Today I’m going to bust two myths about LLMs.
The first one you’ve probably heard before.
The second one — chances are it’s going to surprise you.
Myth 1: LLMs think.
Ask an LLM what the tallest mountain is. It pauses.
Then it responds: “The tallest mountain in the world is Mount Everest…”
Looks like thinking. Feels like thinking. Right?
It isn’t.
What’s actually happening is this:
- First, the model takes your question, processes it through layers of neural networks, and predicts the single most likely next word.
- Then it takes everything — your question plus that word — and predicts the next one.
- Rinse and repeat. One new word at a time, until it predicts an EOS (end-of-sentence) token, then it stops.
That’s it. No *thinking*. No concepts. No understanding of what a mountain is.
Just extraordinarily sophisticated next-word prediction, done billions of times.
Myth 2: LLMs call tools.
As we move into the world of Agentic AI, the combination of LLM+Tool(s) is becoming commonplace.
You’ve probably seen tons of demos where an AI “searches the web” or “sends an email” or “books a meeting.”
It looks like the AI is reaching out and doing things.
It isn’t. Again.
What’s actually happening is this:
LLMs can only generate text.
So when a tool needs to be called, the LLM sends a message back to your code — essentially saying: “I need you to call this tool for me.”
It is your code that calls the tool 🤯
Once your code has the result of the tool-call. Your code reports back to the LLM in plain text with the result. Then the LLM continues.
That’s why it’s important to know how to write code that will accommodate tool-handling for your LLM.
Two myths. One pattern: LLMs always look like they’re doing more than they are.
Understanding what’s actually happening under the hood — that’s what separates builders who get stuck from builders who ship.
Build with clarity,
Kirill
P.S. Next time an AI demo wows you, ask yourself: what is the code actually doing here? The answer is usually more interesting than the magic.