Axe the ass tax

(Maybe if you had some super high dimensional complex class labels it would be good, idk where that’d come up tho)

You could even make a diffusion based classification model. It would be pretty inefficient but nothings stopping you and in principal it would work just as well

I’m a big diffusion booster. The principles behind diffusion generalize to any generation task (or really generically modelling any conditional probability). I wouldnt be surprised if we see more diffusion based reasoning or even agents in the future

Diffusion models make a lot more sense for code generation where often you want to add or modify tokens in place instead of just predicting changes autoregressively. Reasoning about blocks of code all at once instead of sequentially token by token is way more natural

This makes a lot of sense actually. The diffusion training protocol means that instead of learning purely autoregressive correlations it learns to predict words in context, which makes sense would generalize more because a lot of reasoning tasks aren’t linear

This doesn’t even come close to the most ridiculous ramen quest I’ve seen. I’m still searching for a good source for the mysterious Taiwanese wizard ramen I got as a gift for teaching someone Java

No AC in in half of it and it’s kinda a mish mash of a bunch of different expansions and renovations with no consistency