"Run a fair AB test before claiming superiority To prove diffusion language models outperform autoregressive ones, Inception's team insisted on an extremely controlled comparison: 'the same exact neural network architecture with the same number of parameters, trained on the same amount of data,' differing only in the training objective. This isolated the signal so that 'the difference in performance is entirely due to the different modeling paradigm.' Before concluding that one approach, tool, or method beats another, verify that every other variable is held equal — otherwise you're measuring confounds, not the thing you care about. Source: Arash Vahdat, Latent Space Podcast, Diffusion LLMs with Inception AI"