Dagon

• Dagon 27 Jan 2020 23:40 UTC

  8 points

  in reply to: Wei_Dai's comment on: Mo­ral pub­lic goods

  The pub­lic goods idea _does_ help ex­plain things if we think there’s a thresh­old is­sue (not valuable un­less a cer­tain amount is re­dis­tributed) _AND_ a co­or­di­na­tion prob­lem such that many peo­ple would like to donate, but only if they know the to­tal is over the thresh­old.

  It may also ex­plain things if the mo­ti­va­tion to not get full util­ity by an al­tru­ist donat­ing more is some form of pun­ish­ment for free rid­ers (those who don’t donate, but still get value).

  I agree that the more likely ex­pla­na­tion is that poverty al­tru­ism isn’t lin­ear (in util­ity with money donated) for any in­di­vi­d­ual, and most peo­ple are, in fact, giv­ing at the level they want to give. They would like to get some “free” util­ity by en­courag­ing/​forc­ing oth­ers to give more.

  This isn’t at odds with a pub­lic goods model—there are lots of pub­lic goods that go un-pro­vided be­cause they’re not worth it to enough peo­ple to provide it pri­vately or man­date it pub­li­cly. “this is a pub­lic good; there­fore gov­ern­ment must do it” is not a valid ar­gu­ment.

• Dagon 27 Jan 2020 19:58 UTC

  6 points

  on: What re­search has been done on the al­tru­is­tic im­pact of the usual good ac­tions?

  I sus­pect that you’re well into the mea­sure­ment-er­ror range for the things you’re talk­ing about. It would be silly to ex­pect a mea­surable change in QALY for some­thing you spend a tiny frac­tion of your wak­ing year on or a tiny frac­tion of your an­nual in­come on. Let alone the de­bate about how to ad­just the qual­ity mea­sure based on such things.

  For­tu­nately, at these scales, you can use anec­do­tal ev­i­dence of im­prove­ment in YOUR ex­pe­rience, for many things. Your friend’s smile or your co-founder’s con­tinued stream of hor­rible ideas are plenty of re­ward for the low cost of the kind­ness you’re con­sid­er­ing.

• Dagon 27 Jan 2020 18:22 UTC

  5 points

  in reply to: benwr's comment on: Is the Rev­er­sal Test over­rated?

  “helping to de­cide where to find the in­ter­est­ing bits” is ex­actly where this tech­nique shines, and I don’t think it’s over­rated (at least in my cir­cles).

  Note that even in the “yes, this is the right level, for X rea­sons”, there’s still a bunch of value in iden­ti­fy­ing the forces at equil­ibrium that make this the right level. You can then ask “do you want to change some of THOSE val­ues”?

• Dagon 27 Jan 2020 18:13 UTC

  2 points

  in reply to: edoarad's comment on: He­donic asymmetries

  I sus­pect it all comes down to mod­el­ing of out­come dis­tri­bu­tions. If there’s a nar­row path to suc­cess, then both bi­ases are harm­ful. If there are a lot of ways to win, and a few dis­asters, then op­ti­mism bias is very harm­ful, as it makes the agent not loss-averse enough. If there are a lot of ways to win a lit­tle, and few ways to win a lot, then pes­simism bias is likely to miss the big wins, as it’s try­ing to avoid minor losses.

  I’d re­ally en­joy an anal­y­sis fo­cused on your con­di­tions (max­i­mize vs satis­fice, world sym­me­try) - es­pe­cially what kinds of wor­lds and bi­ased pre­dic­tors lead satis­fic­ing to get bet­ter out­comes than op­ti­miz­ing.

• Dagon 27 Jan 2020 18:00 UTC

  4 points

  in reply to: G Gordon Worley III's comment on: G Gor­don Wor­ley III’s Shortform

  [up­voted for talk­ing about some­thing that’s difficult to model and com­mu­ni­cate about]

  Hmm. I be­lieve (with fairly high con­fi­dence—it would take a big sur­prise to shift me) a com­bi­na­tion of empty and closed. Mo­ments of self-ob­served ex­pe­rience are stan­dalone, and wo­ven into a fabric of mem­o­ries in a closed, un-sharable sys­tem that will (sooner than I pre­fer) phys­i­cally de­grade into non-ex­pe­rienc­ing com­po­nents.

