“And very little of what GPT-4 does is fully replicable. Sometimes summarization (aka compression) works (super cool!) and sometimes it doesn’t, as many were quick to point out:”
Our reading, the sentence that was checked:
Very little of what GPT-4 does is fully replicable; its summarization sometimes works and sometimes fails.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, GPT-5 and irrational exuberance, paragraph 16.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-gpt-5-and-irrational-exuberance-da17ab-a1
We recorded 17 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 2023-04-08. 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)
From the speaker
- very little of what GPT-4 does
- fully replicable
- Sometimes summarization (aka compression) works
- sometimes it doesn’t
Our interpretation
- Replicability means reliably reproducing successful task performance under comparable conditions, rather than producing identical wording.
- The claim combines a broad assertion about GPT-4's capabilities with a narrower assertion about summarization failures.
- I use March 2023 through the post date as the latest reporting window for the initial GPT-4 release. Later model versions do not establish its condition then.
- I do not assign an unstated numerical cutoff to 'very little'.
Counts as supported if: Repeated evaluations of the contemporary GPT-4 model show that reliable successes are uncommon across a reasonably broad selection of tasks, including both successful and failed summarization attempts.
Counts as contradicted if: Comparable repeated evaluations establish reliable success across a substantial breadth of contemporary GPT-4 capabilities, inconsistent with successes being only rare exceptions.
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 establish reliability limitations, but they do not settle the much broader assertion that very little of GPT-4's performance was reproducible. The later model-drift study addresses a different question. Neither occasional failures nor selected successes establish the prevalence claimed here.
Limitations
- The task population and meaning of full replication are unspecified.
- The summarization examples lack independently verified repeated trials in the evidence collected.
Evidence (2 items)
Context only: OpenAI,
not fully reliable
Why it matters here: OpenAI acknowledged errors in the contemporary model. This does not measure how many capabilities reproduce reliably or directly test the compression examples.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: Lingjiao Chen, Matei Zaharia and James Zou / arXiv,
vary greatly over time.
Why it matters here: The study compares March and June model versions. Changes between versions do not establish widespread failure to reproduce results within the March model.
- Where
- Abstract, version 3
- Publisher
- Lingjiao Chen, Matei Zaharia and James Zou / arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A broad evaluation of the initial GPT-4 model with repeated trials and disclosed settings, including faithful summarization and the compression-and-reconstruction task.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, GPT-5 and irrational exuberance
And very little of what GPT-4 does is fully replicable. Sometimes summarization (aka compression) works (super cool!) and sometimes it doesn’t, as many were quick to point out:
- Where
- paragraph 16
- Published
- 2023-04-08
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-gpt-5-and-irrational-exuberance-da17ab-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-gpt-5-and-irrational-exuberance-da17ab-a1.