Vladimir_Nesov(Vladimir Nesov)

• Vladimir_Nesov 15 Sep 2020 16:45 UTC

  5 points

  in reply to: swarriner’s comment on: If Star­ship works, how much would it cost to cre­ate a sys­tem of rotable space mir­rors that re­duces tem­per­a­tures on earth by 1° C?

  I think it’s use­ful to dis­t­in­guish knowl­edge of truth from gears-level un­der­stand­ing, these two differ­ent things can oc­cur in any com­bi­na­tion. Your point is that at­tain­ing spe­cific un­der­stand­ing of a plan that’s good enough to make the es­ti­mate in ques­tion is a hope­less en­deavor, and you list par­tic­u­lar is­sues with get­ting such a plan fleshed out.

  But it’s also pos­si­ble to know truths about the world with­out un­der­stand­ing why they are true or how they came to be known (origi­nally). The main ex­am­ple of this is seek­ing ex­pert con­sen­sus in an area you don’t un­der­stand: by find­ing out what the con­sen­sus is, you get a rea­son­able cre­dence in what the truth of the mat­ter is, with­out nec­es­sar­ily un­der­stand­ing why it’s this way, or how speci­fi­cally any­one came to know it’s this way.

  This post asks for a Fermi es­ti­mate, which is an­other way in which a very vague model can yield truths about the world. Even if a de­tailed model is unattain­able, such truths might be in reach.

  (It’s of­ten a lost pur­pose to seek truths about the world in­stead of seek­ing un­der­stand­ing, so it’s nat­u­ral to scorn some forms of pur­suit of truths. I have a lot of sym­pa­thy for this po­si­tion. That doesn’t make such forms of pur­suit of truths un­work­able, just not rele­vant to im­prov­ing un­der­stand­ing of what’s go­ing on.)

• Vladimir_Nesov 12 Sep 2020 15:13 UTC

  16 points

  on: What’s Wrong with So­cial Science and How to Fix It: Reflec­tions After Read­ing 2578 Papers

  A typ­i­cal pa­per doesn’t just con­tain fac­tual claims about stan­dard ques­tions, but also the­o­ret­i­cal dis­cus­sion and a point of view on the ideas that form the fabric of a field. Papers are of­ten refer­enced to clar­ify the mean­ing of a the­o­ret­i­cal dis­cus­sion, or to give credit for in­spiring the di­rec­tion in which the dis­cus­sion moves. This as­pect doesn’t sig­nifi­cantly de­pend on truth of find­ings of par­tic­u­lar stud­ies, be­cause an in­ter­est­ing con­cept mo­ti­vates many stud­ies that both ex­per­i­men­tally in­ves­ti­gate and the­o­ret­i­cally dis­cuss it. Some of the stud­ies will be fac­tu­ally bo­gus, but the the­o­ret­i­cal dis­cus­sion in them might still be rele­vant to the con­cept, and use­ful for sub­se­quent good stud­ies.

  So a clas­sifi­ca­tion of cita­tions into pos­i­tive and nega­tive ig­nores this im­por­tant third cat­e­gory, some­thing like con­cep­tual refer­ence cita­tion.

• Vladimir_Nesov 19 Aug 2020 14:18 UTC

  2 points

  on: On Defin­ing your Terms

  A pro­duc­tive pur­pose of giv­ing defi­ni­tions to your in­ter­locu­tor is to in­cite cog­ni­tive ac­tivity that’s rele­vant to your own thoughts, as lev­ers that bring about think­ing on the right topic. Apt defi­ni­tions are great, when available, but de­mand­ing them when they are not is los­ing sight of the pur­pose of the whole ac­tivity. For­mu­lat­ing defi­ni­tions (or re­search­ing ex­ist­ing ones) might be a good sub­prob­lem to fo­cus on though.

  For vague ideas it makes lit­tle sense to use defi­ni­tions that sharply delineate their in­stances. A defi­ni­tion should in­stead give a de­gree of apt­ness (cen­tral­ity) to its po­ten­tial in­stances, and defi­ni­tions should match based on the whole dis­tri­bu­tion of apt­ness they as­sign.

