abramdemski(Abram Demski)

• abramdemski 28 Nov 2020 5:08 UTC

  3 points

  on: Sny­der-Beat­tie, Sand­berg, Drexler & Bon­sall (2020): The Timing of Evolu­tion­ary Tran­si­tions Suggests In­tel­li­gent Life Is Rare

  Does this have a dooms­day-ar­gu­ment-like im­pli­ca­tion that we’re close to the end of the Earth’s liv­able span, be­cause if life takes a long time to evolve ob­servers then it’s over­whelm­ingly prob­a­ble that con­scious­ness arises shortly be­fore the end of the liv­able span?

• abramdemski 28 Nov 2020 4:44 UTC

  6 points

  on: Trans­par­ent Technologies

  I re­cently finished the book The Dic­ta­tor’s Hand­book. One thing that sur­prised me at the end: af­ter many chap­ters ex­am­in­ing (dys)func­tions of gov­ern­ments, the book sud­denly ap­plied the same con­cepts to cor­po­ra­tions.

  The main thrust of the book is that we can put coun­tries on a spec­trum from dic­ta­tor­ship to democ­ra­cies, by ex­am­in­ing the num­ber of key sup­port­ers needed for the gov­ern­ment to stay in place. Democ­ra­cies pre­dictably de­liver much more value to cit­i­zens. It’s a clas­sic story of in­cen­tives.

  If we look at cor­po­rate in­cen­tives, most (but not all) cor­po­ra­tions have a small num­ber of ma­jor share­hold­ers. They do a pretty good job of de­liv­er­ing value to those share­hold­ers.

  If we look at the cor­po­ra­tions that are more cus­tomer-owned, they do a bet­ter job of de­liv­er­ing value to cus­tomers. (Un­like many of the other claims in the book, this claim was not backed up with a bunch of data, only with a case study AKA anect­dote. But it makes sense!)

  Cus­tomer own­er­ship is im­ple­mented by EG set­ting limits on how many shares any one in­di­vi­d­ual can own, so that you can’t get a bunch of shares ac­cu­mu­lat­ing in a few hands.

  This was an aha mo­ment for me: it’s not eco­nom­i­cally in­evitable that, as you put it,

  The In­visi­ble Hand com­pels com­pa­nies to ex­tract max­i­mum profit from what­ever lev­er­age they have.

  A prepack­aged com­mer­cial product de­signed to make things eas­ier for con­sumers tends to con­tain anti-fea­tures.

  In­stead, it’s a con­se­quence of typ­i­cal cor­po­rate own­er­ship struc­ture.

• abramdemski 27 Nov 2020 23:13 UTC

  2 points

  in reply to: Zack_M_Davis’s comment on: Pon­tor’s Shortform

  Se­conded. Also Skyrms’ “Evolu­tion of the So­cial Con­tract”.

• abramdemski 27 Nov 2020 22:51 UTC

  5 points

  in reply to: Measure’s comment on: Luna Love­g­ood and the Cham­ber of Se­crets—Part 1

  Good catch, I hadn’t thought of that.

• abramdemski 27 Nov 2020 19:40 UTC

  6 points

  on: Is Suc­cess the Enemy of Free­dom? (Full)

  The con­trast be­tween this post and lsusr’s re­join­der re­minds me of the Ob­jec­tivist idea that suc­cess within the sys­tem always cor­rupts (to put it in LW ter­minol­ogy, you can’t feed Moloch with­out feed­ing Moloch), and of Mo­ral Mazes:

  Its the­sis is that this is the re­sult of a vi­cious cy­cle aris­ing from com­pet­i­tive pres­sures among those com­pet­ing for their own or­ga­ni­za­tional ad­vance­ment. Over time, those who fo­cus more on and more value such com­pe­ti­tions win them, gain power and fur­ther spread their val­ues, un­less they are ac­tively and con­tin­u­ously op­posed.

  Once things get bad in an or­ga­ni­za­tion they tend to only get worse, but things in gen­eral get bet­ter be­cause such or­ga­ni­za­tions then de­cay and are re­placed by new ones. Un­for­tu­nately, our so­ciety now slows or pre­vents that pro­cess, with these same or­ga­ni­za­tions and their val­ues in­creas­ingly run­ning the show.

  In­vest­ment and flex­i­bil­ity be­come im­pos­si­ble. Even ap­pear­ing to care about any­thing ex­cept the com­pe­ti­tion it­self costs you your al­lies. Thus things in­evitably de­cay and then col­lapse, flex­i­bil­ity re­turns, cy­cle re­peats.

