We must teach ourselves to think before we teach machines

Today, the systems we have are called GAI — Generative AI. That is, generative artificial intelligence. Even this definition whispers what we should expect from it: To generate. Not to think.

At the core of GAI models are large language models (LLMs). These are systems that extract patterns from massive piles of text on the internet and then produce answers based on those patterns. Deep learning, transformer architectures, billions of parameters… But behind all this technical splendor lies a simple goal: To predict.

What comes as the next word?

How does this sentence continue?

What does a text look like according to this pattern?

That is it. No thinking. No questioning. Just patterns.

----

Apple slapped this reality in our faces with the paper titled The Illusion of Thinking, which it published just before WWDC. They set up a test environment showing that models like GPT, Claude, and Gemini do not "think."

Logic puzzles containing three levels of difficulty:

• Easy: The models excel.

• Medium: Slowing down begins.

• Hard: They all collapse.

But the issue is not just difficult questions. Even more striking is this:

Apple also provided these models with the solution algorithms.

In other words, they said, "solve it this way."

It still could not solve them.

The model that performed 100 moves in the Tower of Hanoi stumbled in 3 moves in the river crossing puzzle.

Because one was in the training set, and the other was not.

-----

What this experiment shows is simple but jarring:

Today's GAI systems still live within rote memorization.

They are very good at generating from memory — but they cannot cross the boundary.

And because of this, even though we perceive them as "intelligent," our expectations have actually changed.

We did not ask them to think. We asked them to generate.

Therefore, their current success is legitimate within their own limits.

But beyond that limit, it is still dark.

-----

A system that thinks, questions, abstracts, and goes beyond the pattern like a human…

In other words, AGI — Artificial General Intelligence — is still a research dream.

It is not yet visible on the horizon.

But this dream is even more valuable because it shows the shortcomings of today's models.

Because this study by Apple implies not only what today's artificial intelligence is not, but also what it needs to become.

-----

A good imitation can replace the real thing for a while.

But in edge cases, the real and the simulation diverge.

What we have today is generative.

But it does not think.

And a system that does not think cannot even pretend to be wrong.

It only goes silent when the pattern breaks.

Perhaps that is why true intelligence still resides in us.

And perhaps, while there is still time, we should teach ourselves to think before we teach machines.