“The underlying technology—often referred to as Foundation Models, because systems are fine-tuned on top of large pretrained (foundation) model—is deeply unstable and thus difficult to engineer into reliable products.”
Our reading, the sentence that was checked:
Foundation models are deeply unstable, which makes them difficult to engineer into reliable products.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, What exactly are the economics of AI?, paragraph 24.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-what-exactly-are-the-economics-of-62834f-a1
We recorded 12 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 2023-08-14. 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
- Foundation Models
- deeply unstable
- thus difficult to engineer into reliable products
- answers can even change from one month to the next
Our interpretation
- The surrounding discussion concerns large language models. I assess foundation models used for language applications, rather than every kind of foundation model.
- The plain reading is that substantial errors, sensitivity to inputs, and changes between model versions create material obstacles to reliable product engineering. It does not require every answer to vary or reliable products to be impossible.
- The claim describes the technology available when the post appeared on 2023-08-14. No measurement window or reporting cycle is specified; the target period therefore has no invented starting date.
- Engineering difficulty means demonstrated integration failures or a need for additional validation, output handling, monitoring, or version management.
Counts as supported if: Contemporaneous controlled evaluations demonstrate consequential instability across inputs or model versions, and documented integration experiments connect that behavior to failures or additional reliability engineering.
Counts as contradicted if: Representative contemporaneous evaluations and integration evidence show reliably stable behavior across relevant inputs and versions, or show that the alleged instability does not materially impede reliable product engineering.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
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 evidence supports the contextual reading: contemporary language models had substantial behavioral weaknesses that complicated reliable integration. Controlled experiments found sensitivity to supplied context, while code and LangChain tests linked formatting behavior directly to execution and parsing failures. This establishes a concrete engineering obstacle. The qualification matters: some failures were repairable, some newer-model results improved, and version pinning reduced update disruption. The evidence supports difficulty, not impossibility, universal deterioration, or unavoidable drift in every fixed model.
Limitations
- The judgment depends on reading the broad wording through its language-model context.
- Deeply unstable is qualitative; the sources do not supply a common threshold for that description.
- The cited experiments were preprints at the time and covered selected tasks rather than representative production deployments.
- Engineering burden is inferred from demonstrated failures and mitigations, not measured through a controlled study of development costs.
Evidence (3 items)
Supports the claim: arXiv; Lingjiao Chen, Matei Zaharia, and James Zou,
manually debugging such issues
Why it matters here: Comparisons of March and June model versions document formatting failures that prevented code execution and caused a LangChain parser to fail, directly connecting model behavior to integration difficulty. The results also limit the claim: GPT-4's agent performance improved, and removing non-code wrappers increased its code-test success from 52% to 70%, so behavioral change did not imply declining underlying capability.
- Publisher
- arXiv; Lingjiao Chen, Matei Zaharia, and James Zou
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv; Nelson F. Liu and colleagues,
performance substantially decreases
Why it matters here: Controlled changes to document order and context length impaired question answering across several model families, supporting consequential input sensitivity in a setup resembling retrieval-based products. Mitigations were task-dependent: placing the query before and after the data fixed the tested synthetic retrieval task but made little difference to the question-answering pattern.
- Publisher
- arXiv; Nelson F. Liu and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: OpenAI,
Each individually pinned model
Why it matters here: Version pinning was an available mitigation, so changes between releases do not establish unavoidable month-to-month drift within a fixed version. The same disclosure acknowledged application disruption from model changes; its stability statement concerns provider updates, not guaranteed correctness on every input.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, What exactly are the economics of AI?
The underlying technology—often referred to as Foundation Models, because systems are fine-tuned on top of large pretrained (foundation) model—is deeply unstable and thus difficult to engineer into reliable products.
- Where
- paragraph 24
- Published
- 2023-08-14
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-what-exactly-are-the-economics-of-62834f-a1: 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-what-exactly-are-the-economics-of-62834f-a1.