The gen­eral an­swer seems to be that re­li­gions (just like “to­tal” poli­ti­cal par­ties) provide value for money, in par­tic­u­lar a so­cial en­vi­ron­ment. Friends, baby sit­ters, group ac­tivi­ties, help when you lose a job or some­one dies. I think aca­demics, in par­tic­u­lar, tend to be such lon­ers, and to be con­tent with such so­cial sup­port as is pro­vided by the gov­ern­ment, that they rad­i­cally un­der­es­ti­mate how hun­gry peo­ple are for this sort of so­cial in­ter­ac­tion.

I haven’t no­ticed a prob­lem with this in my case. Might just not have no­ticed hav­ing this is­sue.

See that un­bi­ased “prior-free” es­ti­mates must be mix­tures of the (un­bi­ased) estimates

I don’t fol­low.

as­sem­bly of (higher-var­i­ance) estimates

What’s an “as­sem­bly of es­ti­mates”?

treat­ing the two as in­de­pen­dent pieces of evidence

But they’re dis­tri­bu­tions, not ob­ser­va­tions.

There’s a long his­tory of de­vel­op­ment of artis­tic con­cepts; for ex­am­ple, the dis­cov­ery of per­spec­tive. There’s also lots of artis­tic con­cepts where the de­pen­dence on the medium is highly sig­nifi­cant, of which the first ex­am­ples that come to mind are cave paint­ings (which were likely de­signed around be­ing viewed by mov­ing fire­light) and ‘highly dis­torted’ an­cient stat­uettes that are prob­a­bly self-por­traits.

So it seems to me like we should ex­pect there to be a similar medium-rele­vance for NN-gen­er­ated art­work, and a similar ques­tion of which artis­tic con­cepts it pos­sesses and lacks. Per­spec­tive, for ex­am­ple, seems pretty well cap­tured by GANs tar­get­ing the LSUN bed­room dataset or re­lated tasks. It seems like it’s not a sur­prise that NNs would be good at per­spec­tive com­pared to hu­mans, since there’s a cleaner in­verse be­tween the per­cep­tive and the cre­ation of per­spec­tive from the GAN’s point of view than the hu­man’s (who has to use their hands to make it, rather than their in­verted eyes).

That is, it seems that the similar­i­ties are ac­tu­ally pretty con­strained to bits where the medium means hu­mans and NN op­er­ate similarly, and the path of hu­man art gen­er­a­tion and the path of NN art gen­er­a­tion look quite differ­ent in a way that sug­gests there are core differ­ences. For ex­am­ple, I think most hu­mans have pretty good fa­cil­ity with cre­at­ing and un­der­stand­ing ‘stick figures’ that comes from train­ing on a his­tory of com­mu­ni­cat­ing with other hu­mans us­ing stick figures, rather than sim­ply gen­er­al­iz­ing from vi­sual image recog­ni­tion, and we might be able to demon­strate the differ­ences through a train­ing /​ fine­tun­ing scheme on some NN that can do both clas­sifi­ca­tion and gen­er­a­tion.

We might want to look for find con­cepts that are eas­ier for hu­mans than NNs; when I talk to peo­ple about ML-pro­duced mu­sic, they of­ten sug­gest that it’s hard to cap­ture the sort of de­pen­den­cies that make for good mu­sic us­ing cur­rent mod­els (in the same way that cur­rent mod­els have trou­ble mak­ing ‘good art’ that’s more than style trans­fer or re­al­is­tic faces or so on; it’s un­likely that we could hook a NN up to a De­vian­tArt ac­count and ac­cept com­mis­sions and make money). But as some­one who can listen to lots of Gan­dalf nod­ding to Jazz, and who thinks there are ap­pli­ca­tions for things like CoverBot (which would do the acous­tic equiv­a­lent of style trans­fer), my guess is that the near-term po­ten­tial for ML-pro­duced mu­sic is ac­tu­ally quite high.

The claim was that A/​​B test­ing was “not as good a tool for mea­sur­ing long term changes in be­hav­ior” and I’m say­ing that A/​​B test­ing is a very good tool for that pur­pose.

