Five quick updates about that Apple reasoning paper that people can’t stop talking about
We recorded 9 claims from this post. 4 were checked against evidence. 5 can’t be checked. 1 more is held. Read the post at its source.
Claims from this post (9)
In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.
Assessment: SupportedSaid
“The core of the paper challenged the idea that LLMs can reason on any deep level, and showed that they broke down on moderately complex versions of basic algorithms (since Tower of Hanoi with 8 discs, first solved with classical techniques around 1957) — casting serious doubt into whether LLMs on their own could ever achieve AGI.”
What we tested: The Apple paper challenged the idea that LLMs can reason on any deep level and showed that they broke down on moderately complex versions of basic algorithms such as Tower of Hanoi with 8 discs, which casts serious doubt on whether LLMs on their own could ever achieve AGI.
What happened: Figure 7 reports near-zero intermediate-solution accuracy for Hanoi at eight or more discs. Section 4.4 reports that supplying the algorithm did not remove the collapse. (Apple researchers, arXiv, )
Assessment: Can’t be checkedSaid
“To some extent they do this better than in 1998, but is still the core challenge.”
What we tested: Although neural networks generalize beyond their training data somewhat better than in 1998, distribution shift is still the core challenge for them.
Assessment: SupportedSaid
““The illusion of the illusion” turned out to be an error-ridden joke.”
What we tested: The rejoinder “The Illusion of the Illusion of Thinking” turned out to be an error-ridden joke.
What happened: Lawsen's firsthand retrospective explicitly describes the work as intended satire and acknowledges serious mistakes. The source-type label reflects the blog format; he is the primary witness to his intent. (Alex Lawsen, )
Assessment: SupportedSaid
“A new paper with a hard coding benchmark called LiveCodeBenchPro designed to resist data contamination shows still more converging evidence for the two core notions behind the Apple paper (and behind my own decades-long critique): (a) the systems are challenged at reasoning, and (b) performance declines as one strays further and further from familiar problems.”
What we tested: LiveCodeBenchPro gives further evidence that LLMs are challenged at reasoning and that their performance declines the further problems are from familiar ones.
What happened: The category analysis reports stronger performance on reusable algorithmic patterns and weaker performance on tasks requiring novel observations. (Zihan Zheng and colleagues, arXiv, )
Assessment: SupportedSaid
“In sum, performance on hard coding problems, an alleged area of LLM strength, drops to zero when you control nicely for contamination.”
What we tested: When data contamination is controlled for, LLM performance on hard coding problems drops to zero.
What happened: Table 1 reports 0.0% Hard pass@1 for every listed model. The quoted caption additionally confirms zero for the tested o4-mini variants with multiple attempts; the protocol reduces contamination and excludes external tools. (Zihan Zheng and colleagues, arXiv, )
Assessment: Can’t be checkedSaid
“What the Apple paper said still looks to be basically correct.”
What we tested: The Apple reasoning paper's conclusions about LLM reasoning limits still look basically correct.
Assessment: Can’t be checkedSaid
“On familiar problems with no wrinkles, they are great. On anything else they are suspect.”
What we tested: Generative AI systems do well on familiar problems with no wrinkles, but their outputs on anything else are suspect.
Assessment: Can’t be checkedSaid
“If we want to get to AGI, we’ll need to do a lot better.”
What we tested: Reaching AGI will require doing much better than current generative AI on distribution shift.
Assessment: Can’t be checkedSaid
“Gary Marcus is sorry to keep writing about how LLMs struggle with distribution shift, but it’s the whole ballgame.”
What we tested: LLMs struggle with distribution shift, and this is the whole ballgame.
Held from this post (1)
These records did not pass every check, so nothing about them is published. Held is not a finding.
- The record does not meet the current rules. Our reading, not assessed: After almost three decades of work, generative AI still struggles with distribution shift.