Lucas Beyer flags 'crazy term drift' as researchers call plain encoders tokenizers
giffmana · x · 2026-10-20
Lucas Beyer (giffmana) debates @vincesitzmann and @TimDarcet over the word "tokenizer": if a DINO-style encoder feeds latent-space generation, it's just an encoder — swap the generator's architecture from transformer to CNN and nobody would call it a tokenizer. He sees this as a crazy drift from the originally sensible VQ-VAE usage. @vincesitzmann notes that in his recent collaborations "tokenizer" is simply used as a synonym for "encoder for latent-space generative modeling".
Related event: Researchers debate whether encoders should be called tokenizers(3 posts)→
More from Research
- Sophont launches Holon-S1 to predict genes and biomarkers from pathology slides — JeffDean · 2026-10-21
- GATOR: generative agentic 3D object reconstruction from casual images — kwangmoo_yi · 2026-10-21
- Claudini: AI autoresearch discovers SOTA jailbreak algorithms with 80% ASR — maksym_andr · 2026-10-21
- Three JEPA variants, one question: how well must a machine predict to plan? — TheTuringPost · 2026-10-21
- COLM 2026 paper shows low-quality training data gives LLMs 'brain rot' — _vztu · 2026-10-21
- Stanford Virtual Embryo Challenge partners with Eigen Labs' open research platform Yukon — anshulkundaje · 2026-10-21