I love the idea of diffusion language models. I think they are potentially superior to autoregressive models, in the way that practically, considering whole systems tends to beat optimizing any one individual detail. To me, the way autoregressive models are sampled feels very fundamentally limited, and diffusion feels much more coherent by comparison.
Comments
I love the idea of diffusion language models. I think they are potentially superior to autoregressive models, in the way that practically, considering whole systems tends to beat optimizing any one individual detail. To me, the way autoregressive models are sampled feels very fundamentally limited, and diffusion feels much more coherent by comparison.