How ChatGPT Actually Works (No PhD Required)

You've used ChatGPT. Maybe you've watched it write your emails, debug your code, explain complex topics. It feels like such a novel experience. But if someone asked you "how does it actually work?"—could you explain it?

Most people can't. Which is wild, because you're probably using it every day.

So what happened? Something fundamental shifted in how we interact with computers. Ten years ago, software couldn't understand you at all. You had to click the right buttons, fill out the right forms, navigate the right menus. Make a typo? Error message. The software was rigid, literal, unforgiving.

Then ChatGPT shows up and you can just... talk to it. Like a person. Be vague, change your mind mid-sentence, make typos—and it gets what you mean.

How did we get from "Click Submit to continue" to that?

The answer is surprisingly simple. Almost suspiciously simple. And understanding it is the key to actually using AI well.

For decades, all software followed the same approach. A programmer sat down and wrote explicit instructions—a recipe—for exactly what the computer should do in every possible situation.

If this happens → do that. If user clicks here → show this. If input equals X → return Y.

Building a calculator? Write rules for addition, subtraction, multiplication, division. Building a tax program? Write rules for every deduction, every form, every scenario.

This worked fine for math, databases, and accounting. The rules were clear. The logic was predictable.

But this approach completely falls apart when you try to build software that understands human communication.

The breakthrough came from Google's translation team.

In 2017, Google Translate was clunky, translating word-by-word, missing context. "Hit the road" became nonsense in Spanish. The problem? Still following recipes.

Then, Google researchers tried something radical: What if we built a machine that could learn to translate by absorbing patterns from millions of examples?

Think about how you learned language. Nobody gave you a grammar manual. You listened, absorbed patterns, made mistakes, got corrected. You became fluent without being able to explain how.

So these researchers built a "Transformer" that looked at entire sentences simultaneously, seeing how words related to each other.

Nobody realized they'd just invented the architecture that would power ChatGPT and kick off the entire AI revolution.

It was like switching from a recipe to a student. And that student was about to learn everything.

Before we explain how this student learns, you need to understand what it actually is.

Picture a gigantic music mixing board—the kind in a recording studio. Except instead of 12 knobs, this one has 100 billion knobs.

That's the model. That's the "AI." A stupidly large collection of adjustable settings in a computer file.

Here's the wild part: When we first create this model, all 100 billion knobs are set to random positions. No pattern. No intelligence.

Ask it "What is the capital of France?" and it outputs gibberish—random noise from randomly-set knobs.

Right now, it's just a very expensive random number generator.

We lock this giant mixing board in a library containing billions of books, articles, websites. And we force it to play Fill-in-the-Blank.

"The capital of France is ___."

Model guesses: "Sandwich."

Wrong. We tell it: "No, it's Paris."

Then we slightly adjust a few million knobs. Tiny twists toward making "Paris" more likely next time.

Again. "The sky is ___." Model: "Potato." Wrong. "It's blue." Twist knobs.

We do this billions of times.

Over weeks, those 100 billion knobs move from random positions into a specific configuration—a compressed, statistical map of how human language works.

Now when we ask "The capital of France is ___," those tuned knobs output: "Paris."

Now that the knobs are set, how does it respond when you use it?

Let's say you type: "Write a professional email declining a meeting because I'm overbooked this week."

The model doesn't have a template stored somewhere. Instead, it runs your words through those 100 billion knobs and gets probabilities for the first word:

It rolls a weighted die. Picks "Thank."

Now it calculates again for the next word after "Thank":

Picks "you." Now it's got "Thank you." Calculates again. Picks the next word. And on and on, building the email word by word.

The model isn't following a script or retrieving pre-written text. It's generating each word fresh, based on patterns it absorbed during training.

This is why it's a chef, not a baker. A baker follows a recipe exactly. A chef tastes and improvises.

Because there's randomness in that selection, you can ask the same question twice and get different responses. "Thank you for thinking of me..." versus "I appreciate the invitation, but..." Both professional. Both coherent. Both slightly different.

Here's where it gets strange.

We trained these models to predict the next word. That's it. Fill in blanks. We weren't trying to teach logic or reasoning.

But something bizarre happened: To get really good at predicting words, the model accidentally learned to reason.

Think about it: If someone writes "The patient had a fever, headache, and tested positive for—" you can only predict "flu" if you understand symptoms and disease patterns.

The model learned that—not because we taught it, but because predicting words well required it.

We built a sentence-completion engine and accidentally got a reasoning machine.

Researchers still don't fully understand how. We see these models reason, code, and translate. But we can't point to specific knobs and say "these do the reasoning." It's distributed across billions of parameters in ways we're still deciphering.

Once you understand the trick, you can see where it shines.

LLMs are extraordinary at pattern-heavy work. Drafting emails, reports, marketing copy. Explaining complex topics in plain language. Translating between styles or tones. Brainstorming variations. Summarizing long documents. Writing code that follows common conventions.

These are all tasks where "what would a good version typically look like?" is exactly the right question. The model has seen millions of examples and can remix them for your situation.

The people getting the most from AI use it for these tasks—then add their own judgment, facts, and finishing touches.

What to watch out for

The trick has a flaw.

LLMs don't know what's true. They know what's probable. Usually those overlap—accurate information appeared more often in training. But sometimes the most probable-sounding answer is completely wrong.

Ask an LLM to cite legal cases and it might invent ones that don't exist—because legal arguments typically include citations, so the model produces something that looks right. Ask for statistics and you might get outdated or fabricated numbers delivered with complete confidence.

This isn't a bug that will get fixed. It's baked into how the system works.

So: use AI to accelerate pattern-heavy work. Double-check anything where facts matter. Treat it less like an oracle, more like a sharp but overconfident colleague who sometimes needs correcting.

That's the real skill emerging—not clever prompts, but knowing when the world's most impressive autocomplete is the right tool for the job.

When an AI confidently states something that isn't true, that's a hallucination. It's not lying (there's no intent) and it's not a bug in the traditional sense. It's the model doing exactly what it was trained to do—predict the most plausible next words—except sometimes the most plausible-sounding answer is wrong.

Why does it happen? Because the model doesn't "know" facts the way you do. It knows patterns. If a question looks like it should have a confident answer, the model produces a confident answer—whether or not it's accurate.

Drop this at your next meeting: "The hallucination problem is fundamental to how LLMs work—they're optimizing for plausibility, not truth. That's why you always need human verification on anything that matters."

Stephen Wolfram's essay "What Is ChatGPT Doing... and Why Does It Work?" is the best plain-English explanation of the technical details. It's long but remarkably clear—Wolfram has a gift for making complex ideas feel almost obvious in hindsight. Worth saving for a weekend read.

That's Gist for this week. See you next week.