[FRIAM] privacy games

[uǝlƃ ☣]

Yes! Collaborative design (both of [un]intentional) could be contextualized with obscurity. (I think I've told the story of how my D&D group evolved into a collaborative fiction group after we all left home. We'd each take over authorship of each others' characters for a "letter" and send the whole packet down the line. Because each of us, as Dungeon Masters took various issues with players' silly characters, the collaborative fiction mode allowed us to do things like humiliate those characters when it was our turn to be author. Not quite the same as the blinding in the exquisite corpse, but a bit adversarial.)

And although I'm blown away by the things we (well someone, not me) can achieve with GAN, it still feels stilted to me ... a bit like the exploitation pitfalls in evolutionary computing (EC, e.g. negative altitude) or overfitted models. It brings to mind /procedural generation <e.g. https://github.com/mxgmn/WaveFunctionCollapse>/ as well. What I think EricS's idea of a multi-method constructing structure brings to the table is that collaboration can take many forms. And it maybe *must* take multiple forms in order to "round out" the composite probability distribution(s). A GAN (or EC) still seems a bit "flat" or "thin" in it's schematic guiding of a trajectory through the possible-needed space ("space" isn't the right word for such a self-constructing, dynamic thing, obviously). A minimal set of structures ... a kind of spanning basis for the collection of constructing/correcting mechanisms would be an ideal goal. And generation (the "G", what I've called Twitch) and discrimination are only 2 of them. Discrimination, in particular, seems ripe for a finer-grained, composite, implementation ... maybe that's why GANs still seem "thin" to me. But "adversarial" is also over-simplified. E.g. in the exquisite corpse (and our bad faith collaborative fiction), any one player's intention is not *entirely* adversarial, only slightly so. In the end all the players *want* some mix of cooperation, competition, syndication, and a sense of "fair play" ... as well as the ability to "game"/"cheese" it in bad faith sometimes.


Pixel-wasting story time: Renee' and I bought a truck on Monday. The finance people made a few errors during the transaction, but it was all resolved in seeming good faith. When I returned with my title for the trade-in, they pulled me aside and said they'd made another error in the contract and we needed to sign a new contract. Well, the new contract has me borrowing *more* money. I did a speedy ethical calculation deciding whether to "play hard ball" and argue that we had a contract and if they screwed themselves, too fscking bad. Nobody would fault me for that. But it's clear the dealership would get pinged by the umbrella (Ford) and whoever made the error would be pinged by their co-workers (maybe even docked pay -- they're not a commission dealership). And although the amount irritates me, it may well *hurt* whoever made the error. So my choice was a) play hard ball and stick to that (which they could also play hard ball on their side and since the relationship is asymmetric, they may well win) versus b) just signing the new contract and taking the hit "for the team". Since I'll be taking the truck into their service department as long as we live here, they're on my team in some sense. Essentially, I had the choice of adversary or in-group correcting collaborator. I chose the latter and gave them a bunch of sh¡t about how I didn't want to borrow that much and how my first big purchase here in WA (with a whopping sales tax -- I could have gone down to Oregon to do this) ... yadda yadda. These are *rich* collaborations that I'd like to see us constructively implement in our AI ... or better named ALife.


On 5/26/20 8:51 PM, Jon Zingale wrote:
> While walking to get a couple of bottles of wine today, I started thinking about
> collaborative games we play that rely on /privacy by obscurity/. The first image to
> pop into mind was the /exquisite corpse <https://www.artsy.net/article/artsy-editorial-explaining-exquisite-corpse-surrealist-drawing-game-die>/. Somehow I think it might be fruitful
> to think about the role various orders of privacy play in even our well-defined
> games. Referring back to our discussion of GANs <https://towardsdatascience.com/using-artificial-intelligence-to-create-people-cars-and-cats-5117189d0625>, I got to thinking about the
> role of privacy in producing more realistic images of cats than it seems possible
> with non-adversarial nets. Any thoughts?


-- 
☣ uǝlƃ