“In the best case (not always reached) they can write python code, supplementing their own weaknesses with outside symbolic code, but even this is not reliable.”
Our reading, the sentence that was checked:
LLMs writing Python code to supplement their weaknesses is not reliable.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, A knockout blow for LLMs?, paragraph 34.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-a-knockout-blow-for-llms-e0245c-a3
We recorded 19 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 2025-06-07. 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)
From the speaker
- write python code
- supplementing their own weaknesses with outside symbolic code
- even this is not reliable
Our interpretation
- The ordinary reading is that generating and executing Python does not by itself make LLM solutions dependably correct across tasks.
- Reliable does not mean mathematically perfect. Material, recurring errors matter; isolated failures would be insufficient.
- Useful code generation and substantial performance improvements are compatible with this claim.
- Use calendar 2025 through June 7 for contemporary capability reports because the statement gives no period and this field has no standard reporting interval.
Counts as supported if: Execution-based evaluations show material errors or unstable task performance when contemporary LLMs use generated Python to solve verifiable problems.
Counts as contradicted if: Broad, representative execution-based evaluations show dependably correct task completion using generated Python, without material correction or task-specific intervention.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
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
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
The evaluation published May 16, 2025 found that executing generated Python substantially improved localization results but generally worsened tracking results, with significant errors from generated algorithms [e1]. For o3-mini, code assistance sometimes increased error by up to 354% [e1]. This supports the claim that generated Python did not by itself ensure dependable correctness across the tested tasks, despite substantial benefits on some tasks [e1].
Limitations
- The decisive evaluation covers spatiotemporal tasks; it does not establish unreliability in every Python application [e1].
- The reported metrics measure numerical error, not a universal task-completion success rate; exact experiment dates are not supplied [e1].
- Successful Hanoi demonstrations were preliminary, used Lua, and were reported June 16, 2025, after the January 1–June 7 observation window [e2].
Evidence (2 items)
Supports the claim: Quan et al., arXiv,
Conversely, for tracking tasks, CI generally degrades performance.
Why it matters here: Executed Python improved localization substantially but worsened tracking; the study attributes significant errors to generated algorithms and optimization inaccuracies. This supports the limited claim that adding Python does not assure dependable solutions.
- Publisher
- Quan et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: A. Lawsen, arXiv,
correctly implement the recursive algorithm
Why it matters here: Preliminary demonstrations found successful Hanoi solutions when models generated Lua functions. This cautions against treating puzzle-output failures as proof that code generation cannot help, but it concerns another language and later tests.
- Publisher
- A. Lawsen, 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, A knockout blow for LLMs?
In the best case (not always reached) they can write python code, supplementing their own weaknesses with outside symbolic code, but even this is not reliable.
- Where
- paragraph 34
- Published
- 2025-06-07
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-a-knockout-blow-for-llms-e0245c-a3: 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.