Now, is it possible that a model can combine advantages of both? Combine fast generation and multidirectional causality of diffusion with precision, capabilities and generalization of autoregression?

Maybe. This paper is research in that direction. So far, it's not a clear upgrade over autoregressive LLMs.