We should be careful how we ask LLM questions because they tend to respond only to what is asked, often in a way that reinforces the user's implied preferences. IMO, in consulting an LLM, we shouldn't ask them to think for us or support what we say but to find missteps and errors in our own thinking. We should be able to stand by our argument in a manner not reliant on the LLMs outputs.

I don't believe in pasting LLM responses but I think it's ok here. Here is what Gemini 2.5 pro says to itself when given the full paper's context:

Accurate/Reasonable Parts of the Statement:

1. "The paper’s findings don't prove reasoning is an illusion..." (in the sense of no reasoning whatsoever): This is largely true. The Apple paper isn't arguing that LRMs do nothing akin to reasoning. It's critiquing the depth, robustness, and generalizability of that reasoning, especially when compared to the elaborate "thinking" traces they produce.
2. "...they prove that probabilistic, pattern-based reasoning is not the same as formal, symbolic reasoning. It is a different kind of cognition.": This is a widely accepted and important distinction. The paper's findings (e.g., failure to consistently apply algorithms, struggles with compositional complexity beyond a certain point) are consistent with the idea that LLMs operate differently from classical symbolic AI or even from how humans might approach some formal reasoning tasks. The paper provides evidence for the limitations of this probabilistic approach in certain contexts.
3. "They are simply two different systems achieving a similar outcome through different means, each with its own strengths and failure points.": In a broad sense, comparing LLM "cognition" to human or symbolic AI cognition, this is true.

The Key Flaw(s):

1. The Bird vs. Airplane Analogy Misrepresents the "Illusion":

   • The statement says: "Calling it an "illusion" is like arguing that because a bird's flight mechanics are different from an airplane's, the bird is creating an "illusion of flight.""
   • This analogy is flawed because it misses the core of the paper's "illusion" argument.
     • A bird achieves flight. An airplane achieves flight. Neither is an "illusion" of the other, nor does a bird produce outputs that mimic the engineering principles of an airplane while actually being unable to sustain flight in the way an airplane does.
     • The "illusion" the Apple paper refers to is about the mismatch between the appearance of deep, step-by-step, human-like thinking (as suggested by the long Chain-of-Thought traces LRMs generate) and the actual brittleness, inefficiency, and scaling limitations of their problem-solving capabilities.
     • It's more like if a system produced incredibly detailed and complex architectural blueprints (the "thinking trace") for a 100-story building but consistently failed to actually construct a stable building beyond 10 stories (the "accuracy collapse"). The detailed blueprints might then create an "illusion" of profound architectural understanding and capability that isn't fully borne out.
     • The paper isn't saying probabilistic reasoning is an illusion because it's different from symbolic reasoning. It's saying the manifestation of LRM reasoning (the verbose "thinking") can create an illusion of greater capability than exists under stress or at scale.

2. Downplaying the "Why" Behind the "Illusion" Title:

   • By focusing only on "different kinds of cognition," the statement sidesteps the paper's specific observations about how current LRMs fail despite their elaborate reasoning traces: the accuracy collapse, the decline in effort at high complexity, the "overthinking" on simple tasks, and the failure to consistently use provided algorithms. These behaviors are what lead the authors to use the term "illusion" – the thinking process looks sophisticated, but the results often don't match that sophistication when pushed.

In essence, the statement correctly identifies that LLMs use a different "kind of cognition" (probabilistic, pattern-based) than formal symbolic systems. However, it fails to grasp that the "illusion" highlighted by the paper isn't about this difference per se, but about the deceptive appearance of depth and robustness in the reasoning processes of current LRMs when contrasted with their actual performance limitations.