I don't argue with anything you say, except to note that for the frozen NN you mention, I can't predict what output Y will result from input X, unless I already tested input X. It's not like a deterministic algorithm where output = input + 3 for all input in {-MAX_INT ... MAX_INT-3}. I need an exhaustive map of allowed inputs to expected outputs for the NN case. And product management never gives me one ;)

There's an utterly rabid response for you. For the record, I'm white. Completely white, for at least the past few hundred years of fairly complete family tree tracing. Baltic white on my maternal side, and British Isles on my paternal. And I suspect I know a metric shitload more about my family (and racial) history than you know about, well, anything much really.

I'm not sure what you're asking here. The generally accepted view is that the human race - almost certainly of dark skin color - originated in Africa, and radiated outwards from there. Through glaciation and tectonics, various populations became isolated from one another. For whatever reason, the populations that stayed closer to home kept their melanin. The populations that wound up in colder climates lost theirs and became whiter of skin. And various mutations related to hair color and texture and eye color also occurred. Due to the aforementioned isolation, there was no interbreeding possible with the humans who remained closer to our origin point. So, divergent races that - after a few generations - didn't know anything about each other any more, until they developed long-distance sailing technology and started killing each other for fun and profit.

You completely misread what I wrote. I was stating that after a long period of the populations being separate (in fact, long enough for two initially identical populations to diverge enough that they have radically different average albedo), when white people met black people, it was a case of whites as invaders and blacks as the invaded. I don't recall any stories of clipper ships sailing from sub-Saharan Africa to raid Finland. It was entirely the other way. Fucking plate tectonics, man.

Well, I'm a QA guy for a company that makes software used in hospitals. So, to me "knowing the expected output for a given input" is closely correlated with "reliable" and "trustworthy", and so yes, I tend to refer to deterministic algorithms positively. I realize the pressures on an advertisement-selecting algorithm are not the same. However I *ALSO* think that the DNA signals Google (or rather Google's advertisers) would be looking for are the less ambiguous things that are more easily detected with a simple deterministic algorithm. Yes, broader conclusions can be reached by letting a neural let look at a bunch of stuff. But an oncology clinic is going to be more interested in yes/no answers to questions like "does the searcher have DNA markers A,B,C,D,E [which are easily searchable deterministically]", and won't actually ask - or will rather assume the answer to - the underlying statistical question "do markers A,B,C,D,E correlate strongly to a predisposition to conditions that we can charge money to treat"?

:D As an Australian (at least for the next few weeks), I'll also point you to the Walibri tribe, who putatively shake penises rather than shaking hands when they greet each other.

"Racism" (not the right word for it) CAN exist in algorithmic search results. Please see my other reply in this same branch of the thread.

See my other reply on this same branch. That link I gave is just one of several, and possibly not the best reference - just the best I could find in a quick, inadequately caffeinated search this morning. (I was originally reading about airline pricing algorithms when I came across the information I was summarizing there. That led down a rabbit hole that gave the specific searching "bonds" = "investments" if you're white, "bail bonds" if you're black, example).

The study, and the related hoopla (this is just one link), was designed to indicate that Google uses a non-obvious, racially skewed signal as an input to search and advertising results. Actually, they probably use more than one such signal. Oversimplified: if I am logged in as "Wilfred Fortescue-Smythe-Smythe III" and search for "bonds" I'll probably see advertising/results for investment vehicles. If I am logged in as "Taneisha Williams" and search for "bonds" I'll more likely see advertising/results for bail bonds. It's important to note that there is NO CHANCE WHATSOEVER that Google designs intentionally racist ad-ranking algorithms; they have too much to lose. What this phenomenon demonstrates is that complex, probably nondeterministic algorithms that sum a buttload of signals, which are designed to exploit demographic/psychographic characteristics to group users by willingness to purchase, can as an epiphenomenon amplify and expose demographic differences in purchasing behavior, some of which differences might cause political trouble with people in a given demographic who compare results with a different demographic. (This works both ways. Some of these "discriminatory" results give higher pricing on things to people in a "rich" demographic).

Participation in community is itself either an evolutionary pressure, or an evolutionary outcome - probably both. The need for language is a direct result of community life. If you don't live in a community, the only things you really need to communicate to others are 1) I'm ready for sexytimes, and 2) fuck off invader, this territory/mate/tree/cave/carcase is mine - both of which can be achieved with physical structures and/or pheremones, no need for language. And a huge amount of the human brain is dedicated to language processing.

Interesting question, though how the "training" occurs is up for debate. It's unlikely to be that, though. It's more likely to be keyed on specific, easily identifiable markers that can be parsed out with deterministic algorithms.

The reason they have this API is basically "because FDA". I don't know the *current* legal status, but at one point 23andme was essentially shut down by the FDA for providing unlicensed diagnostic services. Taking a spit sample and turning it into a written list of base pairs isn't diagnosis, so that part was still FDA-OK. It's comparison against known markers for diseases that's diagnosis. So they turned the "diagnosis" part of the business over to a licensed third party (or rather, they gave you the option of selecting the third party of your choice. Or simply using grep ;))

*Every* white person has "some proportion of non-white heritage", because the human species has pretty damn convincingly been shown to have evolved in Africa.

If everyone was a single homogeneous skin color, the exact same discriminations would be occurring on the basis of eye color, or eyelash color, or penis length. It just happens to be a big, easy to read signal differentiating "us" from "not us". (Actually I think the answer to "why skin color" has a lot to do with the fact that white races were isolated from black races for a long time after they, y'know, evolved from an originally black common ancestor, and when they did meet black people again it was in a conqueror/enslaver vs conquered/enslaved context).

1) Kuwait has introduced a law mandating DNA testing for everyone, already. https://www.hrw.org/news/2015/... - it's required in order to get a government ID, and a government ID is required to live, basically. 2) In America, at least, the real problem with this code (and likely the real use case) is not blocking access to a website as an expression of "racism" (whatever that even means in this context), it would be using that profile to serve up content _selectively_. It's already been shown that Google gives different results to searches that include "black" names vs "white" ones: http://thevisualcommunicationg... - they could with perfect ease include this DNA data as an input signal to ad selection also. Searching for auto insurance? Your color blindness will cause the search to show worse deals. It's axiomatic that you should never, ever make unchangeable information with abuse potential of this sort accessible to anyone, if it can possibly be prevented.