The title is a little click-baity, but the analysis of the paper in the blog is great. A fast download of one (quite hacky, fun) approach to getting scalable test-time compute.