12h ago

Google DeepMind's Mathieu Blondel argues autoregressive models outperform diffusion for discrete sequences due to token-wise softmax independence limitations

Hardware scaling trends favor FLOPS over memory bandwidth

Sentiment

Pos100%

Neg0%

Users praise the autoregression versus diffusion debate for sharing more detailed info and free compute-saving insights than most arXiv papers.

5 comments with sentiment.

Google DeepMind's Mathieu Blondel argues autoregressive models outperform diffusion for discrete sequences due to token-wise softmax independence limitations · Digg