“Like the first generalization, the new ones will eventually run out.”
Our reading, the sentence that was checked:
The new scaling laws (post-training and inference-time) will eventually stop holding.
Assessment: Unresolved No deadline stated on 2026-10-29 Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, The new AI scaling law shell game, paragraph 10.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-a-new-ai-scaling-law-shell-game-baa1e0-a2
We recorded 12 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2024-11-24. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Forecast
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No deadline stated
- Deadline state on 2026-10-29
- No deadline stated (the release’s reference date)
From the speaker
- the new ones
- will eventually run out
- scaling with how long you post-train a model
- scaling with how long you let a given model wrestle with a given problem
Our interpretation
- The forecast covers both post-training and inference-time scaling.
- The ordinary reading is that extending these scaling approaches will eventually cease producing meaningful performance improvements.
- A plateau for one model, task or implementation is relevant but does not by itself establish that both scaling approaches have run out.
- The claim supplies neither a deadline nor a numerical definition of meaningful improvement.
Counts as supported if: Public experiments establish sustained plateaus or breakdowns in both scaling relationships across meaningful task sets as compute increases, with enough controls to distinguish the claimed limits from a particular implementation's failure.
Counts as contradicted if: A decisive result establishes that the claimed eventual breakdown cannot occur within the stated scope. Continued gains over a finite compute range do not refute an eventual limit.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Did the forecast come true?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
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 deadline stated
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
Experiments have identified real limits in particular inference and post-training methods. Other experiments have overcome apparent plateaus. These results do not decisively establish that both scaling approaches have run out. Marcus gave no deadline, so continued progress cannot make this eventual forecast false either.
Limitations
- A narrower reading limited to particular scaling recipes would be easier to support.
- The cited limits and improvements cover finite experimental settings.
Evidence (2 items)
Context only: arXiv,
Resampling cannot decrease this probability, so it imposes an upper bound to the accuracy of resampling-based inference scaling, regardless of compute budget.
Why it matters here: This version reports a bound arising from imperfect verifiers. It is evidence of a specific inference-scaling limit, but does not establish limits for every inference method or for post-training.
- Where
- Abstract
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: NVIDIA,
The divergence was clear: ProRL stagnated, while BroRL delivered steady, significant gains in less time.
Why it matters here: The experiment reports a post-training plateau that a different rollout allocation overcame. This cautions against treating an observed plateau as exhaustion of post-training scaling generally.
- Publisher
- NVIDIA
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
What would settle it
Reproducible compute sweeps showing sustained breakdown of both claimed scaling relationships across meaningful task sets, with controls distinguishing broad limits from implementation-specific plateaus.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, The new AI scaling law shell game
Like the first generalization, the new ones will eventually run out.
- Where
- paragraph 10
- Published
- 2024-11-24
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-a-new-ai-scaling-law-shell-game-baa1e0-a2: Current. Outcome: Unresolved. 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-a-new-ai-scaling-law-shell-game-baa1e0-a2.