  With free will, and similarly with pos­si­ble wor­lds, I think a cru­cial point miss­ing from dis­cus­sion on LW is that these are se­man­tic no­tions, and it’s pos­si­ble to con­sider rad­i­cally differ­ent se­man­tics for the same syn­tac­tic thing. A great illus­tra­tion is se­man­tics of pro­gram­ming lan­guages, where it’s pos­si­ble to un­der­stand what a given pro­gram is do­ing in terms of very differ­ent se­man­tic con­struc­tions. To give some ex­am­ples, there’s the straight­for­ward sets of pos­si­ble val­ues, re­lated wor­lds in Kripke se­man­tics, for­mal the­o­ries of ob­ser­va­tions (that are not even max­i­mal) in Scott do­mains, plays and strate­gies in game se­man­tics. Closer to nor­mal math­e­mat­ics, there’s in­ter­nal lan­guages of cat­e­gories that let us in­ter­pret a term as some­thing in­car­nated in very differ­ent situ­a­tions, no­tably sheaf toposes where you can in­ter­pret a con­struc­tion (as in dis­cus­sion, ar­gu­ment, proof, term, type) as vary­ing con­tin­u­ously on some space and say­ing differ­ent things for differ­ent places, all at the same time.

  So when con­sid­er­ing a thought ex­per­i­ment, as­sum­ing that what’s go­ing on is that there is some dis­crete set of pos­si­ble wor­lds is very limit­ing. There are other ways to think about things, and thus pre­ma­turely lock­ing in a dis­cus­sion un­der a tech­ni­cal defi­ni­tion can be a prob­lem, man­i­fest­ing as failure to no­tice or urge to ridicule the no­tions aris­ing from other defi­ni­tions and in­apt for this one. (Not for­mu­lat­ing tech­ni­cal defi­ni­tions is an­other prob­lem.)

• Vladimir_Nesov 26 Jul 2020 1:23 UTC

  6 points

  on: A pro­posal for train­ing calcu­lus teachers

  We can’t just screen Prof. Leonard for the 5,000 Amer­i­can col­lege calc I classes taught ev­ery year, be­cause many stu­dents want a class­room ex­pe­rience and teach­ers need jobs.

  Re “teach­ers need jobs”: If you get a bet­ter ex­pe­rience with­out teach­ers, you could just hire the teach­ers to do noth­ing.

• Vladimir_Nesov 18 Jul 2020 23:20 UTC

  2 points

  in reply to: Bob Jacobs’s comment on: Bob Ja­cobs’s Shortform

  (I think mak­ing ar­gu­ments clear is more mean­ingful than us­ing them for per­sua­sion.)

• Vladimir_Nesov 18 Jul 2020 20:48 UTC

  2 points

  in reply to: Bob Jacobs’s comment on: Bob Ja­cobs’s Shortform

  It’s not clear what “sub­jec­tive ideal­ism is cor­rect” means, be­cause it’s not clear what “a given thing is real” means (at least in the con­text of this thread). It should be more clear what a claim means be­fore it makes sense to dis­cuss lev­els of cre­dence in it.

  If we are work­ing with cre­dences as­signed to hy­po­thet­i­cals, the fact that the num­ber of dis­joint hy­po­thet­i­cals in­com­pat­i­ble with some hy­po­thet­i­cal S is large doesn’t in it­self make them (when con­sid­ered al­to­gether) more prob­a­ble than S. (A sum of an in­finite num­ber of small num­bers can still be small.)

  Work­ing with cre­dences in hy­po­thet­i­cals is not the only pos­si­ble way to rea­son. If we are talk­ing about weird things like sub­jec­tive ideal­ism, as­sump­tions about epistemics are not straight­for­ward and should be con­sid­ered.

• Vladimir_Nesov 14 Jul 2020 13:34 UTC

  6 points

  on: Null-box­ing New­comb’s Problem

  If the trick­ster god per­son­ally reads the pre­dic­tion, their be­hav­ior can de­pend on the pre­dic­tion, which makes di­ag­o­nal­iza­tion pos­si­ble (ask the trick­ster god what the pre­dic­tion was, then do the op­po­site). This calls the claim of 100% pre­ci­sion of the pre­dic­tor into ques­tion (or at least makes the de­tails of its mean­ing rele­vant).

• Vladimir_Nesov 11 Jul 2020 21:19 UTC

  5 points

  in reply to: jimv’s comment on: Max­i­mal Ventilation

  Not doubt­ing it at all

  I don’t think doubt­ing should be ei­ther so­cially frowned-upon or un­called-for where one doesn’t see an ar­gu­ment that makes a claim ev­i­dent.

