“GPT-4 will solve many of the individual specific items used in prior benchmarks, but still get tripped up, particularly in longer and more complex scenarios.”
Our reading, the sentence that was checked:
GPT-4 will solve many prior benchmark items but still fail, especially on longer, complex scenarios.
Assessment: Supported Settled before its deadline on 2026-10-29 Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, What to Expect When You’re Expecting … GPT-4, prediction 2 of 7.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-what-to-expect-when-youre-expecting-6392ec-a1
We recorded 21 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 2022-12-25. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Forecast
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- Settled before its deadline
- Deadline state on 2026-10-29
- Settled before its deadline (the release’s reference date)
Conditions
- particularly in longer and more complex scenarios (qualifier, stated by the speaker)
From the speaker
- solve many
- individual specific items used in prior benchmarks
- still get tripped up
- particularly in longer and more complex scenarios
Our interpretation
- Prior benchmarks means benchmarks already established before GPT-4's release, consistent with the post's discussion of predecessor evaluations.
- The plain reading predicts substantial benchmark success alongside residual failures that become more pronounced with longer contexts or additional reasoning demands.
- Particularly is an emphasis, not a conditional antecedent that must separately occur.
- The claim does not require performance to decline on every measure whenever a task becomes longer or more complex.
Counts as supported if: GPT-4 solves substantial portions of established benchmarks, while controlled comparisons show remaining failures and poorer performance with fuller contexts or additional reasoning demands.
Counts as contradicted if: GPT-4 fails to solve substantial portions of prior benchmarks, or broad controlled tests show consistently dependable performance as context length and reasoning demands increase.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Did the forecast come true?
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
- Settled before its deadline
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
Both parts of the forecast occurred. GPT-4 solved large shares of established benchmarks. Controlled conversation tests then documented remaining failures, poorer results with fuller context, and difficulty with additional reasoning steps. Prompting improvements and exceptions across question formats qualify the pattern without contradicting Marcus's word particularly.
Limitations
- Benchmark success does not establish whether individual items were learned or memorized.
- The benchmark report evaluates the base model; FANToM evaluates released instruction-tuned versions.
- The length and complexity finding is demonstrated in a specific task family.
Evidence (2 items)
Supports the claim: OpenAI via arXiv,
traditional benchmarks
Why it matters here: The report gives GPT-4 scores of 86.4% on MMLU and 95.3% on HellaSwag, establishing that it solved many established benchmark items. These particular results concern the pretrained base model and do not independently establish deployed-model reliability.
- Publisher
- OpenAI via arXiv
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Supports the claim: arXiv,
significant score gaps
Why it matters here: FANToM tested GPT-4-0314 and GPT-4-0613, finding remaining failures, worse performance with full conversations, and lower scores on binary questions requiring an additional reasoning step. Step-by-step prompting helped, and the complexity pattern was not consistent across every question format.
- Publisher
- 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, What to Expect When You’re Expecting … GPT-4
GPT-4 will solve many of the individual specific items used in prior benchmarks, but still get tripped up, particularly in longer and more complex scenarios.
- Where
- prediction 2 of 7
- Published
- 2022-12-25
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-what-to-expect-when-youre-expecting-6392ec-a1: 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-what-to-expect-when-youre-expecting-6392ec-a1.