Before it
Subword tokenizers assumed text was already split into words by whitespace.
The problem
That assumption fails for Japanese, Chinese, Thai and any language without spaces.
What it introduced
Treat raw text, spaces included, as the input and learn subwords directly from it, reversibly.
Why it matters
It made tokenization language-agnostic, which is why multilingual models are practical.
What modern models inherit
The tokenizer library much of the field still uses.
What came next
Used by T5, Llama and most non-English-first models.