• Vladimir_Nesov 20 Oct 2019 14:44 UTC

  4 points

  in reply to: Clay S’s comment on: Why Ranked Choice Vot­ing Isn’t Great

  Maybe “near-fatal” is too strong a word, the com­ment I replied to also had ex­am­ples. Ex­is­tence of ex­am­ples doesn’t dis­t­in­guish win­ning from sur­vival, see­ing some use. I un­der­stand the state­ment I replied to as mean­ing some­thing like “In 200 years, if the world re­mains mostly as we know it, the prob­a­bil­ity that most elec­tions use car­di­nal vot­ing meth­ods is above 50%”. This seems im­plau­si­ble to me for the rea­sons I listed, hence the ques­tion about what you ac­tu­ally meant, per­haps my in­ter­pre­ta­tion of the state­ment is not what you in­tended. (Is “long run” some­thing like 200 years? Is “win­ning” some­thing like “most elec­tions of some kind use car­di­nal vot­ing meth­ods”?)

• Vladimir_Nesov 20 Oct 2019 6:14 UTC

  2 points

  in reply to: Clay S’s comment on: Why Ranked Choice Vot­ing Isn’t Great

  Car­di­nal vot­ing meth­ods will win in the long run.

  (What kind of long run? Why is this to be ex­pected?) Pop­u­lar­ity is not based on only merit, be­ing more com­pli­cated than the sim­plest most fa­mil­iar method sounds like a near-fatal dis­ad­van­tage. Vot­ing be­ing in­volved with poli­tics makes it even harder for good ar­gu­ments to in­fluence what ac­tu­ally hap­pens.

• Vladimir_Nesov 20 Oct 2019 4:02 UTC

  6 points

  in reply to: Wei_Dai’s comment on: Maybe Ly­ing Doesn’t Exist

  The prob­lem with un­re­strained con­se­quen­tial­ism is that it ac­cepts no prin­ci­ples in its de­signs. An agent that only serves a pur­pose has no knowl­edge of the world or math­e­mat­ics, it makes no plans and main­tains no goals. It is what it needs to be, and no more. All these things are only ex­pressed as as­pects of its be­hav­ior, god­shat­ter of the sin­gu­lar pur­pose, but there is no part that seeks ex­cel­lence in any of the as­pects.

  For an agent de­signed around mul­ti­ple as­pects, its parts rely on each other in dis­similar ways, not as sub­agents with differ­ent goals. Ac­cess to knowl­edge is use­ful for plan­ning and can rep­re­sent goals. Ex­plo­ra­tion and re­flec­tion re­fine knowl­edge and for­mu­late goals. Plan­ning op­ti­mizes ex­plo­ra­tion and re­flec­tion, and leads to achieve­ment of goals.

  If the part of the de­sign that should hold knowl­edge ac­cepts a claim for rea­sons other than ar­gu­ments about its truth, the rest of the agent can no longer rely on its claims as re­flect­ing knowl­edge.

  Of course you’d have to also patch the specification

  In my com­ment, I meant the situ­a­tion where the speci­fi­ca­tion is not patched (and by speci­fi­ca­tion in the pro­gram­ming ex­am­ple I meant the in­for­mal de­scrip­tion on the level of pro­ce­dures or datatypes that es­tab­lishes some prin­ci­ples of what it should be do­ing).

  In the case of ap­peal to con­se­quences, the speci­fi­ca­tion is a gen­eral prin­ci­ple that a map re­flects the ter­ri­tory to the best of its abil­ity, so it’s not a small thing to patch. Op­ti­miz­ing a par­tic­u­lar be­lief ac­cord­ing to the con­se­quences of hold­ing it vi­o­lates this gen­eral speci­fi­ca­tion. If the gen­eral speci­fi­ca­tion is patched to al­low this, you no longer have ac­cess to straight­for­wardly ex­pressed knowl­edge (there is no part of cog­ni­tion that satis­fies the origi­nal speci­fi­ca­tion).

  Alter­na­tively, spe­cific be­liefs could be marked as mo­ti­vated, so the speci­fi­ca­tion is to have two kinds of be­liefs, with some of them sur­viv­ing to serve the origi­nal pur­pose. This might work, but then ac­tual knowl­edge that cor­re­sponds to the mo­ti­vated be­liefs won’t be na­tively available, and it’s un­clear what the mo­ti­vated be­liefs should be do­ing. Will cu­ri­os­ity act on the mo­ti­vated be­liefs, should they be used for plan­ning, can they rep­re­sent goals? A more de­vel­oped ar­chi­tec­ture for re­li­able hypocrisy might ac­tu­ally do some­thing sen­si­ble, but it’s not a mat­ter of merely patch­ing par­tic­u­lar be­liefs.

