““Scaling” alone—building bigger and models until they absorb the entire internet — will prove useful, but only to a point.”
Our reading, the sentence that was checked:
Scaling alone will prove useful only up to a point.
Assessment: Unresolved No deadline stated on 2026-10-29 Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, What to Expect When You’re Expecting … GPT-4, prediction 7 of 7.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-what-to-expect-when-youre-expecting-ba0b73-a2
We recorded 26 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 2022-12-25. 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)
Conditions
- scaling alone (qualifier, stated by the speaker)
From the speaker
- Scaling alone
- building bigger
- absorb the entire internet
- will prove useful
- but only to a point
Our interpretation
- Scaling means increasing language-model size and pretraining data or compute, without adding qualitatively different reasoning tools or system components.
- In the surrounding discussion, only to a point means that scaling improves capabilities but is insufficient by itself for trustworthy general intelligence.
- A temporary benchmark plateau, diminishing returns, or current model failures do not alone establish that lasting limit.
- Scaling alone specifies the development regime to assess. Controlled scaling experiments establish that this regime has been attempted.
- The nearby decade and 2043 predictions concern other propositions and are not imported as this claim's deadline.
Counts as supported if: Evidence establishes useful gains from scaling and a persistent limitation preventing scaling alone from reaching trustworthy general intelligence, rather than merely showing finite experimental saturation.
Counts as contradicted if: A transparently documented system developed through scaling alone demonstrably achieves trustworthy general intelligence under broad, credible independent evaluation.
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
The controlled scaling study available in June 2024 documented predictable performance gains across model sizes and training amounts [e1]. The May 2023 data-constrained study found diminishing returns from repeatedly reusing limited data, establishing a limit within that experimental setting [e2]. Neither study establishes a fundamental ceiling below trustworthy general intelligence [e1, e2]. The June 2026 version of the capability study examined task-dependent saturation using models with contemporary post-training, so it does not isolate scaling alone [e3]. These studies establish neither the required lasting ceiling nor a scaling-only achievement of trustworthy general intelligence, leaving the forecast unresolved [e1, e2, e3].
Limitations
- The claim specifies no deadline, capability threshold, or operational test of trustworthy general intelligence.
- Finite experiments and task-specific saturation do not establish a universal ceiling.
- The boundary between scaling and other training improvements remains unspecified.
- The absence of a qualifying demonstration does not establish that scaling alone must fail.
Evidence (3 items)
Supports the claim: Gadre and colleagues, via arXiv,
scaling laws
Why it matters here: The researchers report a controlled testbed varying model size and training tokens, establishing that the scaling regime was attempted. Predictable gains within that testbed also caution against treating scaling as already exhausted.
- Where
- Abstract
- Publisher
- Gadre and colleagues, via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Muennighoff and colleagues, via arXiv,
eventually decays to zero
Why it matters here: The study finds saturation when compute is added while repeatedly reusing constrained data. That is evidence of a specific scaling limit, not a demonstration that scaling model size and additional data cannot reach trustworthy general intelligence.
- Where
- Abstract
- Publisher
- Muennighoff and colleagues, via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Zhang and colleagues, via arXiv,
task-dependent saturation
Why it matters here: This recent study examines capability boundaries and saturation across benchmarks. Its observational comparisons include contemporary post-training, so they cannot determine the ultimate effect of scaling alone.
- Where
- Abstract
- Publisher
- Zhang and colleagues, via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A validated result establishing a fundamental ceiling that prevents scaling alone from achieving trustworthy general intelligence, or a transparently documented scaling-only system independently shown to achieve it. Either requires an operational definition and evidence separating scale from other development changes.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, What to Expect When You’re Expecting … GPT-4
“Scaling” alone—building bigger and models until they absorb the entire internet — will prove useful, but only to a point.
- Where
- prediction 7 of 7
- Published
- 2022-12-25
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-what-to-expect-when-youre-expecting-ba0b73-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-what-to-expect-when-youre-expecting-ba0b73-a2.