“Every red cell in this graph (e.g, multiplying 20 digit numbers by 19 digit numbers2) is failure to generalize, even after immense amounts of training data:”
Our reading, the sentence that was checked:
LLMs including o1-mini fail to generalize on large multi-digit multiplication despite immense training data.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Ezra Klein’s new take on AGI – and why I think it’s probably wrong, paragraph 18.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-ezra-kleins-new-take-on-agi-and-why-92e534-a2
We recorded 29 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 2025-03-05. 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
- even after immense amounts of training data (qualifier, stated by the speaker)
From the speaker
- Every red cell in this graph
- multiplying 20 digit numbers by 19 digit numbers
- failure to generalize
- even after immense amounts of training data
- o1-mini
Our interpretation
- The linked graph refers to Yuntian Deng's September 2024 multiplication experiment. Its experimental period, rather than March 2025 or later model releases, is the target.
- The ordinary reading is that the tested general-purpose models, including o1-mini, did not reliably apply multiplication to large new inputs.
- Failure to generalize is assessed behaviorally as unreliable exact multiplication. Wrong answers alone do not establish a particular internal mechanism.
- Immense training data is read qualitatively as broad pretrained-model exposure, not an unstated token-count threshold or proof that these exact products appeared in training.
- This is an observation after training, not a conditional forecast.
Counts as supported if: The referenced experiment reports unreliable large multi-digit multiplication by the named models, including o1-mini, despite their broad pretraining.
Counts as contradicted if: The referenced experiment shows reliable exact multiplication throughout the claimed large-input region, or the graph's red cells do not indicate the asserted performance failures.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Supported
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
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
Deng's September 17, 2024 experiment tested multiplication up to 20-by-20 digits and reported unreliable performance as inputs grew, including for o1-mini; o1-preview had similar accuracy, while GPT-4o struggled beyond 4-by-4 digits [e1]. This supports failure to generalize in the behavioral sense of unreliable multiplication on large inputs under the tested conditions [e1]. OpenAI's later December 5, 2024 disclosure describes diverse pretraining sources for o1-mini, supporting broad training exposure without establishing its quantity [e2]. A May 2024 study reported 99% accuracy on 9-by-9-digit multiplication for a specially trained GPT-2 Small, limiting any inference of universal architectural inability [e3].
Limitations
- The retrieved text does not permit a cell-by-cell audit, including the specific 20-by-19-digit example.
- Exact training volume and exposure to the tested products are undisclosed.
- The result concerns the tested models and configurations; it does not establish the internal cause of errors or performance with calculator assistance.
Evidence (3 items)
Supports the claim: Yuntian Deng, reproduced by Thread Reader,
struggles beyond 4x4.
Why it matters here: The original experiment reports a clear limit to reliable multiplication as input length grows. The thread identifies o1-mini and reports similar accuracy for o1-preview, supporting the claim about the tested models.
- Publisher
- Yuntian Deng, reproduced by Thread Reader
- Kind of source
- Independent measurement
- Describes events of
- Public from
Context only: OpenAI,
web data and open-source datasets.
Why it matters here: The disclosure covers o1 and o1-mini and describes broad training sources. It supports the qualitative training background, but does not disclose an exact training-data quantity or exposure to the tested products.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: Yuntian Deng, Yejin Choi and Stuart Shieber, arXiv,
Stepwise Internalization is effective.
Why it matters here: The authors report that a specially trained GPT-2 Small reaches 99% accuracy on nine-by-nine-digit multiplication. Different training and models make this contextual counterevidence against a universal incapacity claim, rather than a contradiction of o1-mini's measured failures.
- Where
- Section 5, Results
- Publisher
- Yuntian Deng, Yejin Choi and Stuart Shieber, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Ezra Klein’s new take on AGI – and why I think it’s probably wrong
Every red cell in this graph (e.g, multiplying 20 digit numbers by 19 digit numbers2) is failure to generalize, even after immense amounts of training data:
- Where
- paragraph 18
- Published
- 2025-03-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-ezra-kleins-new-take-on-agi-and-why-92e534-a2: Current. Outcome: Supported. 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-ezra-kleins-new-take-on-agi-and-why-92e534-a2.