  In­volve­ment with such pat­terns is far more de­struc­tive to hu­mans than is com­monly known.

• abramdemski 27 Nov 2020 19:08 UTC

  2 points

  in reply to: mr-hire’s comment on: Is Suc­cess the Enemy of Free­dom? (Full)

  I don’t think that’s all there is to it. Big firms have R&D. (In­deed, I have been told that big firms tar­get spend­ing spe­cific per­centages on R&D be­cause spend­ing too much or too lit­tle looks bad and would make their stocks go down, or some­thing like that.)

  I think big firms get eaten by Moloch in some fash­ion (lost pur­poses turn ev­ery­thing fake?), whereas star­tups have nicely al­igned in­cen­tives (be­cause it’s worth noth­ing un­less it suc­ceeds, so weird in­ter­nal power dy­nam­ics are not that valuable to fight over com­pared to fight­ing for the com­mon goal).

• abramdemski 27 Nov 2020 18:28 UTC

  6 points

  on: Suc­cess Buys Freedom

  I’m en­joy­ing the counter-alk­jash se­ries! In con­trast to my com­ment on the pre­vi­ous in­stal­l­ment, I think this in­stal­l­ment has a struc­tured model with gears, and told me some­thing about the world. (Still not much men­tal tech to counter failure modes.)

• abramdemski 27 Nov 2020 17:37 UTC

  2 points

  in reply to: jessicata’s comment on: Thoughts on Vot­ing Methods

  OK, I ba­si­cally don’t like the vot­ing sys­tem.

  Scott pointed out to me that the con­dorcet crite­rion makes more sense if we in­clude stochas­tic out­comes. In the cases where the Con­dorcet win­ner is the util­i­tar­ian-worst can­di­date, a mix­ture of other can­di­dates will win over the Con­dorcet win­ner. (So that can­di­date won’t re­ally be the Con­dorcet win­ner, if we in­clude stochas­tic out­comes as “can­di­dates”.)

  But that’s not what’s go­ing on here, be­cause this tech­nique always se­lects a Con­dorcet win­ner, if there is one.

  So (ap­par­ently) it’s not in­clud­ing stochas­tic out­comes in the right way.

  We can do bet­ter by mod­ify­ing the game:

  We spec­ify a sym­met­ric two-player zero-sum game where each player se­lects a dis­tri­bu­tion over can­di­dates. You score points based on how many more votes your pro­posed dis­tri­bu­tion would get against the other player’s. The game’s Nash equil­ibrium (a dis­tri­bu­tion over dis­tri­bu­tion over can­di­dates) is the out­put dis­tri­bu­tion.

  How­ever, I’m a bit sus­pi­cious of this, since I didn’t es­pe­cially like the ba­sic pro­posal and this is the same thing one level up.

  Since this is the unique vot­ing sys­tem un­der some con­sis­tency con­di­tions, I must not agree with some of the con­sis­tency con­di­tions, al­though I’m not sure which ones I dis­agree with.

• abramdemski 27 Nov 2020 14:34 UTC

  2 points

  on: Lat­est & Great­est Wiki-Tag Up­dates: Im­port, Dash­board, and Feed

  Long live the Twiki ;3

• abramdemski 27 Nov 2020 14:17 UTC

  2 points

  in reply to: Gurkenglas’s comment on: Nash Score for Vot­ing Techniques

  STAR is use­less if peo­ple can as­sign real-val­ued scores. That makes me think that if it works, it’s for rea­sons of dis­crete math­e­mat­ics, so we should an­a­lyze the sys­tem from the per­spec­tive of dis­crete math­e­mat­ics be­fore trust­ing it.

  A fair point. If vot­ers were al­lowed real-val­ued scores, they could make scores very close, and things still ba­si­cally de­volve into ap­proval vot­ing.

  \x → max(0, log(x)) does not seem like the op­ti­mal func­tion for any seemly pur­pose.

  Also true. I just don’t know how to con­tinue log into the nega­tive ;p

• abramdemski 27 Nov 2020 1:23 UTC

  LW: 4 AF: 3

  AF

  in reply to: Vika’s comment on: Stable Poin­t­ers to Value: An Agent Embed­ded in Its Own Utility Function

  Ah, looks like I missed this ques­tion for quite a while!

  I agree that it’s not quite one or the other. I think that like wire­head­ing, we can split delu­sion box into “the easy prob­lem” and “the hard prob­lem”. The easy delu­sion box is solved by mak­ing a re­ward/​util­ity which is model-based, and so, knows that the delu­sion box isn’t real. Then, much like ob­ser­va­tion-util­ity func­tions, the agent won’t think en­ter­ing into the delu­sion box is a good idea when it’s plan­ning—and also, won’t get any re­ward even if it en­ters into the delu­sion box ac­ci­den­tally (so long as it knows this has hap­pened).

