-http://zackmdavis.net/blog/2019/05/the-typical-set/
-> once you draw a boundary around a group, the mind starts trying to harvest similarities from the group. And unfortunately the human pattern-detectors seem to operate in such overdrive that we see patterns whether they're there or not; a weakly negative correlation can be mistaken for a strong positive one with a bit of selective memory.
-https://www.lesswrong.com/posts/veN86cBhoe7mBxXLk/categorizing-has-consequences
-[a higher-dimensional statistical regularity in the _conjunction_ of many variables](https://www.lesswrong.com/posts/cu7YY7WdgJBs3DpmJ/the-univariate-fallacy-1)
-96.8% classification from MRI https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6374327/
-]
-[the wrists: http://unremediatedgender.space/papers/yune_et_al-beyond_human_perception_sexual_dimorphism_in_hand_and_wrist_radiographs.pdf]
-[talk about mapping from one distribution to another: e.g. height]
+The typical point in a high-dimensional distribution is _not_ typical along each dimension individually. [In 100 flips of a biased coin](http://zackmdavis.net/blog/2019/05/the-typical-set/) that lands Heads 0.6 of the time, the _single_ most likely sequence is 100 Heads, but they're only one of those and you're _vanishingly_ unlikely to actually see it. The sequences you'll actually observe will have close to 60 Heads. Each such sequence is individually less probable than all Heads, but there are vastly more of them.
+
+Statistical sex differences are like flipping two different collections of coins with different biases, where the coins represent various traits. Learning the outcome of any individual flip, doesn't tell you which which set the coin came from, but if we look at the aggregation of many flips, we can get _godlike_ confidence as to which collection we're looking at.
+
+A single-variable measurement like height is like a single coin: unless the coin is _very_ biased, one flip doesn't tell you much. But there are lots of things about people for which it's not that they can't be measured, but that the measurements require _more than one number_.
+
+Take faces. People are [verifiably very good at recognizing sex from (hair covered, males clean-shaven) photographs of people's faces](/papers/bruce_et_al-sex_discrimination_how_do_we_tell.pdf) (96% accuracy, which is the equivalent of _d_ ≈ 3.5), but we don't have direct introspective access into what _specific_ features our brains are using to do it; we just look, and _somehow_ know. The differences are real, but it's not a matter of any single measurement: [covering up the nose makes people slower and slightly worse at sexing faces, but people don't do better than chance at guessing sex from photos of noses alone](/papers/roberts-bruce-feature_saliency_in_judging_the_sex_and_familiarity_of_faces.pdf).
+
+Notably, for _images_ of faces, we actually _do_ have magical transformation technology. AI techniques like [generative adversarial networks](https://arxiv.org/abs/1907.10786) and [autoencoders](https://towardsdatascience.com/generating-images-with-autoencoders-77fd3a8dd368) can learn the structure of the distribution of face photographs, and use that knowledge to [synthesize faces from scratch](https://thispersondoesnotexist.com/) and
+
+[...]