And the pa­per you linked showed that it wasn’t be­ing done for most of Google’s his­tory. If Google doesn’t do it, I would be doubt­ful if any­one, even a peer like Ama­zon, does. Is it such a good tool if no one uses it?

By 2013 they were cer­tainly already tak­ing into ac­count long-term value, even on mo­bile (which was pretty small un­til just around 2013). This sec­tion isn’t say­ing “we set the thresh­old for the num­ber of ads to run too high” but “we were able to use our long-term value mea­sure­ments to bet­ter figure out which ads not to run”.

Which is just an­other way of say­ing that be­fore then they hadn’t used their long-term value mea­sure­ments to figure out what thresh­old of ads to run be­fore. Whether 2015 or 2013, this is damn­ing. (As are, of course, the other ones I col­late, with the ex­cep­tion of Mozilla who don’t dare make an ex­plo­sive move like ship­ping ad­block­ers in­stalled by de­fault, so the VoI to them is min­i­mal.)

The re­sult which would have been ex­cul­pa­tory is if they said, “we ran an ex­tra-spe­cial long-term ex­per­i­ment to check we weren’t fuck­ing up any­thing, and it turns out that, thanks to all our ear­lier long-term ex­per­i­ments dat­ing back many years which were run on a reg­u­lar ba­sis as a mat­ter of course, we had already got­ten it about right! Phew! We don’t need to worry about it af­ter all. Turns out we hadn’t A/​B-tested our way into a user-hos­tile de­sign by us­ing wrong or short-sighted met­rics. Boy it sure would be bad if we had de­signed things so badly that sim­ply re­duc­ing ads could in­crease rev­enue so much.” But that is not what they said.

com­pa­nies that can keep their UI cur­rent can prob­a­bly, in gen­eral, make bet­ter soft­ware

To the con­trary: com­pa­nies that up­date their UI to be “cur­rent” prob­a­bly, in gen­eral, make worse soft­ware (and not only in virtue of the fact that the UI up­dates of­ten di­rectly make the soft­ware worse).

I’m gen­er­ally pretty ret­rogrouch, and do of­ten pre­fer older in­ter­faces (I live on the com­mand line, code in emacs, etc). But I also rec­og­nize that differ­ent in­ter­faces work well for differ­ent peo­ple …

Do they? It’s funny; I’ve seen this sort of sen­ti­ment quite a few times. It’s always ei­ther “well, ac­tu­ally, I like older UIs, but newer UIs work bet­ter [in un­speci­fied ways] for some peo­ple [but not me]”, or “I pre­fer newer UIs, be­cause they’re [vague hand­wav­ing about ‘mod­ern’, ‘cur­rent’, ‘clean’, ‘not out­dated’, etc.,]”. Much less fre­quent, some­how—to the point of be­ing al­most to­tally ab­sent from my ex­pe­rience—are sen­ti­ments along the lines of “I pre­fer mod­ern UIs, for the fol­low­ing spe­cific rea­sons; they are su­pe­rior to older UIs, which have the fol­low­ing spe­cific flaws (which mod­ern UIs lack)”.

That was how I in­ter­preted your sug­ges­tion that UX peo­ple start to fol­low a “change UIs only when func­tion­al­ity de­mands”. Any­one who tried to do the “re­spon­si­ble” thing would lose out to less re­spon­si­ble folks. Even if you got a large group of UX peo­ple to re­fuse work they con­sid­ered to be chang­ing UIs for fash­ion, com­pa­nies are in a much stronger po­si­tion since the bar­rier to en­try for UX work is rel­a­tively low.

But note that this ob­jec­tion es­sen­tially con­cedes the point: that the pres­sure to­ward “mod­ern­iza­tion” of UX de­sign is a Molochian race to the bot­tom.

The ren­der­ing en­g­ines of Chrome/​​Edge/​​Opera (Blink), Sa­fari (We­bKit), and Fire­fox (Gecko) are all open source and there are many pro­jects that wrap their own UI around a ren­der­ing en­g­ine. The amount of work is re­ally not that much, es­pe­cially on mo­bile (where iOS re­quires you to take this ap­proach).

