“Debugging with deep learning is wicked hard, because nobody (a) really understands how it works, and (b) knows how to fix problems other to collect more data and add more layers and so forth.”
Our reading, the sentence that was checked:
Debugging deep learning is wicked hard because nobody really understands how it works or knows how to fix problems other than collecting more data, adding more layers and so forth.
Assessment: Mixed Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Dear Elon Musk, here are five things you might want to consider about AGI, paragraph 29.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-dear-elon-musk-here-are-five-things-d1f821-a1
We recorded 21 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2022-05-31. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Causal claim
- 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
- Debugging with deep learning is wicked hard
- nobody (a) really understands how it works
- knows how to fix problems other to collect more data and add more layers and so forth
Our interpretation
- Assess the state of knowledge available when the post appeared, on 2022-05-31. The target period has no lower bound because established earlier methods remain relevant to that state.
- The ordinary reading is that limited understanding makes diagnosis and repair difficult, and that practitioners lack effective targeted remedies beyond broadly changing data, training, or model size.
- Assess the difficulty and its proposed explanation separately. A targeted repair need not solve debugging generally to count against the assertion that such remedies were unknown.
- Do not interpret 'understands' as requiring complete mathematical understanding, or 'and so forth' as automatically encompassing every possible repair.
Counts as supported if: Contemporaneous research or engineering evidence shows that model opacity materially obstructed diagnosis and repair, and that effective remedies were generally limited to broad data or training changes.
Counts as contradicted if: Contemporaneous work demonstrates useful understanding of internal mechanisms and effective targeted repairs beyond collecting more data or enlarging models. Evidence that debugging was routinely straightforward would also contradict the difficulty claim.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
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
Contemporary research supports Marcus's point that opacity makes debugging harder. But the stronger explanation overstates the limits of existing methods: before the post, ROME demonstrated causal investigation of factual recall and targeted edits to model weights. The difficulty claim holds in this narrower sense; the assertion that useful understanding and targeted remedies were absent does not.
Limitations
- The evidence concerns neural language models, not every deep learning application.
- The supporting survey gives a qualitative assessment, not a controlled estimate of the causal contribution of opacity.
- The mixed verdict depends on treating the assertion about absent remedies as a substantive commitment rather than loose rhetoric.
Evidence (2 items)
Supports the claim: arXiv; Piyawat Lertvittayakumjorn and Francesca Toni,
opaque deep learning models.
Why it matters here: The survey supports the narrower causal point that opacity makes locating and repairing problems harder. Its review of explanation-based debugging also shows that researchers were already developing remedies.
- Publisher
- arXiv; Piyawat Lertvittayakumjorn and Francesca Toni
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Goes against the claim: arXiv; Kevin Meng, David Bau, Alex Andonian and Yonatan Belinkov,
editing facts stored in model weights.
Why it matters here: The authors used causal interventions to investigate factual recall and evaluated targeted factual edits on CounterFact. This supplies a concrete pre-post example of useful mechanistic understanding and repair without adding data or layers.
- Publisher
- arXiv; Kevin Meng, David Bau, Alex Andonian and Yonatan Belinkov
- 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, Dear Elon Musk, here are five things you might want to consider about AGI
Debugging with deep learning is wicked hard, because nobody (a) really understands how it works, and (b) knows how to fix problems other to collect more data and add more layers and so forth.
- Where
- paragraph 29
- Published
- 2022-05-31
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-dear-elon-musk-here-are-five-things-d1f821-a1: 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-dear-elon-musk-here-are-five-things-d1f821-a1.