  I haven’t found any­one who claims to be open AND is ra­tio­nal enough to con­vince me they’re not just mis­stat­ing what they ac­tu­ally ex­pe­rience. In fact, I’d love to hear some­one talk about what it means to “want” some­thing if you’re ex­pe­rienc­ing all things si­mul­ta­neously.

  I’m quite sym­pa­thetic to the ar­gu­ment that it is what it is, and there’s no rea­son to be sad. But I’m also un­sure whether or why my ac­cep­tance of closed-empty ex­is­tence makes you sad. Pre­sum­ably, if your con­scious­ness in­cludes me, you know I’m not par­tic­u­larly sad over­all (I cer­tainly ex­pe­rience pain and frus­tra­tion, but also joy and op­ti­mistic an­ti­ci­pa­tion, in a bal­ance that seems ac­cept­able).

• Dagon 27 Jan 2020 17:38 UTC

  0 points

  in reply to: DanielFilan's comment on: DanielFilan’s Short­form Feed

  It’s not clear that pos­i­tive-sum in­no­va­tion is lin­ear (or even mono­ton­i­cally pos­i­tive) with to­tal pop­u­la­tion. There al­most cer­tainly ex­ist lev­els at which marginal mouths to feed drive un­pleas­ant and non-pro­duc­tive be­hav­iors more than they do the growth-driv­ing shared in­no­va­tions.

  Whether we’re in a down­ward-slop­ing por­tion of the curve, and whether it slopes up again in the next few gen­er­a­tions, are both de­bat­able. And they should be de­bated.

• Dagon 27 Jan 2020 17:29 UTC

  2 points

  on: What hap­pens if we re­verse New­comb’s Para­dox and re­place it with two nega­tive sums? Doesn’t it kinda maybe af­firm Roko’s Basilisk?

  Wait—did some­one ac­tu­ally show that Roko’s Basilisk is stupid and dumb? I be­lieve that—it’s roughly a bad Pas­cal’s wa­ger, but with a causal hook. But I think the rea­son dis­cus­sion was banned was that it was caus­ing se­vere dis­com­fort, not that it was agreed to be harm­less and mean­ingless.

  In any case, your pro­posal does not have the causal hook that makes the Basilisk so tempt­ing and un­pleas­ant. It doesn’t black­mail you into cre­at­ing the black­mailer.

• Dagon 27 Jan 2020 4:29 UTC

  3 points

  in reply to: paulfchristiano's comment on: Mo­ral pub­lic goods

  Sure, it’s con­sis­tent to pre­fer non-you poor peo­ple get food over non-you rich peo­ple get­ting iphones. But most ac­tual peo­ple pre­fer THEY get an iphone over feed­ing any spe­cific poor per­son. Peo­ple aren’t fun­gible, and no ac­tual hu­mans are fully in­differ­ent to which hu­mans are helped or harmed.

• Dagon 26 Jan 2020 18:19 UTC

  4 points

  on: He­donic asymmetries

  He­donic adap­ta­tion (feel­ing re­ward/​penalty for the rel­a­tive change, much more than the ab­solute situ­a­tion) may be a key strat­egy for this. It ad­justs both up­ward and down­ward, to avoid ei­ther mis­take for very long.

• Dagon 26 Jan 2020 17:13 UTC

  2 points

  in reply to: Viliam's comment on: How Doomed are Large Or­ga­ni­za­tions?

  maybe some­times things suck be­cause there are more peo­ple, but some­times things only suck be­cause mazes have the power to change the law to make things suck.

  We’re in com­plete agree­ment. I’m look­ing for the model that tells me how to know which (or what pro­por­tion) of these is true for ac­tual mazes to­day.

