“We did not see anything like the kind of improvement that ChatGPT has exhibited over math word problems, over the last two years.”
Our reading, the sentence that was checked:
In the authors' body-parts image experiments, they did not see anything like the kind of improvement that ChatGPT has shown on math word problems over the last two years.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Don’t Ride This Bike! Generative AI’s persistent trouble with compositionality and parts, paragraph 8.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-dont-ride-this-bike-generative-ais-d40112-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 2024-12-08. 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
- in the authors' rerun body-parts experiments (scope, stated by the speaker)
- as far as these experiments go (qualifier, stated by the speaker)
- over the last two years (qualifier, stated by the speaker)
From the speaker
- We did not see anything like the kind of improvement
- ChatGPT
- math word problems
- over the last two years
Our interpretation
- The ordinary reading is that gains on the authors' image-part tasks were substantially smaller than ChatGPT's gains on math word problems.
- The image comparison begins with the October 2022 baseline; the endpoint is the state described by the December 2024 post.
- A math benchmark such as GSM8K is a possible operational measure, but the speaker did not select it.
- Benchmark gains must be distinguished from changes caused by prompting, tool use, or different model versions.
Counts as supported if: Audited image-task results and comparable early and late ChatGPT math evaluations show a clearly larger improvement in math under a stated, defensible comparison measure.
Counts as contradicted if: Comparable measurements show image-task improvement of similar or greater magnitude, or fail to show the asserted math improvement.
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
Primary evidence documents a large GSM-8K gain between GPT-3.5 and GPT-4. That makes the math-progress premise plausible, but does not settle the comparison with these image experiments. The image baseline needs auditing, and the claim supplies neither matched ChatGPT versions nor a common measure of improvement.
Limitations
- Related benchmark gains cannot substitute for the complete comparison.
- A different choice of gain measure could change the assessment.
Evidence (1 item)
Context only: OpenAI,
chain-of-thought prompting
Why it matters here: Table 2 reports GSM-8K scores of 57.1% for GPT-3.5 and 92.0% for GPT-4. This establishes a benchmark gain, but the model evaluation and prompting setup do not provide the full matched ChatGPT-versus-image comparison.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Arithmetic
- 92.0 - 57.1 = 34.9 percentage points. (inputs: GPT-3.5 GSM-8K accuracy: 57.1%; GPT-4 GSM-8K accuracy: 92.0%; One decimal place, matching the reported inputs.)
- Describes events of
- Public from
What would settle it
Independently audited image scores and matched ChatGPT math scores, with model versions, prompting procedures, and a justified common measure of improvement.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Don’t Ride This Bike! Generative AI’s persistent trouble with compositionality and parts
We did not see anything like the kind of improvement that ChatGPT has exhibited over math word problems, over the last two years.
- Where
- paragraph 8
- Published
- 2024-12-08
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-dont-ride-this-bike-generative-ais-d40112-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.
About assessment marcus-dont-ride-this-bike-generative-ais-d40112-a1.