I Think, Therefore I Fail

Apple's newly released research paper The Illusion of Thinking has opened a conversation about the technical limitations of artificial intelligence. While the paper is analytical, it also shows a deeper truth about the nature of machine thought.

Understanding Reasoning Models (LRMs)

The paper mainly focuses on large language models (LLMs) with so-called "reasoning" capabilities. Unlike traditional LLMs such as OpenAI's GPT-4o, these models are meant to solve tasks that require some level of reasoning. Mainly used within mathematics and coding, reasoning models use intermediate steps, rather than just the final answer, to arrive at a solution. This is meant to mimic human thought processes. OpenAI's o3 model is the most popular example of this.

The Illusion of Thinking

Apple’s paper shows that what looks to be thought in these models is often only a surface pattern.

When the task is simple, the model writes little and often gets it right. But sometimes it keeps going. It overcomplicates what was already solved. The extra steps introduce noise. Accuracy drops.

As complexity increases, the model writes more. It appears to reflect. It tries paths, revises them, looks again.

Then something changes. At a certain point, the chain shortens. The model gives an answer quickly, without exploring. Accuracy drops again. The structure folds in on itself. Not from the pressure of thought, but from the lack of it.

Mimicking Thought

Reasoning models are mirroring our own mistake. We built models to think like us, and what they reveal the fact that we stopped thinking long ago. We called it “reasoning,” but we meant recitation. Creative thought is more than a series of steps. One must challenge, accept and reject ideas, and then challenge them again.

But thinking has been limited. At school, we are taught to fear mistakes. That correct answers stem from following rules, not necessarily from understanding. We are trained to give the right answer, not to explore the question. Just like reasoning models; we learn to mimic thought, not to think.

Apple's study shows this clearly. When problems get harder, models do not think more, but less. The chains shrink. Reflection drops. Not due to memory constraint or timeouts, but because the model has learned the modern lesson of pretending to try, rather than actually trying.

In other words, this behavior is not artificial, but inherited. It reflects our own educational systems, which trains pupils to produce answers, not understanding.

Mind Without a Soul

It should not be surprising that modern philosophy has abandoned metaphysics. The Enlightenment, with its focus on sole reason and empirical evidence, has led to a view of the mind as a machine. This is the legacy of Descartes, who famously said, "I think, therefore I am."

But what is thinking, really? In our rush of operationalization, we have embraced a reductionist view of the mind. We have come to see it as a set of processes, rather than as a lived experience. This is the illusion of thinking that Apple’s paper reveals.

Materialism, in its philosophical sense, is not a science, but a metaphysical position that refuses to name itself. It states that matter is all there is, and that everything can be explained in terms of matter and its interactions. This view has led to a mechanistic understanding of the mind, where thought is reduced to mere computation.

But this fails to acknowledge what cannot arise from matter alone. Consciousness. Meaning. Love. Logic. These are not functions, nor outputs, nor can they emerge from noise, no matter how complex. They are preconditions; what make thought possible in the first place. A model can mimic the pattern, but not the subject. It can compute, but it cannot care, and therefore not reason. And yet we go on treating simulation as if it were being, thinking and reasoning as if they were the same.

We simply aren’t ready for artificial intelligence, for we still mistake calculation for mind.