• Dagon 26 Jan 2020 15:27 UTC

  −1 points

  on: Mo­ral pub­lic goods

  The first part of this turned seemed like mostly poli­tics—over­sim­ple and flat-out non-real ex­am­ple be­ing used to jus­tify a policy with­out any nu­ance or sense. Point 1 is just un­sup­ported and hard to ar­gue for or against, other than by say­ing your ex­am­ple is wrong and doesn’t jus­tify any spe­cific type or level of re­dis­tri­bu­tion, and you haven’t speci­fied even what “re­dis­tri­bu­tion” means, es­pe­cially in a dy­namic equil­ibrium where wealth and in­come are re­lated but dis­tinct.

  Point 2 is com­pletely miss­ing the fun­da­men­tal ques­tion of what peo­ple want—Fried­man’s point that if peo­ple ac­tu­ally were self-al­igned that they care about feed­ing spe­cific poor peo­ple rather than get­ting a new iPhone, they’d do it. In­stead, they want ab­stract poor peo­ple to get fed, and only if they can force oth­ers to do so (along with them­selves, in many cases, but rarely just unilat­er­ally). You don’t ad­dress this dis­par­ity.

  Point 3 is ac­tu­ally a rea­son­able start to lay­ing out the fun­da­men­tal puz­zle of large-group be­hav­iors. I’ll say that I am a con­se­quen­tial­ist, and that I con­sider my­self some­what al­tru­is­tic, but not an Altru­ist with a cap­i­tal A. And I’m in the first group: I con­sider my­self many or­ders of mag­ni­tude more im­por­tant (to me) than very dis­tant strangers. Not zero, but for some of ’em, more than tens of mil­lions of times more im­por­tant. For some, much less dis­count, maybe half as im­port for very close friends and rel­a­tives.

  Other peo­ple have a de­clin­ing marginal util­ity for me, and for a given level of re­sources, it CAN go nega­tive. There are al­most 8 billion of them cur­rently, and I think that’s prob­a­bly more than I pre­fer while we’re still at to­day’s tech level and limited to one planet. I don’t know if con­cen­tra­tions of re­sources are nec­es­sary to large-scale en­deav­ors, but I sus­pect it, and I don’t worry too much about it.

• Dagon 24 Jan 2020 17:34 UTC

  2 points

  in reply to: Donald Hobson's comment on: Don­ald Hob­son’s Shortform

  What? In this ex­am­ple, the prob­lem is not Carl—he’s harm­less, and Dave car­ries on with the cy­cle (of im­prov­ing the de­sign) as he should. Show­ing a situ­a­tion where Carl’s sen­sa­tion­al­ist mis­state­ment ac­tu­ally stops progress would likely also show that the prob­lem isn’t Carl—it’s EITHER the peo­ple who listen to Carl and in­terfere with Alice, Bob, and Dave, OR it’s Alice and Dave for let­ting Carl dis­cour­age them rather than un­der­stand­ing Bob’s ob­jec­tion di­rectly.

  Your de­scrip­tion im­plies that the prob­lem is some­thing else—that Carl is some­how pre­vent­ing Dave from tak­ing Bob’s anal­y­sis into con­sid­er­a­tion, but your ex­am­ple doesn’t show that, and I’m not sure how it’s in­tended to.

  In the ac­tual world, there’s LOTS of sen­sa­tion­al­ist bad re­port­ing of failures (and of ex­tremely minor suc­cesses, for that mat­ter). And those peo­ple who are ac­tu­ally try­ing to build things mostly ig­nore it, in fa­vor of more rea­son­able pub­li­ca­tion and dis­cus­sion of the un­der­ly­ing ex­per­i­ments/​failures/​calcu­la­tions.

