“Symbolic systems have always been good for outliers; neural networks have always struggled with them. For 30 years, this has been a constant.”
Our reading, the sentence that was checked:
For 30 years, neural networks have struggled with outliers while symbolic systems handled them well.
Assessment: Mixed Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, This one important fact about current AI explains almost everything, paragraph 31.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-this-one-important-fact-about-current-4b8735-a2
We recorded 10 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 2024-08-01. 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
- Symbolic systems have always been good for outliers
- neural networks have always struggled with them
- For 30 years
- a constant
Our interpretation
- I read this as a persistent contrast between the two approaches, rather than perfect symbolic performance on every possible input.
- The contextual comparison concerns applying general rules to unfamiliar cases, including new combinations of familiar elements.
- I use the preceding 30 years, 1994-08-01 through 2024-08-01. Earlier results are context.
- Successful application of a supplied symbolic rule is distinct from discovering an appropriate rule for a new problem.
Counts as supported if: Comparative evidence across the period shows recurring neural difficulty and reliable symbolic rule application on unfamiliar cases, with the contrast remaining broadly stable.
Counts as contradicted if: Controlled comparisons show that the asserted contrast breaks down: neural systems achieve systematic transfer, symbolic systems have material failures, or their relative performance depends substantially on task and training.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Mixed
Part of the claim holds and part does not, and the record says which part is which.
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 neural-difficulty part holds in the cited tests: the 2020 gSCAN study found substantial failures on unfamiliar compositions, and the 2023 comparison found ordinary neural models lacked systematicity [e6; e7]. The 2023 symbolic comparators were systematic but rigid [e7]. The constant contrast does not hold: that same study found a meta-trained neural network achieved both systematicity and flexibility in its generalization tests [e7].
Limitations
- The studies do not provide a continuous, comparable record for August 1, 1994 through August 1, 2024.
- The February 1994 unseen-verb comparison predates the target period [e8].
- The evidence does not establish performance across all kinds of outliers or equate executing supplied rules with learning appropriate rules.
Evidence (3 items)
Supports the claim: arXiv; Ruis and colleagues,
struggle to interpret novel compositions.
Why it matters here: The gSCAN experiments report substantial failures when neural models must apply compositional rules to unfamiliar combinations. This supports the recurring neural-difficulty part of the claim, but does not establish symbolic superiority across the entire period.
- Publisher
- arXiv; Ruis and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Goes against the claim: Nature,
several systematic generalization benchmarks.
Why it matters here: In a comparison of seven models, the meta-trained neural approach combined systematicity and flexibility that the symbolic and ordinary neural comparators did not jointly achieve. This contradicts a stable categorical division between successful symbolic generalization and unsuccessful neural generalization.
- Publisher
- Nature
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Context only: Journal of Artificial Intelligence Research; arXiv copy,
under different representations.
Why it matters here: The paper reports better unseen-verb generalization by its symbolic pattern associator than by the tested neural models. This supports the historical motivation, but precedes the exact target window and cannot establish an unchanged 30-year pattern.
- Publisher
- Journal of Artificial Intelligence Research; arXiv copy
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, This one important fact about current AI explains almost everything
Symbolic systems have always been good for outliers; neural networks have always struggled with them. For 30 years, this has been a constant.
- Where
- paragraph 31
- Published
- 2024-08-01
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-this-one-important-fact-about-current-4b8735-a2: Current. Outcome: Mixed. 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-this-one-important-fact-about-current-4b8735-a2.