[em­pha­sis mine]

I have a hard time be­liev­ing that you are se­ri­ous, here. I find this to be an ab­surd claim.

in large part be­cause of anti-com­pet­i­tive be­hav­ior and gen­eral shadi­ness on the part of Google

Not sure what you’re re­fer­ring to here?

Once again, it is difficult for me to be­lieve that you ac­tu­ally don’t know what I’m talk­ing about—you would have to have spent the last five years, at the very least, not pay­ing any at­ten­tion to de­vel­op­ments in web tech­nolo­gies. But if that’s so, then per­haps the in­fer­en­tial dis­tance be­tween us is too great.

Most com­pa­nies man­age to not run any of those long-term ex­per­i­ments and do things like over­load ads to get short-term rev­enue boosts at the cost of both user hap­piness and their own long-term bot­tom line.

The claim was that A/​B test­ing was “not as good a tool for mea­sur­ing long term changes in be­hav­ior” and I’m say­ing that A/​B test­ing is a very good tool for that pur­pose. That com­pa­nies gen­er­ally don’t do it I think is mostly a lack of long-term fo­cus, in­de­pen­dent of ex­per­i­ments. I’m sure Ama­zon does it.

Note that at the end of a pa­per pub­lished in 2015, for a com­pany which has been around for a while in the on­line ad busi­ness, let us say, they are shocked to re­al­ize they are run­ning way too many ads and can boost rev­enue by cut­ting ad load.

The pa­per was pub­lished in 2015, but de­scribes work on es­ti­mat­ing long-term value go­ing back to at least 2007. It sounds like you’re refer­ring to the end of sec­tion five, where they say “In 2013 we ran ex­per­i­ments that changed the ad load on mo­bile de­vices … This and similar ads blind­ness stud­ies led to a se­quence of launches that de­creased the search ad load on Google’s mo­bile traf­fic by 50%, re­sult­ing in dra­matic gains in user ex­pe­rience met­rics.” By 2013 they were cer­tainly already tak­ing into ac­count long-term value, even on mo­bile (which was pretty small un­til just around 2013). This sec­tion isn’t say­ing “we set the thresh­old for the num­ber of ads to run too high” but “we were able to use our long-term value mea­sure­ments to bet­ter figure out which ads not to run”. So I don’t think “if even Google can fuck that up for so long so badly” is a good read­ing of the pa­per.

Ads are the first, sec­ond, third, and last thing any on­line busi­ness will A/​B test, and if there’s time left over, maybe some­thing else will get tested.

I work in dis­play ads and I don’t think this is right. Where you see the most A/​B test­ing is in fun­nels. If you’re sel­l­ing some­thing the gains from op­ti­miz­ing the flow from “user ar­rives on your site” to “user finishes buy­ing the thing” are of­ten enor­mous, like >10x. Whereas with ads if you just stick AdSense or some­thing similar on your page you’re go­ing to be within, say, 60% of where you could be with a su­per com­pli­cated header bid­ding setup. And if you want to make more money with ads your time is bet­ter spent on ne­go­ti­at­ing di­rect deals with ad­ver­tisers than on A/​B test­ing. I dearly wish I could get pub­lish­ers to A/​B test their ad se­tups!

Have some hor­rible jar­gon: I spit out a ques­tion or topic and ask you for your NeMRIT, your Next Most Rele­vant In­ter­est­ing Take.

Either give your thoughts about the idea I pre­sented as you un­der­stand it, un­less that’s bor­ing, then give thoughts that in­ter­ests you that seem con­cep­tu­ally clos­est to the idea I brought up.

Ex­cel­lent post.

One over­sight I see of­ten in this space, and here, re­lates to a car­bon tax. It is stated that the rev­enue from a car­bon tax can be used to com­pen­sate peo­ple, es­pe­cially lower in­come peo­ple, for the in­creased cost of liv­ing re­sult­ing from the tax. The fatal prob­lem with this is that in a zero emis­sions world, there will be no emis­sions and there­fore no car­bon tax rev­enue.

Of course it may be pos­si­ble to com­pen­sate peo­ple via other means such as other taxes. But a car­bon tax is only re­quired be­cause it is oth­er­wise cheaper to emit car­bon. This means costs will go up over­all and that there will be a net loss (in the short term at least). There is no free lunch and some­one will have to pay.

