“Gary Marcus was right all along that pure scaling would eventually hit a wall, or at least reach a point of massively diminishing returns.”
Our reading, the sentence that was checked:
Pure scaling of LLMs has hit a wall or at least massively diminishing returns.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Hot take: GPT 4.5 is a nothing burger, paragraph 18.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-hot-take-gpt-45-is-a-nothing-burger-fa4620-a2
We recorded 7 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 2025-02-27. 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
- pure scaling
- adding more data and compute
- was right all along
- hit a wall
- massively diminishing returns
Our interpretation
- In context, this asserts that the predicted limitation was already evident by GPT-4.5's release; it is not scored as an unlimited future forecast.
- The ordinary reading is that increasing pretraining scale had stopped producing meaningful capability gains, or was producing sharply smaller gains relative to the additional resources.
- Assess the frontier state at February 27, 2025, using prior generations as comparison context.
- Separate pretraining scale from architectural changes, post-training, and extra computation while answering.
- Neither persistent errors nor diminishing gain per absolute unit of compute alone demonstrates a newly reached wall.
Counts as supported if: Comparable training-input and capability measurements show a plateau or a pronounced deterioration in returns from increased pretraining scale by the post date.
Counts as contradicted if: Comparable measurements show continuing substantial gains from increased pretraining scale without the claimed pronounced deterioration in returns.
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
OpenAI's February 27, 2025 launch results reported GPT-4.5 gains over GPT-4o, including GPQA scores of 71.4% versus 53.6% and AIME 2024 scores of 36.7% versus 9.3%. [e1] Epoch's September 16, 2025 retrospective analysis placed GPT-4.5's gains broadly on the existing compute-performance trend. [e2] These results weigh against a capability plateau or a pronounced deterioration in returns by February 27, 2025. [e1, e2] However, neither comparison isolates pretraining scale with matched training inputs and controls for other changes, so neither settles the claim about pure scaling. [e1, e2]
Limitations
- The launch comparisons do not isolate the contribution of pretraining scale or provide a controlled measure of returns to additional resources. [e1]
- Epoch's trend comparison does not supply a controlled pretraining-only test across the relevant model generations. [e2]
- The 2020 scaling study measures prediction loss; it does not directly establish practical capability returns at GPT-4.5's scale. [e3]
Evidence (3 items)
Context only: OpenAI,
Even by purely scaling up unsupervised learning, GPT‑4.5 shows meaningful improvements over previous models like GPT‑4o.
Why it matters here: Launch scores improve from 53.6% to 71.4% on GPQA and from 9.3% to 36.7% on AIME 2024. These weigh against a literal capability plateau, but the page also describes architecture and optimization changes and does not quantify a controlled return to additional compute.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: Epoch AI,
This is consistent with scaling roughly as expected.
Why it matters here: Epoch's retrospective comparison places GPT-4.5's GPQA and mathematics gains broadly on the existing compute-performance trend. This is material counterevidence to a newly reached wall, but does not provide the controlled pretraining-only comparison needed to settle Marcus's broader formulation.
- Where
- Page 21, footnote 12
- Publisher
- Epoch AI
- Kind of source
- Independent measurement
- Describes events of
- Public from
Context only: arXiv,
We study empirical scaling laws for language model performance on the cross-entropy loss.
Why it matters here: The original scaling study measures prediction loss and reports power-law relationships with compute, data, and model size. Such relationships do not promise equal practical gains per additional unit of compute or the elimination of every error.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Comparable training compute and data disclosures across model generations available by February 27, 2025, with stable capability evaluations and controls for architecture, post-training, and inference settings, assessed against explicit thresholds for a plateau or pronounced deterioration in returns.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Hot take: GPT 4.5 is a nothing burger
Gary Marcus was right all along that pure scaling would eventually hit a wall, or at least reach a point of massively diminishing returns.
- Where
- paragraph 18
- Published
- 2025-02-27
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-hot-take-gpt-45-is-a-nothing-burger-fa4620-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-hot-take-gpt-45-is-a-nothing-burger-fa4620-a2.