How transformers actually work — without the buzz

The old way to process text was like reading a book page by page, left to right. Each word got a turn before the next one arrived. It worked — until the book got long enough that the beginning started to get forgotten.

Transformers throw out that line-by-line rule. Instead, they look at every word at once and ask: how much does each word matter to every other word? The word "it" in the third sentence might point back to the third word in the first sentence. The model figures out those connections without reading anything in order.

Think of a family group chat. Someone drops a photo and says "look at this." Half the replies reference that photo even though they don't repeat the word. That's the attention mechanism — the model learns which words pay attention to which other words, regardless of distance. The word "primos" near the end might still be talking to the word "tienda" from the beginning.

This is why the word "transformer" matters for transformers — it's the name of the architecture, not the machine. The word is also used for a car part that transfers power from the engine to the wheels. Same idea: something that moves what you need to where it's going.

You see transformers everywhere now — in the model that writes your emails, the one that translates Spanish to English, the image generator that turns your prompt into a picture. They all use the same core trick: look at everything at once, weight the connections, and let the important ones win.

What changed was scale. The math existed before, but it needed a lot of data and a lot of compute to actually work. Now that we've got both, the trick has become the engine.

Tip: when someone says a model is "attention-based," that's the transformer — words that pay attention to each other, not the other way around.