Warp specialization is an abomination that should be killed and I'm glad this could be an alternative.

I hope they can minimize the bookkeeping costs because I don't see it gain traction in AI if it hurts big kernels performance.