• Vladimir_Nesov 19 Oct 2019 14:24 UTC

  16 points

  in reply to: Wei_Dai’s comment on: Maybe Ly­ing Doesn’t Exist

  cor­rectly weigh these kinds of con­sid­er­a­tions against each on a case by case basis

  The very pos­si­bil­ity of in­ter­ven­tion based on weigh­ing map-mak­ing and plan­ning against each other de­stroys their de­sign, if they are to have a de­sign. It’s similar to patch­ing a pro­ce­dure in a way that vi­o­lates its speci­fi­ca­tion in or­der to im­prove over­all perfor­mance of the pro­gram or to fix an ex­ter­nally ob­serv­able bug. In the­ory this can be benefi­cial, but in prac­tice the abil­ity to rea­son about what’s go­ing on de­te­ri­o­rates.

• Vladimir_Nesov 24 Sep 2019 3:56 UTC

  LW: 2 AF: 1

  AF

  in reply to: abramdemski’s comment on: Mechanis­tic Corrigibility

  I agree that ex­otic de­ci­sion al­gorithms or prefer­ence trans­for­ma­tions are prob­a­bly not go­ing to be use­ful for al­ign­ment, but I think this kind of ac­tivity is cur­rently more fruit­ful for the­ory build­ing than di­rectly try­ing to get de­ci­sion the­ory right. It’s just that the usual fram­ing is sus­pect: in­stead of ex­plo­ra­tion of the de­ci­sion the­ory land­scape by con­sid­er­ing clearly bro­ken/​in­sane-act­ing/​use­less but not yet well-un­der­stood con­struc­tions, these things are pitched (and cho­sen) for their per­ceived use in al­ign­ment.

• Vladimir_Nesov 22 Sep 2019 5:55 UTC

  2 points

  in reply to: TurnTrout’s comment on: TurnTrout’s short­form feed

  I sus­pect that it doesn’t mat­ter how ac­cu­rate or straight­for­ward a pre­dic­tor is in mod­el­ing peo­ple. What would make pre­dic­tion morally ir­rele­vant is that it’s not no­ticed by the pre­dicted peo­ple, ir­re­spec­tive of whether this hap­pens be­cause it spreads the moral weight con­ferred to them over many pos­si­bil­ities (giv­ing in­ac­cu­rate pre­dic­tion), keeps the rep­re­sen­ta­tion suffi­ciently baroque, or for some other rea­son. In the case of in­ac­cu­rate pre­dic­tion or baroque rep­re­sen­ta­tion, it prob­a­bly does be­come harder for the pre­dicted peo­ple to no­tice be­ing pre­dicted, and I think this is the ac­tual source of moral ir­rele­vance, not those things on their own. A more di­rect way of get­ting the same re­sult is to pre­dict coun­ter­fac­tu­als where the peo­ple you rea­son about don’t no­tice the fact that you are ob­serv­ing them, which also gives a form of in­ac­cu­racy (imag­ine that your pre­dict­ing them is part of their prior, that’ll drive the coun­ter­fac­tual fur­ther from re­al­ity).

• Vladimir_Nesov 17 Sep 2019 20:04 UTC

  4 points

  in reply to: Pattern’s comment on: Open & Wel­come Thread—Septem­ber 2019

  This still puts these com­ments in Re­cent Com­ments on GreaterWrong, and the fact that they can’t be seen on the LessWrong All Com­ments page is es­sen­tially a bug.

• Vladimir_Nesov 15 Sep 2019 14:11 UTC

  LW: 10 AF: 3

  AF

  in reply to: Wei_Dai’s comment on: A Cri­tique of Func­tional De­ci­sion Theory

  By the way, self­ish val­ues seem re­lated to the re­ward vs. util­ity dis­tinc­tion. An agent that pur­sues a re­ward that’s about par­tic­u­lar events in the world rather than a more holo­graphic val­u­a­tion seems more like a self­ish agent in this sense than a max­i­mizer of a util­ity func­tion with a small-in-space sup­port. If a re­ward-seek­ing agent looks for re­ward chan­nel shaped pat­terns in­stead of the in­stance of a re­ward chan­nel in front of it, it might tile the world with re­ward chan­nels or search the world for more of them or some­thing like that.

• Vladimir_Nesov 15 Sep 2019 13:54 UTC

  6 points

  in reply to: Gurkenglas’s comment on: Prov­ing Too Much (w/​ ex­er­cises)

  “I think, there­fore I am.”

  (This is also in­cor­rect, be­cause con­sid­er­ing a think­ing you in a coun­ter­fac­tual makes sense. Many UDTish ex­am­ples demon­strate that this prin­ci­ple doesn’t hold.)

• Vladimir_Nesov 14 Sep 2019 18:38 UTC

  LW: 4 AF: 2

  AF

  in reply to: Vanessa Kosoy’s comment on: For­mal­is­ing de­ci­sion the­ory is hard

  I was never con­vinced that “log­i­cal ASP” is a “fair” prob­lem. I once joked with Scott that we can con­sider a “pre­dic­tor” that is just the sin­gle line of code “re­turn DEFECT” but in the com­ments it says “I am defect­ing only be­cause I know you will defect.”

