“Just as I argued here in April 2024, LLMs have reached a point of diminishing returns.”
Our reading, the sentence that was checked:
LLMs have reached a point of diminishing returns from scaling.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, CONFIRMED: LLMs have indeed reached a point of diminishing returns, paragraph 6.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-confirmed-llms-have-indeed-reached-65340b-a2
We recorded 13 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2024-11-09. 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
- LLMs have reached a point of diminishing returns
- Just as I argued here in April 2024
- adding more data and more compute
- without making fundamental architectural changes
Our interpretation
- Read diminishing returns as smaller useful capability improvements from comparable proportional increases in pretraining scale.
- The April reference identifies an earlier assertion; the present-tense claim describes the state by November 9.
- Use calendar 2024 through publication as the latest development period indicated by the context, with older generations as baselines.
- Distinguish a practical slowdown in capabilities from the mathematical curvature of prediction-loss scaling.
Counts as supported if: Consistent evaluations and resource measurements establish that successive frontier scaling increases yielded smaller useful capability gains during the target period.
Counts as contradicted if: Comparable evaluations establish sustained or increasing useful capability gains for comparable scaling increases across the relevant frontier generations.
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 October 1, 2024 research describes predictable capability scaling and differences in training efficiency across model families, but does not establish declining frontier capability gains per proportional resource increase [e1]. Reuters reported pretraining plateau concerns on November 12, 2024, after the November 9 cutoff; the reported experiments lack resource-normalized measurements and precise dates [e2]. Neither item establishes whether useful returns from scaling diminished during January 1–November 9, 2024 [e1, e2].
Limitations
- The evidence lacks a comparable series of frontier training runs measuring capability gains against proportional increases in compute and data [e1, e2].
- The reported private experiments do not separate scaling effects from other training changes [e2].
- Predictable capability improvement alone does not establish whether marginal returns increased or decreased [e1].
Evidence (2 items)
Context only: Ruan, Maddison and Hashimoto, arXiv,
training compute efficiencies
Why it matters here: The study models predictable capability changes across many models while accounting for differences between families. Its fitted relationships do not establish a broad late-2024 decline in returns at the frontier.
- Where
- Abstract
- Publisher
- Ruan, Maddison and Hashimoto, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Reuters, syndicated by AsiaOne,
Scaling
Why it matters here: The report documents contemporary concerns consistent with diminishing returns. The private results lack resource-normalized measurements and sufficiently precise event dates for the target period.
- Publisher
- Reuters, syndicated by AsiaOne
- Kind of source
- News report
- Describes events of
- Public from
What would settle it
Comparable measurements of successive frontier training runs during January 1–November 9, 2024, with older generations as baselines, showing useful capability gains against proportional increases in compute and data while accounting for benchmark saturation, data quality, architecture and post-training.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, CONFIRMED: LLMs have indeed reached a point of diminishing returns
Just as I argued here in April 2024, LLMs have reached a point of diminishing returns.
- Where
- paragraph 6
- Published
- 2024-11-09
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-confirmed-llms-have-indeed-reached-65340b-a2: 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-confirmed-llms-have-indeed-reached-65340b-a2.