New AI text diffusion models break speed barriers by pulling words from noise
1 min read
Summary
Inception Labs has launched Mercury Coder, a AI model that uses text diffusion to generate text, which is faster than conventional language models.
Unlike regular language models that create text word by word, diffusion models like Mercury produce text by refining it from an initially masked state, resulting in whole responses being generated simultaneously.
Text diffusion models use a masking-based approach, beginning with fully obscured content, and gradually removing the “noise” to reveal the complete response.
The introduction of noise during the training phase allows text diffusion models to learn how to generate the most likely completion of partially obscured data.
Inception Labs’ approach allows the model to refine outputs and address mistakes by not being limited to considering only previously generated text.
This allows Mercury Coder to generate 1,000-plus tokens per second.
