Seven replies to the viral Apple reasoning paper – and why they fall short
We recorded 17 claims from this post. 8 were checked against evidence. 9 can’t be checked. 3 more are held. Read the post at its source.
Claims from this post (17)
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: UnresolvedSaid
“The real takeaway from the Apple paper is that LLMs can’t be trusted to run algorithms as complexity and distance from the training distribution grows (just as humans shouldn’t serve as calculators).”
What we tested: LLMs cannot be trusted to run algorithms as problem complexity and distance from the training distribution grow.
Assessment: Can’t be checkedSaid
“If we want to get to AGI, we will have to better.”
What we tested: Reaching AGI will require doing better than current LLMs.
Assessment: SupportedSaid
“The LRMs failed on Tower of Hanoi with 8 discs, where the optimal solution is 255 moves, well within so-called token limits;”
What we tested: The LRMs failed Tower of Hanoi with 8 discs, whose optimal solution of 255 moves is well within so-called token limits, so the token-limit objection does not explain the overall pattern of results.
What happened: The authors’ retrospective account of the original study reports errors around move 40 for eight-disc puzzles requiring 255 moves. This directly supports failures occurring before the required answer approaches the output ceiling. (arXiv; Shojaee and colleagues, )
Assessment: Can’t be checkedSaid
“The length limit on LLM is a bug, and most certainly not a feature.”
What we tested: The output length limit of LLMs is a bug, and most certainly not a feature.
Assessment: Can’t be checkedSaid
“One is left simply having to test everything, all the time, with little guarantees of anything. Some model might be big enough for task T of size S and fail on the next size, or on Task T’ that is slightly different, etc. It all becomes a crapshoot.”
What we tested: Because some model might be big enough for a task of one size yet fail on the next size or on a slightly different task, one is left having to test everything all the time with little guarantee of anything, making it all a crapshoot.
Assessment: UnresolvedSaid
“But here’s the thing, we can’t know in advance what model is big enough (if any) for any given problem.”
What we tested: We can't know in advance which model, if any, is big enough for any given problem.
Assessment: Can’t be checkedSaid
“Huge vindication for what I have been saying all along: we need AI that integrates both neural networks and symbolic algorithms and representations (such as logic, code, knowledge graphs, etc).”
What we tested: AI should integrate neural networks with symbolic algorithms and representations.
Assessment: Can’t be checkedSaid
“True in some cases, and a huge win for neurosymbolic AI, given that they can’t reliably solve the puzzles without code, and given that code is symbolic.”
What we tested: That the systems can solve the puzzles with code, which is true in some cases, is a huge win for neurosymbolic AI, given that they can't reliably solve the puzzles without code and code is symbolic.
Assessment: UnresolvedSaid
“But downloading code without conceptual understanding is of less help in the case of new problems, dynamically changing environments, and so on.”
What we tested: LLMs retrieving code without conceptual understanding help less on new problems and dynamically changing environments.
Assessment: UnresolvedSaid
“But also, we need to do so reliably, and in a general way and we haven’t yet crossed that threshold.”
What we tested: Current AI does not yet integrate neural and symbolic methods reliably and generally.
Assessment: SupportedSaid
“I doubt any of them are perfect, but the four together provide converging evidence with dozens of other prior papers (including some of my own), and I am confident many more examples will be uncovered.”
What we tested: Many more examples of LLM/LRM failures at applying algorithms, like those in the Apple paper, will be uncovered.
What happened: The published study reports complexity-related declines across nine validated reasoning tasks, including satisfiability, cryptarithmetic, graph coloring, water jugs, Sudoku, and Rubik's Cube. It supplies additional algorithmic and constraint-following examples beyond Apple's original task set. (arXiv; Utsho and colleagues, )
Assessment: UnresolvedSaid
“Mark my words: in time, we will see scores of papers reinforcing the Apple results.”
What we tested: Over time, scores of papers will reinforce the Apple paper's findings on reasoning model limitations.
Assessment: Can’t be checkedSaid
“The real news here, aside from the fact that this was a clever study nailing down an important point, is that people are finally starting to pay attention, to (one of the) two biggest Achilles’ Heels of generative AI, and to appreciate its significance.”
What we tested: People are finally starting to pay attention to poor generalization, one of the two biggest Achilles' heels of generative AI, and to appreciate its significance.
Assessment: Can’t be checkedSaid
“If people like Sam Altman are sweating, it’s because they should.”
What we tested: If people like Sam Altman are sweating, it is because they should be.
Assessment: Can’t be checkedSaid
“None of the rejoinders are compelling.”
What we tested: None of the roughly seven rejoinders to the Apple paper reviewed in this post are compelling.
Assessment: Can’t be checkedSaid
“Taking the SalesForce report together with the Apple paper, it’s clear the current tech is not to be trusted.”
What we tested: Taking the Salesforce report together with the Apple paper, it is clear that the current tech is not to be trusted.
Assessment: UnresolvedSaid
“Update: The most popular reply to the Apple paper turned out to be an error-ridden joke.”
What we tested: The most popular rebuttal to the Apple reasoning paper was full of errors.
Held from this post (3)
These records did not pass every check, so nothing about them is published. Held is not a finding.
- The two reviews read the claim differently
- The two reviews read the claim differently
- The two reviews reached different assessments. Our reading, not assessed: In many cases LLMs are a step backwards compared with existing systems that run such algorithms without errors, whereas AGI should be a step forward.