What it means
The transformer, introduced in 2017, is the design nearly all current language models share. Its key move was abandoning the sequential processing that earlier architectures required. Previous approaches read text one word at a time, which made training slow and long-range connections weak.
A transformer processes an entire sequence at once, using attention to let every position look at every other position directly. That change made training parallelizable, which made it economical to train on far more data than before — and it is that scale, unlocked by the architecture, that produced modern capabilities.
The design turned out to be strikingly general. The same architecture now underpins image, audio, video and protein models. The paper's title, "Attention Is All You Need," has aged well.
Why it matters
The transformer is why the current era happened when it did. It also explains a persistent constraint: attention cost grows sharply with sequence length, which is why context windows were historically small and why extending them is a genuine engineering achievement rather than a settings change.
In practice
Mostly background knowledge, but it clarifies why context is expensive and why nearly all vendors' models behave similarly at a fundamental level — they are largely the same architecture trained differently.