Every few months someone in my life asks me to explain how ChatGPT actually works. Usually it’s after they’ve seen it do something that felt a little too smart (write a cover letter, debug their code, explain a medical term better than their doctor did). They want to know: is this thing actually intelligent?

I’ve tried a lot of explanations. The one that sticks, every time, starts with your brain.

1. Your Brain Is a Network of Tiny Switches#

You’ve got about 86 billion neurons in there. Each one is a tiny cell with essentially one job: fire or don’t fire. That’s it. No neuron is smart on its own, it’s just a switch.

What makes your brain remarkable isn’t the neurons themselves, it’s the connections between them. When a neuron fires, it sends a signal across a tiny gap called a synapse to the neurons it’s connected to. Those connections aren’t all equal, some are strong and fast, practically screaming, others are so faint they barely register. And the key thing, the thing that makes all of this interesting, is that every time two neurons fire together, their connection gets a little bit stronger.

There’s a principle in neuroscience called Hebb’s rule: “neurons that fire together, wire together.” Donald Hebb wrote about it in 1949 and it’s held up pretty well. The idea is that learning is just reinforcement. Do something repeatedly, and your brain physically restructures itself to make it easier next time.

Think about learning to ride a bike. The first time you tried, your brain was a mess (random signals firing, guessing, correcting, falling over). Every attempt was your nervous system running a slightly better version of the experiment. Over days and weeks, the right connections got stronger and the useless ones faded. Eventually your neurons had built something like a highway. You stopped consciously thinking about balance. You just rode.

Learning isn’t installing a program. It’s paving a road.

2. Language Worked the Same Way#

Nobody taught you grammar. You were never sat down at age two with a textbook on sentence structure and told to get on with it. You just… heard things. All the time, for years.

“Milk.” You got milk. “More.” You got more. Thousands of sentences went into your brain with no explanation attached, and your neurons quietly started noticing: certain words keep showing up near each other, certain structures repeat, questions have a different shape than statements. You didn’t learn any of this consciously. You absorbed enough examples that the patterns burned themselves in.

By the time you were four or five, you could produce sentences you’d never heard before, and they came out grammatically correct. Not because you knew the rules. Because breaking them sounded wrong. That sense of wrongness is just your wiring telling you the pattern doesn’t fit.

Language, to your brain, is pattern recognition running so deep it feels like something else entirely.

3. You Weren’t the Only One Paying Attention#

Here’s where it gets interesting.

While your brain was doing all of this (wiring itself from exposure, building pattern highways through repetition), something else was doing the exact same thing. Just with incomparably more data and no body to put it in.

Large language models (the technology behind ChatGPT, Claude, Gemini, and most AI tools you’ve heard of lately) were trained on enormous amounts of text. Books, websites, academic papers, forum arguments, code repositories, Wikipedia articles. Hundreds of billions of words, processed over weeks of computation. And they learned through the same basic mechanism your brain uses: see a pattern, make a prediction, get feedback, adjust.

The thing doing the adjusting isn’t a neuron, it’s a number called a weight. Every connection in the model has one, and during training those numbers get nudged slightly up or down depending on whether the model’s guess was right. Strong connection, high weight. Weak or irrelevant connection, low weight. Exactly like synaptic strength.

The input isn’t processed word by word either. The model slices text into chunks called tokens (not quite full words, more like syllables or word fragments), which is surprisingly close to how your auditory cortex breaks down speech into phonemes before meaning kicks in. And it processes those tokens through dozens or hundreds of stacked layers, each one catching more abstract patterns than the last. Your visual cortex does the same thing: the first layer detects edges, the next detects shapes, eventually you’re recognizing faces. The model is just doing that with language instead.

Nobody programmed it to understand language. It absorbed patterns from exposure, the same way you did.

4. So What’s Actually Happening When You Talk to It#

You type something into ChatGPT. What does the model actually do?

It’s not thinking, not the way that word usually means. There’s no internal monologue, no deliberation, no moment of comprehension. What’s happening is prediction, the model looks at your tokens and, based on everything baked into its weights from training, picks the most likely next token. Then the next one. Then the next. One token at a time, the whole response gets assembled.

What makes this feel so unnervingly smart is a mechanism called attention. When you read “she sat by the bank watching the water,” your brain immediately knows which “bank” is meant, not because you consciously checked, but because you weighed the surrounding words automatically. Attention in language models works the same way: for each token, the model calculates which other tokens in the conversation are most relevant to it right now. That’s why it can hold a long conversation without losing the thread. It’s constantly recalculating context.

The result looks like understanding. Feels like it too. But what’s actually happening underneath is pattern completion, just at a scale and sophistication that makes the distinction easy to forget.

5. And Here’s Where the Analogy Breaks Down#

I want to be straight with you here, because I think the brain comparison is genuinely useful but can go too far if you let it.

Your brain lives in a body. You know what “cold” means not because you’ve read the word cold near the word winter a million times, but because you’ve been cold. You’ve felt it in your fingers. Every abstract concept you hold is anchored to physical memory, sensation, emotion. The model learned the word “cold” from text. That’s it. There’s no experience underneath, just co-occurrence statistics.

The thing that gets me is the why. You don’t just know that touching a hot stove is bad, you understand why, through a chain of knowledge that connects heat to tissue damage to pain to the reflex to pull back. LLMs know that certain tokens tend to follow other tokens. At the bottom of all that impressive output, that’s genuinely what it is. Correlation, stacked so high it looks like reasoning. It’s not the same.

There’s also the frozen thing, which surprises people. The model you used last month has the same weights it had when training ended. It won’t update tonight. It won’t incorporate what happened today. Your brain rewires a little while you sleep; the model just sits there with the same parameters it’s always had, until someone decides to retrain it from scratch. It’s not growing. It’s not learning from your conversations. It’s static.

Is it conscious? I genuinely don’t know, and I’m suspicious of anyone who answers that with total confidence in either direction. But the practical answer for how you should think about it is no. It has no inner experience. There’s nobody home, just a very sophisticated process that produces text shaped like thought.

Which, to be fair, is a remarkable thing to have built. Honestly kind of staggering when you think about how it works (all those weights, all that training, all those patterns absorbed from more human writing than any human will ever read). The result is a mirror of human language so good that it fools us constantly.

Just don’t mistake the mirror for a face. Appreciate the engineering. That part deserves appreciation.