Thought Is All You Need
On minds, machines, and the strange habit of confusing one for the other.
“I think, therefore I am.” — René Descartes
Descartes reached for that line in 1637. He wanted one fact he couldn’t doubt away, and he found it in the act of thinking itself, not in any particular thought. Everything else might be illusion: the world, the body, other people. But the doubting itself proved something was there to do the doubting. Thought wasn’t just something a mind does. It was proof that a mind exists at all.
Three and a half centuries later, machines are doing something that looks dangerously like thought. They write articles. They output code, construct arguments, and synthesize research. Watching this happen, it’s easy to feel the ground shifting under our feet, like we are being quietly outpaced by something faster and more tireless than us.
But we aren’t being replaced. We are looking at an optimization problem and mistaking it for a mind.
The Math Disguised as Metaphor
To understand the illusion, you have to look at the 2017 Google research paper that triggered the current AI boom: “Attention Is All You Need.”
Before this paper, models processed text clumsily, one word at a time, losing the thread of long sentences. This architecture gave models “attention”—a mathematical mechanism for weighing an entire passage at once and calculating which words statistically mattered most to which others.
It is a landmark piece of engineering. But the name is a trap. We borrowed the language of human consciousness to label a mathematical operation, and now we are falling for our own vocabulary.
When you or I pay attention to something, the word carries an entire interior life inside it: focus, intent, care, the sense of someone behind the noticing. When a Large Language Model “attends” to a prompt, none of that is happening. It is running a staggering number of matrix multiplications, layering linear algebra to achieve exactly one goal: minimizing statistical error to predict the next token in a sequence.
It is not reading. It is calculating. And once you strip away the anthropomorphic vocabulary, the real question stops being how smart the machine is, and becomes what a mind does that a prediction engine never could.
The Mechanics of the Unprompted
The most defining characteristic of human cognition is that it cannot be turned off. We think because we are biologically forced to.
Strip away the sleek chat interface of an AI, and what remains is a function in a state of absolute null. It requires an input—a prompt. Take that away, and it does not sit in the dark quietly turning a problem over. There is no dark, and there is no sitting. It simply ceases to execute. It has no baseline state of existence.
A human mind cannot manage that kind of silence. Thoughts are generated by the friction of living in the physical world. Frustration arrives uninvited at a red light. A solution to a complex coding problem surfaces while you are trying to sleep. A child asks why the sky is blue without being prompted to wonder.
This spontaneous generation of thought is not a bug; it is the engine of human progress. The most ordinary thing a mind does is synthesize new ideas completely unprompted—something no machine has ever done, not once.
The Stakes of Memory
This reveals the other fundamental gap between an algorithm and a human: the concept of stakes. An algorithm does not experience the data it processes.
For a model, memory is a static matrix of weights. During training, every piece of text it consumes is flattened into numerical values based on frequency and proximity. It has no skin in the game.
Human memory, by contrast, plays favorites. It is aggressively biased by what matters to us. Think about how you process a conversation with a close friend versus a brief exchange with a stranger on the same day. To a machine, those are just two transcripts of equal weight. But your mind assigns them entirely different priority. You will carry your friend’s words with you for days, while the stranger is forgotten before the day is over. That uneven, emotional filtering is not a biological flaw. It is the exact mechanism that turns a series of events into a unique, unrepeatable perspective.
Crucially, that perspective is constantly shifting so If you ask a person the same question ten days apart, you might get two different answers. Not because their memory is faulty, but because in those ten days, they lived. They acquired new context, felt new pressures, and evolved.
If you ask a model the same question ten days apart, nothing has moved. It hasn’t lived through anything in between. It only runs the same numbers through the same frozen weights to arrive, with total mathematical confidence, at the same static answer.
The Origin of Intent
AI will inevitably speed up almost any workflow you can name. But it cannot replace the only part of the work that actually matters: the intent behind it.
Writing a script, drafting a legal brief, painting a canvas, or building a system—all of these tasks have a phase of pure execution. But before the execution happens, there is the noticing. Someone has to decide that a specific problem is worth solving. Someone has to feel that a specific story needs to be told. That desire comes from a specific history, a frustration, or a need that nobody handed to you in a prompt window.
A model can write the function or generate the image. What it cannot do is care whether it exists. It has studied human output to learn the probability of what comes next, which is entirely different from having a reason to speak in the first place.
Thought Is All You Need
None of this means AI is a passing fad. It is an extraordinary tool that will eliminate the friction between having an idea and bringing it into reality. It will handle the heavy lifting, the boilerplate, and the repetitive grinding that used to eat the hours of our days.
But execution is not conception. What a machine cannot do is ask the first question. It cannot look at the world, feel that something is missing, and decide that it matters enough to change.
When Descartes was looking for proof of his own existence, he didn’t find it in his ability to process information or recall facts. He found it in the undeniable, spontaneous act of thinking itself. The machines currently drafting our emails and compiling our code have inherited none of that. They have only ever had to prove they can convincingly predict the end of a sentence.
That has always been enough to be incredibly useful. It has never been enough to be a mind.