  But the hard prob­lem of delu­sion box would be: we can’t make a perfect model of the real world in or­der to have model-based avoidance of the delu­sion box. So how to we guaran­tee that an agent avoids “gen­er­al­ized delu­sion boxes”?

• abramdemski 27 Nov 2020 0:49 UTC

  3 points

  in reply to: jimv’s comment on: Nash Score for Vot­ing Techniques

  I think part of the point (for me) of the Nash bar­gain­ing anal­ogy is that “no con­fi­dence” isn’t like other can­di­dates… but, yeah, that be­ing said, treat­ing it as a can­di­date would pro­duce more rea­son­able re­sults here. I agree that “as­sume con­fi­dence + STAR” with an ex­tra no-con­fi­dence can­di­date would be pretty rea­son­able com­pared to what I’ve come up with so far.

  Still hold­ing out hope for a more the­o­ret­i­cally jus­tified solu­tion if the game the­ory can be solved for the “col­lec­tive no con­fi­dence” bar­gain­ing game, though.

Nash Score for Vot­ing Techniques

• abramdemski 26 Nov 2020 15:24 UTC

  27 points

  on: Pain is the unit of Effort

  I note that Alk­jash’s post

  • had a struc­tured model with gears

  • told me some­thing about why the world is the way it is

  • pro­vided men­tal tech­niques to counter a problem

  I don’t think this post did any of these things. At least I didn’t ex­tract them if they were there.

  I’m not say­ing the mes­sage here is wrong or that a post like this couldn’t provide those three things. I just think this post didn’t achieve that.

  In what way is pain the unit of effort?

  What are peo­ple miss­ing about the world when they don’t see this?

  What TAPs can we im­ple­ment in light of these things?

• abramdemski 26 Nov 2020 14:34 UTC

  2 points

  in reply to: jessicata’s comment on: Thoughts on Vot­ing Methods

  Ah, I see. What I was miss­ing from this de­scrip­tion:

  What is the rule? Take a sym­met­ric zero-sum game where each player picks a can­di­date, and some­one wins if their can­di­date is preferred by the ma­jor­ity to the other, win­ning more points if they are preferred by a larger ma­jor­ity. This game’s Nash equil­ibrium is the dis­tri­bu­tion.

  was un­der­stand­ing that we con­struct a two player game, not a game where the play­ers are the vot­ers (or even a game where the can­di­dates are the play­ers).

De­con­struct­ing 321 Voting

• abramdemski 26 Nov 2020 2:08 UTC

  4 points

  in reply to: jungofthewon’s comment on: Embed­ded In­ter­ac­tive Pre­dic­tions on LessWrong

  Ah, a lot of this makes sense!

  So you’re from Ought?

  We speci­fi­cally want some peo­ple to make lots of pre­dic­tions so that other peo­ple can reuse the pre­dic­tions we house to an­swer new ques­tions.

  Yep, OK, this makes sense to me.

  It’s un­clear how ex­actly the LW com­mu­nity will use this in­te­gra­tion but if they use it to de­com­pose ar­gu­ments or op­er­a­tional­ize com­plex con­cepts, we can start to as­so­ci­ate rea­son­ing or ar­gu­men­ta­tive con­text with pre­dic­tions. It would be very cool if, given some para­graph of a LW post, we could pre­dict what fore­cast should be em­bed­ded next, or how a cer­tain claim should be op­er­a­tional­ized into a pre­dic­tion.

  Right, OK, this makes sense to me as well, al­though it’s cer­tainly more spec­u­la­tive.

  When Elicit has nice ar­gu­ment map­ping (it doesn’t yet, right?) it might be pretty cool and use­ful (to both LW and Ought) if that could be used on LW as well. For ex­am­ple, some­one could make an ar­gu­ment in a post, and then have an Elicit map (in­volv­ing sev­eral ques­tions linked to­gether) where LW users could re­veal what they think of the premises, the con­clu­sion, and the con­nec­tion be­tween them.

• abramdemski 25 Nov 2020 16:59 UTC

  LW: 6 AF: 4

  AF

  in reply to: romeostevensit’s comment on: Normativity

  Why not both? ;3

  I have noth­ing against jus­tifi­ca­tions be­ing cir­cu­lar (IE the same ideas re­cur­ring on many lev­els), just as I have noth­ing in prin­ci­ple against find­ing a foun­da­tion­al­ist ex­pla­na­tion. A cir­cu­lar ar­gu­ment is just a par­tic­u­larly sim­ple form of in­finite regress.

