Nostalgia thread: learning BERT's WordPiece took a whole month back in 2019

MaziyarPanahi · x · 2026-09-19

Responding to Hugging Face's old "BERTology" docs page, a researcher reminisces about the BERT era: it took him a full month in early 2019 to learn WordPiece tokenization while porting BERT to Java, and the community was still debating the role of [CLS] in classifiers until late 2020.

Related event: 'BERTology' Page Sparks Nostalgia for BERT's Golden Era(2 posts)→

Original post →

More from Fun

Fun channel →