“And, crucially, the failure to generalize adequately outside distribution tells us why all the dozens of shots on goal at building “GPT-5 level models” keep missing their target. It’s not an accident. That failing is principled.”
Our reading:
Failure to generalize outside the training distribution is the principled reason efforts to build GPT-5-level models keep missing their target.
Assessment: Unresolved
What happened
Does the evidence support the cause?
Outcome: Unresolved
We checked, and the evidence so far doesn’t settle it.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
The ASU paper version available August 5, 2025 reported declining reasoning effectiveness as training and test transformations differed more in controlled synthetic tasks [e1]. An April 2025 study reported improved generalization on tested code-reasoning tasks, including mutated programs [e2]. Apple's June 2025 paper reported accuracy collapse as puzzle complexity increased, without establishing distribution shift as the cause [e3]. These studies document task-specific failures and improvements, but do not connect them to identified development efforts and their missed capability targets [e1, e2, e3]. They therefore neither establish nor displace the proposed explanation for recurring frontier-model development shortfalls [e1, e2, e3].
Caveats
- The packet does not identify the relevant development efforts, their capability targets or their measured shortfalls.
- Results on selected tasks do not establish a common cause across development efforts.
- The code study does not establish commercial models' training exposure.
- Publication dates do not establish exact experiment dates.
Evidence (3 items)
Context only: Chengshuai Zhao and colleagues, Arizona State University; arXiv,
effectiveness of CoT reasoning decreases
Why it matters: The controlled synthetic experiments support the proposed distribution-shift mechanism within their tested setting. They do not identify the claimed development efforts, establish their targets, or show why those targets were missed.
- Publisher
- Chengshuai Zhao and colleagues, Arizona State University; arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Rem Yang and colleagues, MIT and Brown University; arXiv,
substantial increase in LLM generalization
Why it matters: Tests across ten models found improved generalization on code-reasoning tasks, including mutated and generated programs intended to probe distribution differences. This limits a blanket inference of generalization failure, but success on these tasks does not establish adequate broad generalization or resolve the claimed cause of missed development targets.
- Publisher
- Rem Yang and colleagues, MIT and Brown University; arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Parshin Shojaee and colleagues, Apple; arXiv,
complete accuracy collapse
Why it matters: The study establishes failures on controlled puzzle tasks as complexity increases. It does not isolate training-distribution mismatch as their cause or connect them to the development efforts Marcus describes.
- Where
- Abstract
- Publisher
- Parshin Shojaee and colleagues, Apple; arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
How we judged it
- Kind of claim
- Causal claim
- Deadline
- None given. We don’t add one.
Where we had to interpret
- The plain reading is that a systematic weakness in handling unfamiliar problems explains recurring failures to achieve the broad capabilities sought from frontier models.
- The claim concerns development efforts up to the post on 2025-08-09. It is not a forecast that no future model can overcome this weakness.
- “Principled” means a recurring technical limitation, rather than an accidental failure. It need not mean that distribution shift is the only contributing cause.
- Marcus supplies neither a capability threshold for “GPT-5 level” nor a list of the development efforts and their targets. Selected benchmark failures can test the proposed mechanism but cannot alone establish the full causal attribution.
Right if: Controlled evaluations show that distribution shift systematically causes relevant reasoning failures, and evidence connecting those failures to the development efforts establishes that this weakness explains their recurring shortfalls against documented capability targets.
Wrong if: Evidence from the relevant development efforts shows that their missed targets occurred despite adequate generalization on the capabilities at issue, with controlled tests or development records identifying a different explanation that displaces the proposed common cause.
We wrote this test after the outcome was known.
What would settle it
Document the relevant development efforts and targets through August 9, 2025. Obtain controlled tests and development records showing whether distribution-shift failures explain their shortfalls, or whether adequate generalization and another demonstrated cause displace that explanation.
We found 33 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, GPT-5: Overdue, overhyped and underwhelming. And that’s not the worst of it., paragraph 67.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-28. The post may have been edited after 2025-08-09.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, GPT-5: Overdue, overhyped and underwhelming. And that’s not the worst of it.
And, crucially, the failure to generalize adequately outside distribution tells us why all the dozens of shots on goal at building “GPT-5 level models” keep missing their target. It’s not an accident. That failing is principled.
- Where
- paragraph 67
- Published
- 2025-08-09
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Unresolved. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
The final check failed: a cited source didn’t say what the write-up said, and our summary didn’t match their words.