Loading slide
Within a few years, the transformer had pushed almost every other approach out. It is worth pausing on why, because it was not one advantage but several arriving together, most of which this module has already shown you.
It read across distance with no chain to fade through, so it held long passages the old way never could. It ran in parallel, so it drank up the GPUs that the going-deeper module showed had become cheap and plentiful. It stacked into deep towers, building meaning in stages the way every powerful thing in this course has. And it trained on raw text with no human labelling, predicting the next word, so the entire internet became its textbook for free.
Stack those up and the transformer was not just a bit better. It was better and faster and hungrier for data and able to use all the data and hardware the field suddenly had. Competing approaches could win on one of those at a time. None could win on all of them at once.
But the deepest reason came last, and it is the one that turned a strong architecture into the center of the whole field. The transformer kept getting better as you made it bigger. Add more floors, more words of training text, more compute, and it did not plateau the way earlier methods did. It just improved, and then it started doing things no one had built into it at all.
We have a name for that from the going-deeper module: emergence. New behaviour appearing simply from scale, the way wetness appears from enough water molecules though no single one is wet. We glimpsed it with vision networks. With transformers it became the entire story. What appeared when these models were scaled up beyond anything imagined, and why it unsettled even the people building them, is the whole of the next module.