• Dagon 24 Jan 2020 0:56 UTC

  10 points

  in reply to: Viliam's comment on: How Doomed are Large Or­ga­ni­za­tions?

  be­ware the un­stated al­ter­na­tive: is there rea­son to be­lieve that un-co­or­di­nated in­di­vi­d­u­als grow in power lin­early (or any faster/​slower than cor­po­rate/​gov­ern­ment ag­gre­ga­tions)?

  If this power calcu­la­tion holds, then things suck be­cause there are more peo­ple, not be­cause of how they’re or­ga­nized. I’d call that con­clu­sion fairly re­pug­nant.

• Dagon 23 Jan 2020 19:55 UTC

  2 points

  on: Terms & liter­a­ture for pur­posely lossy communication

  Not sure there’s a gen­eral term for it, but “psy­choa­cous­tic com­pres­sion” is the term for mod­el­ing the im­por­tance of in­for­ma­tion in lossy au­dio en­cod­ing such as MP3.

• Dagon 23 Jan 2020 19:48 UTC

  3 points

  on: To­kenis­ing hu­man ver­ifi­ca­tion in or­der to de­rive in­for­ma­tion from the re­sult­ing markets

  I’d ar­gue that “per­son­hood” is rarely what these things ac­tu­ally care about—it’s just a cheap-to-mea­sure proxy for “like­li­hood of con­ver­sion to sale” or “amount I’d get paid for an ad” or the like. A bot that can en­ter into con­tracts and is more likely than a real per­son to make a pur­chase would be wel­comed, but there are few of them and there’s no good test of it.

  For ac­tu­ally valuable things, a bot could just pay hu­mans to pass the captcha and all would be well. Shadier bots could man-in-the-mid­dle pretty eas­ily if they just pass through a captcha on their cat pic­ture site.

  For im­ple­men­ta­tion, it’s worth look­ing at the OAuth specs and com­mon fed­er­ated au­then­ti­ca­tion sys­tems that google, face­book and a num­ber of other sites provide—those do NOT as­sert hu­man-ness, they as­sert au­then­ti­cated ac­count iden­tity, but for most uses, that’s a bet­ter proxy any­way. In cases where it’s not, you could build a provider that uses OAuth to as­sert hu­man­ity us­ing what­ever ver­ifi­ca­tion it likes.

• Dagon 23 Jan 2020 18:28 UTC

  4 points

  in reply to: johnswentworth's comment on: (A → B) → A in Causal DAGs

  In­ter­est­ing take. When I see “agenty” used on this site and re­lated blogs, it usu­ally seems to map to some­thing like self-ac­tu­al­iza­tion or per­cieved lo­cus of con­trol, more psy­cholog­i­cal frame­works. I’d not thought too much about how differ­ent (or similar) it was to “agent” in de­ci­sion the­ory and game-the­o­ret­i­cal us­age, which is not about the feel­ing of con­trol, but about be­hav­ior se­lec­tion ac­cord­ing to leg­ible rea­son­ing.

• Dagon 23 Jan 2020 16:43 UTC

  2 points

  in reply to: johnswentworth's comment on: (A → B) → A in Causal DAGs

  I think there’s a con­tra­dic­tion here. Ideal­ized game-the­o­retic agents are the OPPOSITE of what we call “agenty be­hav­ior”. Ex­e­cut­ing a know­able strat­egy is the triv­ial part. Agenty-ness in the hu­man ra­tio­nal­ist sense is about cases where the model is too com­pli­cated to re­flec­tively know or an­a­lyze for­mally.

• Dagon 23 Jan 2020 0:54 UTC

  2 points

  in reply to: crl826's comment on: How Doomed are Large Or­ga­ni­za­tions?

