As it has gotten longer since that nice-neighbourhood pathway initially opened, my refuge has become busier and buser. While the nebulous construction fence reclaimed silence for a while, those not willing to creatively jump their way by have mainly found that a few blocks deeper in there is another path that links up. The plethora of dogwalkers are very genteel but I am not particularly after passing pedestrian dog conversations. The edges of one mown section fringes into occasional fairy paths through the surrounding layer of Tradescantia luring incautious feet into the swamp beneath. One of these paths is wholly dry, except where trees have fallen on or slid through it. Occasional human and dog prints affirm the way. About 50 metres on there is 5 metres of clear sand/silt riverside beach, riven by a tiny rivulet down from 6 metres of cliff. The odd shape and scant primary succession imply a treefall into the river. Excuse me, I do go on. This week has been both busy and unfulfilling. I itinerantly teach in some small way. Lately I have been getting university alumni to migrate from Notepad++ on Microsoft Windows to Emacs orgmode on a liferaft. The university and town are absolutely owned by the Microsoft sales rep except for an oily sheen of Google acquisitions. People that find me come with a sense of being born behind enemy lines. Enjoining people to orgmode is against my personal current. A hulking monolith of writing, publishing, scheduling, tabulation, inline rendering of images and TeX ensconsing the tangle and weave of code. I wonder if Knuth is for or against this particular progeny of his. What I had been planning to scribble on in this resounding new code silence from me was constructive solid geometry (CSG). Since I first came upon it in the context of electromagnetic finite element analysis, I largely inherited Gmsh. Gmsh sits preferentially upon oce, being the lgpl product bait of yet another cad software company. I like that Gmsh furnishes a light C api rather than miring you in C++ idioms (cf cgal and its boost corollary). It is hard to say what the best choice of CSG is. There are a handful of 3D programs and their libraries that are GPL3+ rather than the commercial enticement LGPL. Counter to that is the temptation to - since I know particular algorithms and standard approaches - just defpackage myself a little lisp rather than introducing a hard dependency on ECL's sffi. These packages are the life's work of their major contributors though. Eventually I will settle down but for now I am a stone doomed to rolling. Pivoting a second time I deeply enjoyed Boris Shminke's post about deep learning automated proofs. Deep learning success story articles often present proofs of the nature and destination of their convergence, those being the controversial bits of deep learning*. Deep learning converges somewhere, eventually, and for any particular destination there is a way to converge there, but whether the place you got and time it took to get there was really that great can be dubious. *I think. I would like to pre-defer to Boris on the issue. I am from classical machine learning around receiver operating characteristic statistics for foreground segmentation algorithms (slightly tongue in cheek, what this means is that you come up with a completely arbitrary algorithm then search for / invent a new receiver operating characteristic metric that happens to make whatever your algorithm was look good. Two of my favourites from about 10 years ago were a Chinese research group who proposed some metric they didn't settle on, but tried something like standard ROC metrics within subsequent dilations of gold standard foreground segmentations and counting how the results changed. Less tantilizingly, I saw an article that claimed to have the highest true positive fraction foreground segmentation ever recorded! Despite its miserable false positive fraction. (If you just say 'true' no matter what your input was you can achieve this)).