noun · AI architecture
Transformer
also called transformer network
The blueprint almost every modern AI language model is built from. It is not a robot, it is a way of arranging layers.
A neural network architecture built from stacked attention and feed-forward layers, introduced in 2017 and used by most current language models.
A stack of repeated layers, each combining attention, which mixes information across tokens, with a feed-forward step that processes each token. It can be trained in parallel on huge text, which is a big reason models scaled up.
Real-life example
You read that two very different chat products both run on transformers. It does not mean they are the same. The word describes how the model is laid out inside, like saying two buildings both use steel frames. The data they were trained on and the care taken afterward still differ a lot.
Where the analogy breaks
-
It has nothing to do with robots or electrical parts. The name describes how the layers transform the input.
-
It is not the only possible design. Other architectures exist and new ones are tried, though transformers are dominant for language today.
-
Using a transformer does not make a model good. Quality depends heavily on data, size and training choices.