AI models that generate text can now run 30% faster

AI models that generate text can now run much faster. A new method lets these models skip unnecessary steps, cutting processing time by about 30% without losing accuracy.

Running large AI models is expensive. This paper shows how to make a specific type of generative AI model, Diffusion Language Models, significantly cheaper to operate. Companies using or building these models can now get the same results with less computing power, or generate more text for the same cost.