As some readers may know, this write has long maintained a practice of writing book reviews to distill insights from complex ideas. Years ago, while reviewing a book on rationality (https://geninnov.ai/s/PbebLi), an inescapable conclusion emerged: rationality, at its core, is inherently irrational. It’s a Gödelian loop, an axiomatic truth that defies easy definition, a concept usually best left unexamined in polite company. In simpler terms, as my review concluded, “it does not pay to overthink rationality.”

When a research paper such as Apple’s 2025 study, “The Illusion of Thinking,” asserts that AI reasoning models fail under complex conditions, it prompts a revisitation of this perspective. Released just before Apple’s WWDC 2025, the paper coincides with an event that offered limited advancements in generative AI, raising questions about the company’s strategic direction. While the paper’s findings on AI limitations are partially valid, they reflect a broader mindset at Apple—one that prioritizes precision over adaptability. This mindset, akin to Albert Einstein’s insistence on understanding the rationality behind the uncertainties of quantum mechanics, may explain Apple’s cautious approach to generative AI, posing risks to its market position and offering lessons for investors.

Apple’s historic strengths—meticulous design, ecosystem control—have fueled its success, but now clash with AI’s inherent messiness. This article examines how Apple may be epic in its struggles, but it is far from alone, and the main consequence of the giants’ missteps is the relative stagnation of Edge Computing.

The Glass Half-Empty View of Reasoning Models

Apple’s paper, "The Illusion of Thinking", critiques LLMs for collapsing under problem complexity, overthinking easy tasks, and ignoring explicit solution templates.

The study finds that models perform well on simpler tasks but exhibit a “complete accuracy collapse” on complex versions, scoring 0% on challenges like the 7-disk Tower of Hanoi. It notes tendencies to overthink easy problems, fail to apply provided algorithms, and struggle with task variations, concluding that AI’s “reasoning” may be superficial pattern-matching.

It is undoubtedly true that general LLMs are far from perfect and require significant improvement. At its core is the "incomprehensibility issue". This term symbolizes our limited ability to understand exactly how these black-box models arrive work. Like with quantum mechanics, the incomprehensibility issue may not be resolved for a long time, if not forever, for a host of reasons, but it is a problem that demands continuous attention for scores of reasons. 

The situation echoes Einstein’s struggle with quantum physics. He famously resisted the probabilistic and counterintuitive nature of quantum mechanics, insisting that "God does not play dice." He grappled with epistemological issues, the fundamental meaning of particles and fields, while the rest of the physics community, sometimes simply opting to "shut up and calculate," propelled the science forward. Quantum mechanics, despite its inherent strangeness and unresolved philosophical questions, has delivered astounding technological progress over the past century.

Apple and many other proponents of theories like statistical parrots are in a similar bind, if not worse. Generative AI models present an even more complex challenge. They are, in a way, much like us humans: they can solve some of the most complex problems imaginable, like math Olympiad questions at expert levels, and yet might stumble on more elementary tasks, like simply counting the number of “r”s in "strawberry" or identifying basic shapes.

Crucially, not all models fail on all of these tasks. Indeed, there are models, or even refined versions of LLMs, that can perform some of the very tasks Apple’s paper cites as examples of "complete reasoning failure." The paper is correct that the "reasoning" observed in chain-of-thought models is not true reasoning in a human sense. Rather than waiting for an explanation or a structured output of their internal computational processes, critics may want to focus somewhat on what works and what is improving.

In other words, the paper is factually correct in detail, with the cherry-picked examples, but it does not provide justification for organizational inaction.

Table: It is not a straightforward good or bad scoresheet

Everyone Has Their Blind Spot: A Cross-Tech Snapshot

Apple is not alone. Here’s how other tech giants struggle with generative AI, each for reasons rooted in their legacy strengths:

We have written about the struggles of Apple (https://geninnov.ai/s/FEJTLc), Google (https://geninnov.ai/s/NcfdpU) and Micrsoft/Salesforce/Adobe (https://geninnov.ai/s/mquHyb) separately. 

Investment Implications: The Lagging Promise of Edge AI

It is good that these organizations are engaging in rigorous research and attempting to understand the limitations of AI. Such efforts are crucial for building more robust and reliable systems. However, when such intellectual pursuits translate into organizational intransigence and impede strategic goals, investors must take note. Apple’s protracted cautiousness is a prime example. It’s almost as if the company is wishing for the world to revert to a pre-generative AI era, hoping that the current advancements are merely a bad dream. But the reality is stark, even after WWDC.

For investors, a key theme that has significantly tempered our enthusiasm over the past year and a half, despite its earlier promise, is the rise of edge computing and a massive replacement cycle in smartphones and consumer devices driven by generative AI. We once held considerable faith in this narrative. Yet, over the last few quarters, we have actively reduced our focus on this specific theme.

Apple's recent event, coupled with Google's continued struggle to properly integrate smaller, powerful generative AI models onto Android phones, underscores a stark reality: edge computing, particularly in the consumer device space, is still significantly lagging in the generative AI era. It's a theme that undoubtedly holds tremendous long-term potential and will have its great days ahead. However, for now, we have another piece of evidence that this is unlikely to be a theme of the near future.