“That, in a nutshell, is why we should never trust pure LLMs; even under carefully controlled circumstances with massive amounts of directly relevant data, they still never really get even the most basic linear functions.”
Our reading:
Even with controlled conditions and massive relevant data, pure LLMs never truly learn even the most basic linear functions.
Assessment: Unresolved
What happened
Was it true for the period described?
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
MathGLM's declining accuracy supports Marcus's concern about reliability, but does not establish his categorical conclusion. Contemporary experiments also demonstrated meaningful extrapolation. One reported near-perfect performance through 50 digits using attention calibration. Treating that as decisive would require resolving whether a specialized calibrated transformer counts as a pure LLM. Treating ordinary held-out accuracy as decisive would ignore his emphasis on abstraction and extrapolation. The available evidence therefore supports the narrower criticism without settling the full claim.
Caveats
- The assessment concerns capabilities existing by 2023-10-19, not subsequent model improvements.
- The experimental reports are primary preprints; no independent replication is established here.
- Neither the boundary of pure LLMs nor a numerical reliability threshold is specified.
Evidence (3 items)
Supports the claim: Zhen Yang and colleagues, arXiv,
decline in accuracy
Why it matters: Supports the narrower observation that MathGLM remained unreliable as arithmetic inputs grew. The experiment added 50,000 training records within the 12-digit range, so its declining performance does not itself establish a universal inability to extrapolate beyond training lengths.
- Publisher
- Zhen Yang and colleagues, arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Samy Jelassi and colleagues, arXiv,
generalize to longer sequences
Why it matters: Material counterevidence to an absolute absence of extrapolation: models trained on five-digit addition reached 99.9% accuracy at ten digits and 98.3% at fifteen. Accuracy fell to 21.3% at twenty digits, so this does not establish the sustained reliability emphasized by Marcus.
- Publisher
- Samy Jelassi and colleagues, arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Shaoxiong Duan and Yining Shi, arXiv,
up to 50 digits
Why it matters: The original report describes substantial extrapolation, including addition trained through six digits, using attention patterns extracted, extended, and incorporated into retraining. This challenges the broader claim about transformer learning, but whether these specialized calibrated models qualify as pure LLMs is unresolved.
- Publisher
- Shaoxiong Duan and Yining Shi, arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Conditions
- even under carefully controlled circumstances with massive directly relevant data (their words)
Where we had to interpret
- Read as a claim about capabilities demonstrated by the post date, not a forecast that learning will remain impossible forever.
- The context makes reliable extrapolation central: success on unseen examples within the training range alone does not contradict the claim.
- Operationally, learning requires consistently correct answers across substantially longer inputs or other meaningful departures from training coverage, with evidence of a reusable rule. A few successful examples are insufficient.
- Pure LLMs means learned language models producing answers without an external calculator, executed program, or symbolic arithmetic module. Whether specialized arithmetic transformers with an additional attention-calibration procedure qualify remains ambiguous.
- Assess elementary arithmetic in the context supplied, including multiplication. Do not replace this with an easier test of approximate linear regression merely because the sentence says linear functions.
- The target is the state of published capabilities through 2023-10-19. No beginning date is supplied, and this research claim has no standard quarterly reporting period.
Right if: Controlled studies show that relevant pure LLMs still fail reliable extrapolation despite extensive arithmetic training. Supporting the categorical claim additionally requires evidence broad enough to justify its generalization beyond particular tested models; MathGLM failures alone are insufficient.
Wrong if: A qualifying pure LLM existing by the post date demonstrably learns a reusable elementary arithmetic rule from data and reliably applies it substantially beyond training coverage. Within-range accuracy alone, an external arithmetic engine, or a different model class does not suffice.
We wrote this test after the outcome was known.
What would settle it
A documented contemporaneous pure LLM showing sustained exact arithmetic performance substantially beyond training coverage, with methods establishing that its learned rule performs the computation, would contradict the claim. Supporting the categorical conclusion would require broader evidence than failures of selected models.
We found 6 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, “Math is hard” — if you are an LLM – and why that matters, paragraph 8.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-29. The post may have been edited after 2023-10-19.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, “Math is hard” — if you are an LLM – and why that matters
That, in a nutshell, is why we should never trust pure LLMs; even under carefully controlled circumstances with massive amounts of directly relevant data, they still never really get even the most basic linear functions.
- Where
- paragraph 8
- Published
- 2023-10-19
- 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
A reviewer said our test wasn’t fair to what was actually said.