“In short, diminishing returns are what we can expect.”
Our reading, the sentence that was checked:
Scaling data for current AI models will yield diminishing returns.
Assessment: Unresolved No deadline stated on 2026-10-29 Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Breaking news: Scaling will never get us to AGI, paragraph 4.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-breaking-news-scaling-will-never-ba4979-a1
We recorded 6 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 2024-04-08. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Forecast
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No deadline stated
- Deadline state on 2026-10-29
- No deadline stated (the release’s reference date)
From the speaker
- diminishing returns
- what we can expect
Our interpretation
- The surrounding discussion makes this a forecast about increasing training data for then-current model approaches.
- I read diminishing returns as smaller performance gains per additional unit of comparable training data.
- This does not necessarily mean that performance stops improving or that scaling becomes useless.
- The post supplies no deadline, performance measure, data range, or minimum decline.
Counts as supported if: Representative controlled scaling studies establish declining marginal performance gains as comparable training data increases across the forecast's intended model scope.
Counts as contradicted if: Representative controlled scaling studies establish sustained non-diminishing marginal gains across that scope.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Did the forecast come true?
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 deadline stated
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
Published experiments show diminishing returns in particular regimes, including a later controlled dataset-size study. Other research shows that training choices can postpone saturation. These findings are relevant but do not conclusively settle the broad, undated forecast. No decisive event meeting the early-resolution rule was found by October 28, 2026.
Limitations
- A narrower claim about a specified loss curve could be supported by individual experiments.
- Diminishing marginal gains are compatible with continued substantial progress.
Evidence (3 items)
Context only: arXiv; Muennighoff and colleagues,
with more repetition
Why it matters here: The experiment demonstrates saturation when data is repeatedly reused. It predates the forecast and concerns repeated tokens and compute rather than unrestricted additions of new data.
- Where
- Abstract
- Publisher
- arXiv; Muennighoff and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Wiegand and colleagues,
clear diminishing returns
Why it matters here: This later experiment directly reports diminishing returns from increased dataset size. Its deliberately tiny attention-only decoder does not establish the broad trend for the forecast's intended model population.
- Where
- Abstract
- Publisher
- arXiv; Wiegand and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Wu and colleagues,
sparse training postpones diminishing returns
Why it matters here: The study finds that sparsity can delay saturation in data-constrained training. This qualifies extrapolation from a fixed training setup, although it does not show that diminishing returns disappear.
- Where
- Abstract, finding 2
- Publisher
- arXiv; Wu and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A replicated scaling program covering representative model approaches and larger data ranges, with a consistent performance measure and controlled data quality, model size and compute, could establish the trend under an explicit operational reading.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Breaking news: Scaling will never get us to AGI
In short, diminishing returns are what we can expect.
- Where
- paragraph 4
- Published
- 2024-04-08
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-breaking-news-scaling-will-never-ba4979-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-breaking-news-scaling-will-never-ba4979-a1.