“Deep learning is indeed finally hitting a wall, in the sense of reaching a point of diminishing results.”
Our reading, the sentence that was checked:
Deep learning is hitting a wall in the sense of reaching diminishing returns.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Deep Learning, Deep Scandal, paragraph 1.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-deep-learning-deep-scandal-e25773-a2
We recorded 8 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 2025-04-07. 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
- Deep learning
- finally hitting a wall
- reaching a point of diminishing results
- clear for months
Our interpretation
- The ordinary reading is that increasing training scale was producing smaller capability gains, not that all progress had stopped.
- The surrounding examples concern frontier language models. I assess that domain rather than every application of deep learning.
- I use January 1 through April 7, 2025 for the recent months described, with earlier generations providing the comparison baseline.
- Testing diminishing results requires comparing performance gains with training inputs. Disappointment, release delays, and benchmark scores alone do not establish that relationship.
Counts as supported if: Comparable measurements across frontier model generations show a clear reduction in capability gains relative to added training compute and data during the target period.
Counts as contradicted if: Comparable measurements show sustained or increasing gains relative to added training inputs across the relevant frontier models and tasks.
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
METR's February 27, 2025 evaluation placed GPT-4.5 between GPT-4o and o1 on its autonomy and research-engineering evaluations, but did not measure gains relative to training inputs (e1). Epoch's September 16, 2025 report found GPT-4.5's gains on selected benchmarks consistent with expected compute scaling (e2). That finding does not exclude diminishing marginal returns (e2). Neither item supplies the broad, matched comparison needed to establish or rule out declining gains per added training input during January 1–April 7, 2025 (e1; e2).
Limitations
- METR evaluated an earlier GPT-4.5 checkpoint in a short, preliminary evaluation.
- Epoch's selected benchmark comparisons do not establish a trend across frontier models and tasks.
- The evidence does not cover deep learning beyond frontier language models or settle which definition of diminishing returns applies.
Evidence (2 items)
Context only: METR,
somewhere between GPT-4o and o1.
Why it matters here: METR found limited relative performance on its autonomy and research-engineering evaluations. This does not measure gains per unit of training input.
- Publisher
- METR
- Kind of source
- Independent measurement
- Describes events of
- Public from
Context only: Epoch AI,
scaling roughly as expected.
Why it matters here: The footnote compares GPT-4.5 benchmark improvements with historical gains per order of magnitude of compute. This challenges a new breakdown in scaling, but expected scaling can itself have diminishing marginal returns.
- Publisher
- Epoch AI
- Kind of source
- Independent measurement
- Describes events of
- Public from
What would settle it
A reproducible comparison of frontier model generations before and during January 1–April 7, 2025, using consistent capability evaluations and disclosed or defensibly estimated training inputs. It would need to define returns and account for uncertainty, architecture, data quality, post-training, and inference compute.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Deep Learning, Deep Scandal
Deep learning is indeed finally hitting a wall, in the sense of reaching a point of diminishing results.
- Where
- paragraph 1
- Published
- 2025-04-07
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-deep-learning-deep-scandal-e25773-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-deep-learning-deep-scandal-e25773-a2.