“Diminishing returns doesn’t mean zero improvement, it means that the “laws” of scaling no longer hold, and that the ROI on adding more data and compute are no longer are predictive in the way that they used to be, hence the need for new approaches.”
Our reading, the sentence that was checked:
The diminishing returns he describes in AI do not mean zero improvement; they mean the scaling "laws" no longer hold and the returns on adding more data and compute are no longer predictive in the way they used to be, which is why new approaches are needed.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Casey (Newton) Strikes Out, paragraph 12.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-casey-newton-strikes-out-a6ac38-a1
We recorded 10 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 2024-12-11. 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
- diminishing returns doesn't mean zero improvement (qualifier, stated by the speaker)
- no longer predictive in the way that they used to be (qualifier, stated by the speaker)
From the speaker
- Diminishing returns doesn’t mean zero improvement
- the “laws” of scaling no longer hold
- adding more data and compute
- no longer are predictive in the way that they used to be
- the need for new approaches
Our interpretation
- The ordinary technical reading is that established relationships between training resources and model performance had lost predictive accuracy.
- Interpret ROI as capability gains from training resources, because no company financial return is specified.
- Use the six weeks before publication identified in the immediately preceding sentence as the period of the claimed current breakdown.
- Smaller marginal gains alone do not establish failed predictions. The claim requires evidence that outcomes departed from the relationships' predictions.
Counts as supported if: Dated frontier training results in the target period show material, repeated departures from previously specified scaling predictions, sufficient to establish the claimed loss of predictive usefulness.
Counts as contradicted if: Comparable frontier training results in the target period show that established scaling relationships retained their predictive accuracy despite smaller marginal gains.
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
Reports of disappointing results establish concern, not a measured failure of scaling predictions. Smaller gains can occur without predictions becoming inaccurate. Meta documented successful predictions earlier in 2024, but those experiments cannot settle a claimed breakdown during the six weeks discussed here. The available records do not establish the stronger assertion.
Limitations
- Important frontier training results were private.
- The claim mixes loss prediction, capability gains, and the usefulness of new methods.
- The narrower late-2024 period prevents earlier results from being treated as decisive counterexamples.
Evidence (3 items)
Context only: Kaplan et al., arXiv,
cross-entropy loss.
Why it matters here: The original study models loss through power-law relationships with resources. Consequently, a claim about disappointing benchmark gains is not automatically a test of those relationships.
- Publisher
- Kaplan et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Meta Llama team, arXiv,
outperforms smaller models
Why it matters here: The paper also reports an accurate two-stage ARC Challenge prediction across four orders of magnitude. This weighs against treating diminishing returns as proof of failed prediction, but the experiment predates the claimed recent breakdown.
- Publisher
- Meta Llama team, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Reuters, republished by Yahoo Finance,
limitations
Why it matters here: Reuters reports plateau concerns, disappointing training outcomes, and exploration of inference-time methods. It does not disclose the forecasts, measured losses, or errors needed to establish that scaling relationships stopped predicting results.
- Publisher
- Reuters, republished by Yahoo Finance
- Kind of source
- News report
- Describes events of
- Public from
What would settle it
Public records of the relevant training runs, including previously specified scaling forecasts and corresponding observed results, would establish whether predictive accuracy materially deteriorated.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Casey (Newton) Strikes Out
Diminishing returns doesn’t mean zero improvement, it means that the “laws” of scaling no longer hold, and that the ROI on adding more data and compute are no longer are predictive in the way that they used to be, hence the need for new approaches.
- Where
- paragraph 12
- Published
- 2024-12-11
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-casey-newton-strikes-out-a6ac38-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.