How LLMs Work

The Magic Trick

When you talk to ChatGPT, it feels like there's a person on the other side. It answers questions, tells jokes, and even writes poetry. But under the hood, it's doing something much simpler: Predicting the next word.

The "Autocomplete" on Steroids

You know how your phone tries to guess the next word when you're texting?

• You type: "I am going to the..."
• Phone suggests: "store", "park", "movies".

Large Language Models (LLMs) are just incredibly powerful versions of that.

They have read almost everything on the internet—books, websites, code. Because they've read so much, they know exactly which words usually follow each other.

If you say: "The capital of France is..."
The LLM knows, with 99.9% certainty, the next word is "Paris".

Tokens: The Language of AI

Computers don't read words like we do. They break text into little chunks called Tokens.

• "Apple" might be one token.
• "Ingenious" might be split into "In", "gen", "ious".

The AI processes these tokens as numbers. It's basically doing complex math to figure out which number comes next in the sequence.

Is it Thinking?

This is the big question.

• It looks like thinking: It can solve math problems and write essays.
• It's actually matching patterns: It's seen millions of math problems and essays before. It's remixing what it has learned to create something new.

It doesn't have feelings, beliefs, or a physical body. It's a mathematical model designed to be a helpful conversation partner.

Conclusion

LLMs aren't magic. They are "prediction machines" trained on a massive amount of text. They are great at sounding human, but remember: they are just guessing the next word, one at a time, very, very quickly.