“The core claims that Ernie and I made in Technology Review still all, so far as I can tell, hold true, nearly two years and many even larger models later. Yes, some of the specific examples we pointed out have been remedied, but the general problems have not.”
Our reading, the sentence that was checked:
So far as Marcus can tell, all the core claims he and Davis made in Technology Review still hold true nearly two years and many even larger models later; some of their specific examples have been remedied, but the general problems have not.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, What does it mean when an AI fails? A Reply to SlateStarCodex’s riff on Gary Marcus, paragraph 29.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-what-does-it-mean-when-an-ai-fails-d02c78-a1
We recorded 14 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 2022-06-07. 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)
Conditions
- so far as I can tell (qualifier, stated by the speaker)
- nearly two years and many even larger models later (scope, stated by the speaker)
From the speaker
- core claims
- still all
- so far as I can tell
- nearly two years
- some of the specific examples ... have been remedied
- the general problems have not
Our interpretation
- The context identifies six general problems: biological, physical, social and practical reasoning, object and individual tracking, and non sequiturs.
- The ordinary reading is that these remained substantive weaknesses, despite improvements on particular examples. It does not require the models to fail every test.
- Because this describes the current model generation without a reporting period, I use 2022 through the post date as the latest public evaluation period.
- The qualification about Marcus's knowledge does not remove the factual assertion. However, evidence about only some categories cannot establish his assertion about all six.
Counts as supported if: Public evaluations of the relevant model generation establish continuing substantive failures across all six categories, while allowing improvements on individual examples.
Counts as contradicted if: Robust evaluations establish that at least one of the six general problems had been resolved across varied tests by the post date. Better average scores or corrected examples alone do not suffice.
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 records support continuing mistakes and also show substantial progress. They do not establish Marcus's stronger assertion that all six general problems remained unresolved. Nor do the improved benchmark results establish that any whole category had been solved. The breadth of the claim remains unverified.
Limitations
- This is a gap in the retrieved evidence, not evidence that the categories had been solved.
- Benchmark averages do not directly measure the six categories as Marcus defined them.
- The dated disclosures identify public capability snapshots, not the exact dates each experiment ran.
Evidence (2 items)
Supports the claim: OpenAI researchers / arXiv,
InstructGPT still makes simple mistakes.
Why it matters here: The reported model still accepted false premises, invented facts and struggled with multiple constraints. This supports persistent reliability problems, but does not establish failures across all six categories.
- Publisher
- OpenAI researchers / arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Google Research,
average performance
Why it matters here: Google reported strong average performance on a subset of 58 BIG-bench tasks. This is material evidence of progress, but the aggregate does not establish that any entire Marcus category had been resolved.
- Publisher
- Google Research
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
What would settle it
Independent tests of the models available by June 7, 2022, covering all six categories with varied examples, or robust evidence that one complete category had been resolved.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, What does it mean when an AI fails? A Reply to SlateStarCodex’s riff on Gary Marcus
The core claims that Ernie and I made in Technology Review still all, so far as I can tell, hold true, nearly two years and many even larger models later. Yes, some of the specific examples we pointed out have been remedied, but the general problems have not.
- Where
- paragraph 29
- Published
- 2022-06-07
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-what-does-it-mean-when-an-ai-fails-d02c78-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-what-does-it-mean-when-an-ai-fails-d02c78-a1.