Diffusion LMs denoise a canvas which I personally find more interesting. I don't really disagree that human cognition is essentially a predictive task though,

Diffusion LMs denoise a canvas which I personally find more interesting. I don't really disagree that human cognition is essentially a predictive task though, as I understand it, predictive coding and related theories based on the Bayesian brain hypothesis are fairly popular these days (though maybe not clearly dominant over alterative models? IDK I'm not a neuroscientist). I imagine most people would draft a few tokens before refining them like MTP or diffusion though, if we do decide to use LMs as an analogy to human cognition.