  I’m lean­ing this way as well, but I think it’s an im­por­tant clue to figur­ing out com­mit­ment races. ASP Pre­dic­tor, Defec­tBot, and a more gen­eral agent will make differ­ent com­mit­ments, and these things are already al­gorithms spe­cial­ized for cer­tain situ­a­tions. How is the cho­sen com­mit­ment re­lated to what the thing mak­ing the com­mit­ment is?

  When an agent can ma­nipu­late a pre­dic­tor in some sense, what should the pre­dic­tor do? If it starts schem­ing with its thoughts, it’s no longer a pre­dic­tor, it’s just an­other agent that wants to do some­thing “pre­dic­tory”. Maybe it can only give up, as in ASP, which acts as a pre­com­mit­ment that’s more the­mat­i­cally fit­ting for a pre­dic­tor than for a gen­eral agent. It’s still a com­mit­ment race then, but pos­si­bly the mean­ing of some­thing be­ing a pre­dic­tor is pre­served by re­strict­ing the kind of com­mit­ment that it makes: the com­mit­ment of a non-gen­eral agent is what it is rather than what it does, and a gen­eral agent is only com­mit­ted to its prefer­ence. Thus a gen­eral agent loses all knowl­edge in an at­tempt to out-com­mit oth­ers, be­cause it hasn’t com­mit­ted to that knowl­edge, didn’t make it part of what it is.

• Vladimir_Nesov 13 Sep 2019 16:48 UTC

  9 points

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

  (By “be­lief” I meant a be­lief that talkes place in some­one’s head, and its ex­is­tence is not nec­es­sar­ily com­mu­ni­cated to any­one else. So an ut­tered state­ment “I think X” is a dec­la­ra­tion of be­lief in X, not just a be­lief in X. A be­lief in X is just a fact about that per­son’s mind, with­out an ac­com­pa­ny­ing dec­la­ra­tion. In this fram­ing, the ver­sion of the norm about be­liefs (as op­posed to dec­la­ra­tions) is the norm not to think cer­tain thoughts, not a norm to avoid shar­ing the ob­ser­va­tions about the fact that you are think­ing them.)

  I think a salient dis­tinc­tion be­tween dec­la­ra­tions of “I think X” and “it’s true that X” is a bad thing, as de­scribed in this com­ment. The dis­tinc­tion is that in the former case you might lack ar­gu­ments for the be­lief. But if you don’t en­dorse the be­lief, it’s no longer a be­lief, and “I think X” is a bug in the mind that shouldn’t be called “be­lief”. If you do en­dorse it, then “I think X” does mean “X”. It is plau­si­bly a true state­ment about the state of the uni­verse, you just don’t know why; your mind in­scrutably says that it is and you are in­clined to be­lieve it, pend­ing fur­ther in­ves­ti­ga­tion.

  So the state­ment “I think this is true of other peo­ple in spite of their claims to the con­trary” should mean ap­prox­i­mately the same as “This is true of other peo­ple in spite of their claims to the con­trary”, and a mean­ingful dis­tinc­tion only ap­pears with ac­tual ar­gu­ments about those state­ments, not with differ­ent place­ment of “I think”.

• Vladimir_Nesov 12 Sep 2019 15:55 UTC

  10 points

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

  crit­i­ciz­ing peo­ple who don’t jus­tify their be­liefs with ad­e­quate ev­i­dence and arguments

  I think jus­tifi­ca­tion is in the na­ture of ar­gu­ments, but not nec­es­sary for be­liefs or dec­la­ra­tions of be­liefs. A be­lief offered with­out jus­tifi­ca­tion is a hy­poth­e­sis called to at­ten­tion. It’s con­cise, and if han­dled care­fully, it can be suffi­cient for com­mu­ni­ca­tion. As ev­i­dence, it’s a claim about your own state of mind, which holds a lot of in­scrutable ter­ri­tory that nonethe­less can chan­nel un­der­stand­ing that doesn’t yet lend it­self to ar­gu­ments. Seek­ing ar­gu­ments is cer­tainly a good thing, to re­fac­tor and con­vey be­liefs, but that’s only a small part of how hu­man in­tel­li­gence builds its map.

  What links here?

  • Vladimir_Nesov's comment on G Gor­don Wor­ley III’s Shortform by G Gordon Worley III (13 Sep 2019 16:48 UTC; 9 points)