Continuous Diffusion Language Models (CDLM's)
SMRTR summary
Continuous diffusion language models (CDLMs) are making a comeback after being largely abandoned in 2023, when discrete methods took over due to better performance at scale. CDLMs are regaining interest mainly because they can be distilled into faster models, opening new possibilities for post-training and controlled text generation.
SMRTR provides this summary for quick context. The original article belongs to Hacker News.
Read the original article