How LLMs Actually Work

For the past 20+ days, I’ve been fully immersed in an AI Engineering course.
I’ve gone through AI engineering fundamentals, prompt engineering, model-related topics, and token generation concepts.

But all of that was truly put to the test when a friend casually asked me:

“How do these AI systems actually work?”

Honestly, I didn’t expect to struggle with the explanation.
I’d read about AI long before this course and thought I fully understood it.

So I started confidently:

When you send a prompt, it’s converted into tokens. Tokens are basically chunks of text. Each token becomes a number, the model analyzes how those numbers relate to each other using context it learned during training, and then it generates a response by predicting the next most likely token repeatedly.

That’s the textbook explanation.

It makes sense to me, but I could immediately tell it didn’t really land for him.
Especially the part about “predicting the next token”. It sounded abstract.

That’s when I paused and asked myself a more important question:

How do I explain this so that anyone can instantly get it?
That’s actually why I decided to write this blog post in the first place.

So I stepped back and stripped away all the AI jargon.

Forget “intelligence”.
Forget “thinking”.
Forget “understanding”.

At its core, an LLM is just a prediction machine.

Not a thinking machine.
Not a reasoning being.
Just a system that is extremely good at predicting what comes next.

Here’s the example that finally made it click for me.

If I say:

“On Sunday afternoon, nothing beats a hot plate of jollof rice and…”

You probably already completed it in your head.

Chicken.
Cold drink.
Fried plantain.

You didn’t “think” about it deeply.
You didn’t analyze it consciously.

Your brain simply predicted what usually comes next, based on experience.

That’s exactly what an LLM does.

The difference is scale.

Where humans learn patterns from daily life, conversations, and culture, LLMs learn patterns from massive amounts of text: books, articles, code, documentation, websites, and conversations.

So when you send a prompt, the model doesn’t understand it the way a human does.

It doesn’t know what jollof rice is.
It doesn’t know what Sunday means.
It doesn’t know what food tastes like.

What it sees is something more like this:

“When these tokens appear together, based on my training data, this token usually comes next.”

That’s what “predicting the next token” actually means.

The model looks at your input, calculates probabilities, and picks the most likely next token.
Then it adds that token to the text…
…and repeats the process.

Over and over.
Very fast.

Each new token becomes part of the context for predicting the next one.

That’s how a single prompt turns into a full paragraph, an explanation, or a conversation.

It feels intelligent because the predictions are:

• context-aware

• grammatically correct

• often surprisingly accurate

But under the hood, there’s no awareness, no intent, and no understanding.

Just:
patterns → probabilities → next token

Once I truly internalized this, a lot of things suddenly made sense.

Why models can sound confident and still be wrong.
Why hallucinations happen.
Why wording your prompt differently changes the output.
Why these systems feel smart but don’t actually know anything.

LLMs are not thinking.

They are guessing, incredibly well.

And that’s not a weakness.
That’s the entire design.

Understanding this shifted how I approach AI engineering completely.
Instead of treating models like intelligent agents, I now treat them like powerful statistical tools that need:

• good input

• clear constraints

• safety checks

• and human judgment

That mental shift, more than any technical concept, is what helped everything else click.

And if this explanation helped you even a little, then this blog has done its job.