking) Hard Fork’s Kevin Roose
Here Roose has displayed the cardinal sin of looking only at regurgitation rather than generalization. History has told us over and over that systems can be good at the former and still face enormous challenges at the latter.
- Where
- paragraph 15
- Published
- 2025-03-02
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-decoding-and-debunking-hard-forks-e70745-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-decoding-and-debunking-hard-forks-e70745-a1.