“Scaling here was largely AWOL; even the best model was only at 80.6%, and for most models, scaling had at best a neglible effect.”
Our reading, the sentence that was checked:
On the Ruis et al. pragmatics task, the best model reached only 80.6%, and for most models scaling had at best a negligible effect.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, What “game over” for the latest paradigm in AI might look like, paragraph 15.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-what-game-over-for-the-latest-paradigm-30ad12-a1
We recorded 13 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 2022-10-29. 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
- on the Ruis et al. implicature task (qualifier, stated by the speaker)
From the speaker
- Scaling here was largely AWOL
- even the best model was only at 80.6%
- for most models
- scaling had at best a neglible effect
Our interpretation
- The target is the 2022 Ruis et al. study as available when the post appeared, not subsequent models or paper revisions.
- The ordinary reading of 80.6% is the paper's headline accuracy averaged across prompt templates, including its few-shot evaluations.
- Scaling means increasing model size within the tested model families. 'Most models' is read as most families, since scaling requires comparisons between sizes.
- Negligible means little or no consistent improvement across the sizes tested, judged against the paper's reported comparisons. This is not a claim that every family had a flat curve or that a permanent ceiling was proved.
- The target period is the study's 2022 experimental results. The post does not specify individual experiment dates.
Counts as supported if: The original manuscript reports a best overall mean accuracy of 80.6%, and its size comparisons show little or no consistent improvement for most tested model families.
Counts as contradicted if: The original manuscript reports a higher best overall mean accuracy under its main evaluation protocol, or substantial consistent gains with size for most tested model families.
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 original manuscript supports the claim as a description of its headline results. Its best accuracy averaged across templates was 80.6%, and most tested model families showed little or no consistent benefit from greater size. The wording needs qualification: some prompt-specific scores exceeded 80.6%, and several families benefited from scaling. These results do not establish an enduring 80% ceiling.
Limitations
- Support depends on reading 80.6% as the overall mean, not the maximum across every prompt configuration.
- The study tested a finite collection of model families and sizes.
- Observed weak scaling in this experiment does not establish that future scaling cannot improve performance.
Evidence (1 item)
Supports the claim: arXiv; Laura Ruis and colleagues,
80.6% ± 1.22
Why it matters here: The best overall mean is 80.6% with 30 examples; Section 4.1 reports near-zero or decreasing slopes for many families, with only Cohere, T0 and InstructGPT-3 performing best at their largest tested size. The same source qualifies the headline: structured prompts average 81.7%, and Table 58 includes an individual-template score of 82.8%.
- Publisher
- arXiv; Laura Ruis 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, What “game over” for the latest paradigm in AI might look like
Scaling here was largely AWOL; even the best model was only at 80.6%, and for most models, scaling had at best a neglible effect.
- Where
- paragraph 15
- Published
- 2022-10-29
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-what-game-over-for-the-latest-paradigm-30ad12-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-what-game-over-for-the-latest-paradigm-30ad12-a1.