“• AGI should have reliable ability to reason over abstractions. (Ongoing unpredictable errors in LLMs reflect this.)”
Our reading, the sentence that was checked:
LLMs still lack a reliable ability to reason over abstractions, as their ongoing unpredictable errors show.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Ezra Klein’s new take on AGI – and why I think it’s probably wrong, paragraph 20.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-ezra-kleins-new-take-on-agi-and-why-236597-a2
We recorded 29 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 2025-03-05. 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
- reliable ability to reason over abstractions
- Ongoing unpredictable errors in LLMs
Our interpretation
- The empirical claim is that contemporary LLMs still make recurring errors when applying abstract rules. The normative statement about what AGI should possess is not separately scored.
- Reliable means dependable performance across relevant new examples and task variations, not literal perfection or an absence of any reasoning ability.
- Unpredictable is read as unexpected or inconsistent failures in use, not as a mathematical claim that aggregate error rates cannot be predicted.
- With no reporting period supplied, the target is the latest pre-publication evaluation period in 2025, through March 5. Older experiments are context only.
Counts as supported if: Controlled evaluations of contemporary LLMs document recurring failures in applying or inferring abstract rules across new examples, despite success on easier or familiar cases.
Counts as contradicted if: Broad, controlled contemporary evaluations establish dependable abstract-rule reasoning across relevant variations, with the alleged recurring failures explained by invalid tests or otherwise not reproduced.
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
MIR-Bench, published February 14, 2025, reports recurring failures in inferring rules and applying them to new inputs across 15 models [e1]. GSM-∞, published February 7, 2025, finds high accuracy on simpler problems and declining accuracy as required operations increase [e2]. These results support limited reliability in the pre-March 5, 2025 period [e1, e2]. MIR-Bench also finds robustness to incorrect examples, showing successful reasoning alongside the failures [e3]. Earlier context includes an o3 configuration scoring 87.5% on ARC-AGI in December 2024 [e4]. These successes qualify the finding without establishing dependable performance across task variations [e1, e3, e4]. GSM-∞ finds predictable aggregate declines, so its results do not support a literal claim that error rates are statistically unpredictable [e2].
Limitations
- The February studies are preprints and do not provide complete execution timestamps.
- The evidence does not cover every contemporary model or configuration or establish a universal reliability threshold.
- Behavioral failures do not establish that abstract representations or reasoning abilities are absent.
- The December 2024 ARC-AGI result predates the target period.
Evidence (4 items)
Supports the claim: Kai Yan and colleagues, arXiv,
open
Why it matters here: MIR-Bench tests fifteen models on deriving outputs from examples of underlying functions. Its controlled evaluations find substantial remaining failures, including in o1-family models, supporting the limited reliability claim.
- Publisher
- Kai Yan and colleagues, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: Yang Zhou and colleagues, arXiv,
LLM accuracy declines
Why it matters here: The generated problems permit controlled changes in reasoning complexity and show recurring failures across models. The predictable aggregate decline also cautions against interpreting unpredictable literally.
- Publisher
- Yang Zhou and colleagues, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Kai Yan and colleagues, arXiv,
quite robust
Why it matters here: The same study finds resilience to incorrect examples. This is material evidence of successful reasoning behavior, limiting any broader inference that LLM abstraction is uniformly fragile or absent.
- Publisher
- Kai Yan and colleagues, arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: ARC Prize,
scored 87.5%.
Why it matters here: The benchmark organizers report substantial success on novel abstract tasks, alongside remaining easy-task failures. This is material counterevidence to a claim of no abstraction ability, but it does not establish general reliability and precedes the target period.
- Publisher
- ARC Prize
- Kind of source
- Benchmark result
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Ezra Klein’s new take on AGI – and why I think it’s probably wrong
• AGI should have reliable ability to reason over abstractions. (Ongoing unpredictable errors in LLMs reflect this.)
- Where
- paragraph 20
- Published
- 2025-03-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-ezra-kleins-new-take-on-agi-and-why-236597-a2: 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-ezra-kleins-new-take-on-agi-and-why-236597-a2.