Large language models (LLMs) like the ones behind modern chatbots can feel like magic. They're not — they're a clever combination of a few ideas taken to an enormous scale. Here's the intuition, without a single equation.
It all comes down to prediction
At heart, an LLM does one thing: predict the next piece of text. Given "The capital of France is", it learns that "Paris" is a very likely continuation. Everything else — answering questions, writing code, translating — emerges from doing this prediction extraordinarily well.
Words become tokens
Models don't see words exactly as we do. Text is chopped into tokens — chunks that are often whole words but sometimes fragments ("un", "believ", "able"). Each token is turned into a list of numbers that captures something about its meaning.
Attention: reading in context
The breakthrough that made modern LLMs possible is called attention. As the model reads, each token can "look at" every other token and decide which ones matter for understanding it. In "The trophy didn't fit in the suitcase because it was too big," attention helps the model figure out that "it" refers to the trophy, not the suitcase.
This mechanism, stacked in many layers, lets the model build a rich, contextual understanding of a passage before predicting what comes next.
Training: learning from the world's text
During training, the model reads staggering amounts of text and repeatedly plays a fill-in-the-blank game. When it guesses wrong, its internal settings (billions of them) are nudged slightly. Repeat trillions of times and the model gradually encodes grammar, facts, reasoning patterns, and style.
Why it sometimes makes things up
Because the model predicts plausible text, it can produce confident-sounding statements that are wrong — so-called hallucinations. It isn't looking anything up; it's generating what a good answer tends to look like. That's why grounding techniques (giving the model real documents to work from) matter so much.
Fine-tuning and alignment
A raw model that just predicts text isn't very helpful or safe. A second stage — using human feedback — teaches it to follow instructions, be helpful, and refuse harmful requests. This is what turns a text-predictor into an assistant.
The takeaway
An LLM is a next-token predictor trained at massive scale, made context-aware by attention, and shaped by human feedback into something useful. Understanding this helps you use it well: give it context, verify facts, and treat it as a brilliant, occasionally overconfident collaborator.