I didn’t in­tro­duce Wikipe­dia as an ex­am­ple of a site with poor UI. I think it’s pretty good aside from, as I said, the line width is­sue. It’s also in a space that peo­ple have a lot of ex­pe­rience with: dis­play­ing tex­tual in­for­ma­tion to peo­ple. Wikipe­dia could likely benefit from some A/​B tests to op­ti­mize their page load times, but that’s all be­hind the scenes.

users look­ing at it will have a low im­pres­sion of it

Mis­tak­enly, of course. This is a well-at­tested prob­lem, and is fun­da­men­tal to this en­tire topic of dis­cus­sion.

I’m not sure that this is mis­taken: com­pa­nies that can keep their UI cur­rent can prob­a­bly, in gen­eral, make bet­ter soft­ware. This prob­a­bly only holds for large com­pa­nies: since small com­pa­nies face more of a choice of what to pri­ori­tize while large com­pa­nies that look like they’re from 2005 are more likely to be en­vi­ron­ments that can’t get any­thing done.

I’m gen­er­ally pretty ret­rogrouch, and do of­ten pre­fer older in­ter­faces (I live on the com­mand line, code in emacs, etc). But I also rec­og­nize that differ­ent in­ter­faces work well for differ­ent peo­ple and as more peo­ple start us­ing tech I get farther and farther from the norm.

you can’t make it go away just by unilat­er­ally stop­ping play­ing

I never said that you could.

That was how I in­ter­preted your sug­ges­tion that UX peo­ple start to fol­low a “change UIs only when func­tion­al­ity de­mands”. Any­one who tried to do the “re­spon­si­ble” thing would lose out to less re­spon­si­ble folks. Even if you got a large group of UX peo­ple to re­fuse work they con­sid­ered to be chang­ing UIs for fash­ion, com­pa­nies are in a much stronger po­si­tion since the bar­rier to en­try for UX work is rel­a­tively low.

how do I move to a browser with which I can effec­tively browse ev­ery web­site, but whose UI stays static? I can’t.

The ren­der­ing en­g­ines of Chrome/​Edge/​Opera (Blink), Sa­fari (We­bKit), and Fire­fox (Gecko) are all open source and there are many pro­jects that wrap their own UI around a ren­der­ing en­g­ine. The amount of work is re­ally not that much, es­pe­cially on mo­bile (where iOS re­quires you to take this ap­proach). If this was some­thing that many peo­ple cared about it would not be hard for open source pro­jects to take it on, or com­pa­nies to sell it. That no one is pri­ori­tiz­ing a UI-sta­ble browser re­ally is strong ev­i­dence that there’s not much de­mand.

in large part be­cause of anti-com­pet­i­tive be­hav­ior and gen­eral shadi­ness on the part of Google

Not sure what you’re refer­ring to here?

1. Some­thing like MVP. Don’t start by throw­ing a brand new game out there—even if you have the edge in the game, you have to get peo­ple to play it. Get­ting the stuff for a new game + ad­ver­tis­ing costs money. Test it out a lit­tle (small scale). If you lose money test­ing it*, you paid a lit­tle bit of money to find out you’d have lost a lot of money if you’d tried it out big. (More nat­u­rally—big com­pa­nies are at times known for stay­ing the same, with star­tups com­ing in with new ideas. If you copy ideas from other peo­ple that haven’t bankrupted them...)

This seems to as­sume the peo­ple who did the origi­na­tion here were cas­inos or ex­plicit en­trepreneurs, in­stead of peo­ple who started gam­bling in­for­mally and then started with some sense of which games had which pay­offs.

(Or rather, maybe you’re ex­plic­itly not as­sum­ing that and that’s your point. But the way I’d make the same point you seem to be mak­ing here is not “they op­er­ated like a startup” and more like “they op­er­ated like a group of friends/​ri­vals/​com­mu­ni­ties in­cre­men­tally ex­per­i­ment­ing, and by the time some­one con­sid­ered start­ing an ex­plicit busi­ness, good gam­blers had some in­tu­itive sense of how games worked.”)

Cu­rated.

