“The problems in physical and psychological reasoning that we pointed in the Fall of 2020 very much remain.”
Our reading, the sentence that was checked:
The physical and psychological reasoning problems in GPT-3 that the authors identified in fall 2020 still remain as of February 2023.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, How Not to Test GPT-3, paragraph 42.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-how-not-to-test-gpt-3-419899-a2
We recorded 12 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 2023-02-17. 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
- physical and psychological reasoning
- that we pointed in the Fall of 2020
- very much remain
Our interpretation
- The claim concerns persistence of the types of physical and psychological reasoning failures identified in the authors' 2020 work, including close variants.
- It does not assert that every original example remains incorrect or that no improvement occurred.
- Both named domains must be assessed; unrelated reasoning failures alone do not settle the claim.
- I use early 2023 through the post date as the latest reportable period for the assertion that the problems remain.
Counts as supported if: Independent tests in the target period document substantial recurring failures in both domains on the original problems or demonstrably close variants.
Counts as contradicted if: Independent, sufficiently broad matched testing establishes reliable resolution of the identified failure types in both domains.
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
The physical-task example published on December 19, 2022 records an unsuccessful GPT-3 solution, but predates the January 1–February 17, 2023 target period and does not map to the specified 2020 failures [e1]. The February 4, 2023 report found 93% success for davinci-003 on selected false-belief tasks, but does not establish resolution of those earlier failure types [e2]. These items do not settle whether the identified problems persisted in both domains.
Limitations
- Neither item supplies matched testing of both domains during the target period.
- The evidence does not measure how extensively the specified failure types persisted.
- Success on selected false-belief tasks does not establish reliable psychological reasoning more broadly.
Evidence (2 items)
Context only: arXiv; Webb, Holyoak and Lu,
gumballs
Why it matters here: This is a recorded model response, not the researchers' proposed solution. The paper describes the proposed transfer mechanisms as physically unsound, but this earlier, different task does not establish persistence of the specified 2020 failures during the target period.
- Publisher
- arXiv; Webb, Holyoak and Lu
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Michal Kosinski,
solved 93%
Why it matters here: This original report found substantial success on selected false-belief tasks. Those tasks do not establish resolution of the authors' particular failures, but the result cautions against reading the claim as an absence of psychological reasoning successes.
- Publisher
- arXiv; Michal Kosinski
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
Independent testing from January 1–February 17, 2023 covering the original physical and psychological problems or clearly mapped variants, with model versions, prompts and repeated responses showing substantial recurring failures or reliable resolution in both domains.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, How Not to Test GPT-3
The problems in physical and psychological reasoning that we pointed in the Fall of 2020 very much remain.
- Where
- paragraph 42
- Published
- 2023-02-17
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-how-not-to-test-gpt-3-419899-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.