  But my main ar­gu­ment against only the cir­cu­lar rea­son­ing ex­pla­na­tion is that at­tempted ver­sions of it (“co­her­en­tist” po­si­tions) don’t seem very good when you get down to de­tails.

  Pure co­her­en­tist po­si­tions tend to rely on a stipu­lated no­tion of co­her­ence (such as prob­a­bil­is­tic co­her­ence, or weighted con­straint satis­fac­tion, or some­thing along those lines). Th­ese no­tions are them­selves fixed. This could be fine if the co­her­ence no­tions were suffi­ciently “as­sump­tion-lite” so as to not be nec­es­sar­ily Good­hart-prone etc, but so far it doesn’t seem that way to me.

  I’m pre­dict­ing that you’ll agree with me on that, and grant that the no­tion of co­her­ence should it­self be up for grabs. I don’t ac­tu­ally think the co­her­en­tist/​foun­da­tion­al­ist/​in­fini­tist trilemma is that good a char­ac­ter­i­za­tion of our dis­agree­ment here. My claim isn’t so much the clas­si­cal claim that there’s an in­finite regress of jus­tifi­ca­tion, as much as a claim that there’s an in­finite regress of un­cer­tainty—that we’re un­cer­tain at all the lev­els, and need to some­how man­age that. This fits the ship-of-the­seus pic­ture just fine.

  In other words, one can un­roll a ship of the­seus into an in­finite hi­er­ar­chy where each level says some­thing about how the next level down gets re-ad­justed over time. The rea­son for do­ing this is to achieve the foun­da­tion­al­ist goal of un­der­stand­ing the sys­tem bet­ter, with­out the foun­da­tion­al­ist method of fix­ing foun­da­tional as­sump­tions. The main mo­tive here is am­plifi­ca­tion. Tak­ing just a ship of the­seus, it’s not ob­vi­ous how to make it bet­ter be­sides run­ning it for­ward faster (and even this has its risks, since the ship may be­come worse). If we un­roll the hi­er­ar­chy of want­ing-to-be­come bet­ter, we can EG see what is good and bad about merely run­ning it for­ward faster, and try to run it for­ward in good ways rather than bad ways (as well as other, more rad­i­cal de­par­tures from sim­ple fast-for­ward am­plifi­ca­tion).

  One dis­agree­ment I have with your story is the ar­gu­ment “given the fini­tude of hu­man brain ar­chi­tec­ture”. The jus­tifi­ca­tion of a be­lief/​norm/​al­gorithm needn’t be some­thing already pre­sent in the head. A lot of what we do is given to us by evolu­tion. We can no­tice those things and ques­tion whether they make sense by our cur­rent stan­dards. Cal­ling this pro­cess finite is kind of like call­ing a Tur­ing ma­chine finite. There’s a finite core to it, but we can be sur­prised by what this core does given more work­ing tape.

• abramdemski 24 Nov 2020 17:23 UTC

  4 points

  in reply to: habryka’s comment on: Embed­ded In­ter­ac­tive Pre­dic­tions on LessWrong

  That is in­deed what I read. A quote:

  In our vi­sion, Elicit learns by imi­tat­ing the thoughts and rea­son­ing steps users share in the tool. It also gets di­rect feed­back from users on its sug­ges­tions.

  It just seems like LW pig­gy­backs on Elicit with­out re­veal­ing to Elicit any of the more com­plex stuff that goes into pre­dic­tions. Elicit wants to get data about (as I un­der­stand it) prob­a­bil­is­tic ar­gu­ment-map­ping. In­stead, it’s just get­ting point prob­a­bil­ities for ques­tions. That doesn’t seem very use­ful to me.

• abramdemski 23 Nov 2020 22:15 UTC

  8 points

  on: Embed­ded In­ter­ac­tive Pre­dic­tions on LessWrong

  I’m cu­ri­ous what this means for Ought. Is Ought plan­ning on do­ing any­thing with the data, and if so, what are the cur­rent spec­u­la­tions about that? How will Elicit evolve if the pri­mary user base is LessWrong?

  I looked at some of the stuff about Elicit on the Ought web­site, but I don’t see how LessWrong em­beds ob­vi­ously helps de­velop this vi­sion.

  (To be clear, I’m not com­plain­ing—if Elicit in­te­gra­tion is good for LW with no di­rect benefit for Ought that’s not a prob­lem! But I’m guess­ing Ought was mo­ti­vated to work with LW on this due to some per­ceived benefit.)