Sys­tem 1 and 2 have got­ten a lot of at­ten­tion on LessWrong over the years, in­form­ing many ex­plicit mod­els as well as vague folk-the­o­ries cir­cu­lat­ing the com­mu­nity. Given that the origi­na­tor of the con­cept has de­p­re­cated it, it seems par­tic­u­larly im­por­tant for the LW com­mu­nity to have com­mon knowl­edge of the state of the liter­a­ture.

be­cause they hap­pened not to bankrupt cas­inos that ran them,

1. Some­thing like MVP. Don’t start by throw­ing a brand new game out there—even if you have the edge in the game, you have to get peo­ple to play it. Get­ting the stuff for a new game + ad­ver­tis­ing costs money. Test it out a lit­tle (small scale). If you lose money test­ing it*, you paid a lit­tle bit of money to find out you’d have lost a lot of money if you’d tried it out big. (More nat­u­rally—big com­pa­nies are at times known for stay­ing the same, with star­tups com­ing in with new ideas. If you copy ideas from other peo­ple that haven’t bankrupted them...)

2. It seems like it’s pos­si­ble to make it by on “this is un­likely” or set­ting things up so you always win. (I no­tice snake eyes doesn’t come up a lot. (Per­haps I check this by rol­ling dice a bunch of times.))

Sim­plest case: you buy a place, and pool equip­ment, and you rent it out to peo­ple. If they make bets with each other on the out­come, you don’t care—they’re just pay­ing you so they can play pool.

Slightly more com­pli­cated: you offer to han­dle the bet­ting on the game. Peo­ple pay you a lit­tle to be able to bet (and later, to win big), but the money all comes from them, and you don’t care who wins—you make money off peo­ple play­ing, peo­ple watch­ing, and peo­ple bet­ting!

Did cas­inos even know what they were do­ing, or did they some­what rou­tinely fail af­ter pick­ing the wrong game de­sign?

3. Were cas­inos a thing be­fore prob­a­bil­ity was un­der­stood?

*One game night with a few peo­ple, maybe you and your friends? If you have peo­ple who are happy to try out a new game, with­out real money, (For Free! per­haps?), that’s a place to start ini­tially—and all you lose is the time to run it. If you have fun, then maybe that’s a small price to pay. And if peo­ple are will­ing to pay to play a game with fake money, then you can just print more monopoly money if you run out—no odds calcu­la­tion needed for a sure bet.

I re­ally like the green-un­read on post pages. On Re­cent Dis­cus­sion I have so much of it that I think I don’t re­ally pay at­ten­tion to it.

Meta/​UI:

I cur­rently be­lieve it was a mis­take to add the “un­read green left-bor­der” to posts and com­ments in the Re­cent Dis­cus­sion sec­tion – it mostly makes me click a bunch of things to re­move the green that I didn’t re­ally want to mark as read. Cu­ri­ous if any­one has opinions about that.

And, as that pa­per in­ad­ver­tently demon­strates (among oth­ers, in­clud­ing my own A/​B test­ing), most com­pa­nies man­age to not run any of those long-term ex­per­i­ments and do things like over­load ads to get short-term rev­enue boosts at the cost of both user hap­piness and their own long-term bot­tom line.

That in­cludes Google: note that at the end of a pa­per pub­lished in 2015, for a com­pany which has been around for a while in the on­line ad busi­ness, let us say, they are shocked to re­al­ize they are run­ning way too many ads and can boost rev­enue by cut­ting ad load.

Ads are the core of Google’s busi­ness and the core of all A/​B test­ing as prac­ticed. Ads are the first, sec­ond, third, and last thing any on­line busi­ness will A/​B test, and if there’s time left over, maybe some­thing else will get tested. If even Google can fuck that up for so long so badly, what else are they fuck­ing up UI-wise? A for­tiori, what else is ev­ery­one else on­line fuck­ing up even worse?

users look­ing at it will have a low im­pres­sion of it

Mis­tak­enly, of course. This is a well-at­tested prob­lem, and is fun­da­men­tal to this en­tire topic of dis­cus­sion.

