Every article in this section starts with "large language model" — but before those, there's a simpler idea underneath. "Machine learning" is the foundation, and LLMs are one kind of machine learning. So what actually is it?
The normal way vs. the learning way
Normally, when you write a program, you spell out the rules. If you want a program that tells whether an email is spam, you write a list of clues: contains "winner", has a suspicious link, uses all caps. You code the rules, the computer follows them.
Machine learning flips that around. Instead of you writing the rules, you show the computer lots of examples — thousands of emails already labelled "spam" or "not spam" — and the computer figures out the rules by itself. You give it the answers, and it hunts for the patterns that connect them.
What "learning" actually means
A model is a program with a lot of adjustable numbers inside it. "Learning" is just the process of tweaking those numbers so the model's guesses get closer to the real answers. It works in a loop:
- Try: the model makes a guess on an example.
- Check: compare the guess to the correct answer — how far off was it?
- Adjust: nudge the numbers a little to be slightly less wrong next time.
- Repeat: do this millions of times over millions of examples.
After enough repetition, the model has learned the patterns — not because it was told "this is a rule", but because adjusting toward the answers is the only thing that made its mistakes shrink.
A useful mental model: training is just guess, measure the error, adjust — repeated so often that the errors stop shrinking. What's left is the pattern.
Training data: the real limit
A model is only as good as the examples it learned from. If you train a weather model on data from a desert only, it'll do poorly in a rainy climate. This is why the training data matters so much: the model can't know more than what its examples contained, and it can't tell the difference between a real pattern and a coincidence in the data.
Not all machine learning is the same
- Supervised: the examples come with correct answers (spam/not spam). This is the most common starting point.
- Unsupervised: the examples have no labels — the model looks for hidden groups or structure on its own.
- Reinforcement: the model learns by trying things in a system and getting rewards or penalties, like training a dog with treats.
LLMs are a form of supervised learning, where the "answer" to every example is just the next piece of text. But the idea — learn from examples, then predict — is the same machine learning idea.
Why it still gets things wrong
A model doesn't understand anything; it just generalizes from examples. So it can confidently match a new situation to a pattern it saw before, even when the match is wrong. That's why a model can misclassify a legitimate email as spam, or get a fact wrong. It's not thinking — it's extrapolating from what it learned.
The practical takeaway: machine learning is powerful when it learns from good, relevant, representative examples. When the data is thin or biased, the model's guesses will be too.
The one-sentence summary
Machine learning is a way of teaching computers by showing them lots of examples and letting them figure out the patterns themselves — through guess, measure, and adjust — instead of handing them a written list of rules.