“Silly errors certainly do, and problems of abstraction are widespread in large language models.”
Our reading, the sentence that was checked:
Large language models still make silly errors, and problems of abstraction are widespread in them.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Happy Groundhog Day, The AI Edition, paragraph 5.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-happy-groundhog-day-the-ai-edition-d7f85a-a2
We recorded 12 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 2023-02-03. 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
- Silly errors certainly do
- problems of abstraction
- widespread
- large language models
Our interpretation
- Silly errors means objectively incorrect answers on elementary tasks, such as arithmetic, rather than stylistic disagreements.
- Abstraction means applying a rule or relation beyond familiar examples or content.
- Widespread means recurring across multiple contemporary models and tasks, not that every model fails every test.
- Use evaluations from the latest completed research year, 2022, and the pre-post portion of 2023 to assess the ongoing state.
Counts as supported if: Contemporary controlled evaluations document elementary errors and recurring failures to apply rules across multiple models, model sizes or task settings.
Counts as contradicted if: Representative contemporary evaluations establish reliable elementary accuracy and robust rule transfer, with the alleged abstraction problems confined to isolated or invalid examples.
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 May 2022 evaluation of GPT-J and two GPT-Neo sizes found arithmetic errors and performance linked to training-term frequency [e1]. The July 2022 study of 7-billion- and 70-billion-parameter models found weaker performance on unrealistic or abstract content across three logical reasoning tasks [e2]. These results support recurring elementary errors and difficulty applying rules across different content in the tested models and settings [e1, e2]. The latter study also reported above-chance performance and error patterns similar to humans; it does not establish a complete absence of abstraction [e2].
Limitations
- The studies do not provide a population-wide percentage of LLMs with abstraction problems [e1, e2].
- Training-frequency correlations do not by themselves establish the mechanism behind the errors [e1].
- The reasoning study does not establish that these difficulties are unique to language models [e2].
Evidence (2 items)
Supports the claim: arXiv; Yasaman Razeghi and colleagues,
strong effect
Why it matters here: Experiments on GPT-J and two GPT-Neo sizes find errors in arithmetic and strong dependence on training-term frequency. That supports elementary errors and fragile transfer, although frequency dependence alone does not prove an absence of reasoning.
- Publisher
- arXiv; Yasaman Razeghi and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv; Ishita Dasgupta and colleagues,
many imperfections.
Why it matters here: The study tests 7-billion- and 70-billion-parameter models on three reasoning tasks and finds sensitivity to whether content is familiar or abstract. It also finds above-chance ability and similarities to human errors, which qualify any inference that the models cannot abstract at all.
- Publisher
- arXiv; Ishita Dasgupta and colleagues
- 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, Happy Groundhog Day, The AI Edition
Silly errors certainly do, and problems of abstraction are widespread in large language models.
- Where
- paragraph 5
- Published
- 2023-02-03
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-happy-groundhog-day-the-ai-edition-d7f85a-a2: 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-happy-groundhog-day-the-ai-edition-d7f85a-a2.