What's This Actually About?

Understanding the Complex: How Does an LLM Work?

---

On the night of November 30, 2022, Sam Altman posted a tweet. "Today we launched ChatGPT. Try talking with it." No product announcement. No press release. Just a link.

The next morning, 10,000 people were using it.

Within five days, a million.

Within two months, a hundred million — the fastest user growth in internet history. Reuters analyst Krystal Hu documented the milestone in February 2023: no consumer product had ever hit that scale that quickly.

What happened that night wasn't just a product launch. Something shifted.

---

People who used it didn't just describe it as "a really good chatbot." They said it felt different. It felt like talking to someone who had read everything. A high school student asked it to explain the French Revolution in terms of a sports rivalry. It did. A nurse in Glasgow asked it to translate a patient's medical report into plain English. It did that too. A novelist in Tokyo asked it to write the next chapter of her book in the style of Haruki Murakami. It wrote something she genuinely couldn't tell apart from the real thing.

Within weeks, universities were panicking about plagiarism. The New York Times ran a front-page story in January 2023 documenting how professors across the US had already received AI-written essays. Law firms were debating whether they could use it to draft contracts. Radiologists were testing it on medical images. Programmers discovered they could describe what they wanted in plain English and it would write the code.

None of this should have been possible.

Not because the technology was magical. But because the company had trained the system to do one thing: predict the next word in a sequence. That's it. Predict the next word. Over and over.

That's the puzzle at the center of this course.

---

How does a machine trained to predict the next word end up:

• Writing essays that pass for human
• Explaining quantum physics to a ten-year-old
• Translating between languages it was never explicitly taught
• Solving math problems
• Writing code
• Knowing the capital of France

None of those capabilities were programmed in. None of them were designed. They emerged — from a single, almost absurdly simple task, run at an almost incomprehensible scale.

That emergence is what we're here to understand.

---

Three years on, nearly everyone in the developed world has encountered a system like ChatGPT, Claude, or Gemini. Millions use them daily. They're embedded in email clients, search engines, customer service systems, medical software, and legal tools.

And yet: if you asked most users how these systems actually work, you'd get one of three answers.

The first: "It's like Google, but better." Wrong. Google retrieves. These systems generate. Completely different.

The second: "It understands us. It thinks. Some people think it's conscious." Possibly partly true — and also not an explanation.

The third: "It's a probabilistic language model based on transformer architecture with self-attention mechanisms." Technically accurate. Explains nothing to most people.

All three answers leave you where you started: knowing that something significant is happening, not knowing what.

---

This course gives you a fourth option.

By the end of it, you'll be able to explain — mechanically, honestly, without magic or jargon — what happens between the moment you type a question and the moment an answer arrives. You'll understand who built these systems, why, and whose money made it possible. You'll know what the experts are actually fighting about. And you'll have frameworks for thinking about what comes next — not certainty, but calibrated thinking.

The central question we're answering:

What actually happens when you give a machine a question and it answers? And how much of that is understanding — and how much is very sophisticated guessing?

That question has two halves. The first is mechanical: what does the system compute? We can answer that. Not completely, but well enough that the magic disappears and something more interesting replaces it.

The second half is harder: does the system understand anything? Do these models know what they're saying? The honest answer is that the world's best researchers genuinely disagree. We'll show you the strongest arguments on both sides. Then you can decide.

One last thing before we start.

You probably have opinions about AI. Maybe you think it's going to save the world. Maybe you think it's going to destroy jobs, spread misinformation, or worse. Those opinions are welcome. This course won't try to replace them. It'll give you better tools to examine them.

Let's begin.

---

Next lesson: Why should I care? — Three reasons this topic is worth 150 minutes of your time.

---

Reading time: approx. 8–9 minutes