I think that diffusion models might be a better way to generate code than autoregressive left-to-right generation