Before it
Vision belonged to convolutional networks; transformers were for text.
The problem
Two different architectures meant no shared representation between modalities.
What it introduced
Cut an image into patches, treat each as a token, and feed them to a plain Transformer encoder.
Why it matters
One architecture for pixels and words is the precondition for every multimodal model that followed.
What modern models inherit
Patches as tokens, the standard way images enter an LLM.
What came next
CLIP, Flamingo, LLaVA and the native multimodal frontier models.