  This is a new as­ser­tion—mazes only oc­cur in mo­nop­o­lies? And I guess the an­swer for why peo­ple would par­ti­ci­pate in the maze is that they only hap­pen in la­bor monop­sony con­di­tions? It’s pos­si­ble, in which case the solu­tion is sim­pler (to state; not always to do): break up the monopoly. I don’t think that’s what Zvi and oth­ers are claiming, though (ex­cept maybe in the fi­nance in­dus­try, which may be an effec­tive monopoly on em­ploy­ment: there are no op­tions which aren’t mazes), and it doesn’t match my ex­pe­riences or sec­ond-hand sto­ries of ac­quain­tances close enough that I’ve got­ten de­tails. Even in cases where it _is_ cur­rently a monopoly, you have to an­swer WHY there are no com­pet­ing op­tions to do it bet­ter and more pleas­antly at the same time. (note: if pressed, I will ad­mit that this para­graph was writ­ten mostly for me to in­tro­duce the phrase “cul­tural monop­sony”).

  Oh, wait—you said “if mazes are in­evitable”. They’re not uni­ver­sal to­day. I don’t know about even­tual in­evita­bil­ity, but there are large or­ga­ni­za­tions that are not en­tirely maze-like, at least not to the de­gree de­scribed in this se­ries. I have in­di­rect ex­pe­rience (not my­self, but rel­a­tively close friends and/​or rel­a­tives) with GM, IBM, and the US Navy, and none are all that bad for mid­dle man­agers—there’s poli­tics, but there’s also ac­tual pro­duc­tion and re­ward­ing work im­pact.

  I don’t think I’d claim that “good Moloch” ex­ists or is pos­si­ble. I make the much weaker claim that Moloch hasn’t ac­tu­ally op­ti­mized very far, so you CAN beat ‘em and don’t have to join ’em. For some time, at least—per­haps decades or gen­er­a­tions. I re­ally have no pre­dic­tion about the long-term be­yond “to­day isn’t a sta­ble equil­ibrium”, but I don’t see any­thing that over­all beats com­pe­ti­tion as a mo­tive for op­ti­miz­ing on leg­ible di­men­sions over illeg­ible ones, in a finite uni­verse with in­finite po­ten­tial de­sires.

• Dagon 22 Jan 2020 23:11 UTC

  3 points

  on: Safety reg­u­la­tors: A tool for miti­gat­ing tech­nolog­i­cal risk

  How do you deal with the knowl­edge prob­lem? Typ­i­cally, the ac­tual, ex­pe­rienced pain in steps 2 and 3 is crit­i­cal to the safety mea­sures im­ple­mented in 3 and en­joyed in 4. The progress is not de­layed for all pos­si­ble prob­lems, but the worst of them get ad­dressed—the in­cen­tive to be safe (re­duce pain) al­igns with the in­cen­tive to use the tech­nol­ogy at all.

  This works for pain (risk that’s short-term enough to mea­sure the cost and in­ci­dence of). It’s not clear that it works for rarer but more se­vere risks (x-risk or just gi­ant eco­nomic risk).

  In other words, the reg­u­la­tors are part of the tech­nol­ogy in the first place—what’s the guaran­tee (or even the mechanism to start) that the reg­u­la­tors are ad­dress­ing only the crit­i­cal risks?

• Dagon 22 Jan 2020 21:25 UTC

  2 points

  in reply to: crl826's comment on: How Doomed are Large Or­ga­ni­za­tions?

  I think this is the cen­tral puz­zle on the topic: where is the money com­ing from to pay the rats who are in (and cre­at­ing) the mazes? Why wouldn’t cus­tomers pre­fer a more effi­cient provider?

  My cur­rent spec­u­la­tion is that there’s a ton of slack on the scale were talk­ing about. Mazes aren’t ac­tu­ally less effi­cient than non-mazes, they just spend the slack on un­pleas­ant things rather than pleas­ant. To the ex­tent this is true, my ad­vice will ac­tu­ally re­duce over­all slack—the win­ners still will have to work harder and longer than they like. But they’ll en­joy it (both the work, and the re­main­ing slack) more. So, less over­all non-work en­ergy, but bet­ter able to use it for non-work pur­poses.

  Moloch still wins even­tu­ally, as even­tu­ally you have to com­pete with other hard-work­ing non-maze-waste orgs. But that can take a long time, and the ramp is far more pleas­ant.