Introducing Mercury

Inception's Mercury generates text through repeated refinement that can change several output positions in parallel. Its launch explanation contrasts this with the sequential process of generating one text unit after another. A neural network progressively improves the developing answer over successive passes.

Mercury Coder applies that approach to code generation. Inception's stated aim is to avoid the delay of sequential generation, since each refinement pass can work on multiple positions at once. The article's contribution is a concrete account of how diffusion, commonly associated with image, video and audio generation, can also produce text and code.