I don’t think fash­ion is the main mo­ti­va­tor here

No, the halo effect is the main mo­ti­va­tor.

you can’t make it go away just by unilat­er­ally stop­ping play­ing

I never said that you could. (Although, in fact, I will now say that you can do so to a much greater ex­tent than peo­ple usu­ally as­sume, though not, of course, com­pletely.)

The “ear­lier un­der­stand­ing” of many prob­lems in UX de­sign was more cor­rect. Knowl­edge and un­der­stand­ing in the in­dus­try has, in many cases, de­gen­er­ated, not im­proved.

How so? I can think of cases where ear­lier UX was a bet­ter fit for ex­pe­rienced users and newer UXes are “dumbed down”, is that what you mean?

In part. A full treat­ment of this ques­tion is be­yond the scope of a tan­gen­tial com­ment thread, though in­deed the ques­tion is wor­thy of a full treat­ment. I will have to de­cline to elab­o­rate for now.

If there was mas­sive value de­struc­tion then users would move to soft­ware that changed UI less.

In prac­tice this is of­ten im­pos­si­ble. For ex­am­ple, how do I move to a browser with which I can effec­tively browse ev­ery web­site, but whose UI stays static? I can’t (in large part be­cause of anti-com­pet­i­tive be­hav­ior and gen­eral shadi­ness on the part of Google, in part be­cause of other trends).

The fact is that such sim­plis­tic, spher­i­cal-cow mod­els of user be­hav­ior and sys­temic in­cen­tives fail to cap­ture a large num­ber and scope of “Molochian” dy­nam­ics in the tech in­dus­try (and the world at large).

If that’s the most se­vere (or one of the most se­vere) prob­lems with Wikipe­dia’s UI that you can think of, then this only proves my point. As you say, Wikipe­dia is gen­er­ally pretty good—which can­not be said for the over­whelming ma­jor­ity of mod­ern web­sites, even—es­pe­cially!—those that (quite cor­rectly and rea­son­ably) con­form to the “limit text column width” ty­po­graphic guideline.

-Ba­si­cally, since pre­dic­tions are laws of na­ture, wrong pre­dic­tions are mir­a­cles. We should ex­pect to them to ex­ist but to be rare.

Due to in­dex­i­cal un­cer­tainty, we can always be sur­prised by low prob­a­bil­ity events. I don’t see these as ev­i­dence of God though.

-Talk­ing about aliens. In­finite hy­poth­e­sis “God with such at­tributes ex­ists” can be used only as ap­prox­i­ma­tion (that is, ba­si­cally, our un­der­stand­ing of it). The finite hy­poth­e­sis “aliens fake us to be­lieve that God with such at­tributes ex­ists” also can be used only as ap­prox­i­ma­tion, (that is our un­der­stand­ing of God + as­sump­tion that it is faked by aliens). Thus such ap­prox­i­ma­tion is longer and should be given smaller prob­a­bil­ity.

Around in cir­cles again, but is there a differ­ence this time? Do we agree “fake alien God hy­poth­e­sis” dom­i­nates “in­finite God hy­poth­e­sis”? When us­ing in­duc­tion? You don’t seem to be dis­put­ing it. But is “ap­prox­i­mate God” sim­pler than “fake alien God”? That de­pends! How good is your ap­prox­i­ma­tion of “in­finite”? How com­plex are your aliens?

But if you want to ar­gue for a non-in­finite God, that’s OK with me, but even if you con­vince me, it won’t be the in­finite God you have con­vinced me of, but the finite ap­prox­i­ma­tion: Some­thing more pow­er­ful than mankind, but not in­finitely pow­er­ful. Some­thing more knowl­edge­able than mankind, but not in­finitely know­ing… this sounds like you’re de­scribing ad­vanced aliens. They’re the same thing. I would then ar­gue that the aliens are the re­al­ity and the “in­finite God” is the ap­prox­i­ma­tion of them made by ig­no­rant hu­mans.

Even I would be will­ing to call such aliens “gods” given cer­tain con­di­tions, but we’re us­ing your defi­ni­tion of “God”.

Can you con­vince me of ap­prox­i­mately-God aliens? Maybe. My prior is not zero, but like the pet pur­ple dragon from Mars, it would take a lot of ev­i­dence to con­vince me.