shminux

• shminux 20 Mar 2019 1:39 UTC

  4 points

  in reply to: Pattern's comment on: Ask LW: Have you read Yud­kowsky’s AI to Zom­bie book?

  Definitely Griffiths Quan­tum Me­chan­ics for un­der­grad. Lu­cid, clear, con­cise ex­po­si­tion. Also eas­ily found on­line.

• shminux 19 Mar 2019 4:00 UTC

  5 points

  in reply to: CaiwitzAzaria's comment on: Ask LW: Have you read Yud­kowsky’s AI to Zom­bie book?

  There are definitely cool parts in that se­quence, as you said. Non­lo­cal­ity, en­tan­gle­ment, the Liou­ville’s the­o­rem and some other con­cepts are pre­sented lu­cidly and un­der­stand­ably, and with the Bayesian view. They are other parts, that are ba­si­cally ad­vo­cacy of Many Wor­lds, that are much weaker, but still mostly fun to read. It’s just none of it is re­ally needed for learn­ing ra­tio­nal­ity.

  I am not a good au­thor­ity on pop­u­lar liter­a­ture about QM, as I had learned it aca­dem­i­cally from un­der­grad and grad-level texts, and it’s hard to go back to my pre­vi­ous self to do a fair eval­u­a­tion of a pop­u­lar book or a video se­ries.

• shminux 18 Mar 2019 15:14 UTC

  4 points

  on: Ask LW: Have you read Yud­kowsky’s AI to Zom­bie book?

  I read the Se­quences as they came out and went back and read those that were out be­fore I stum­bled upon LW. They were in many ways, eye-open­ing, ex­cept for the Quan­tum Physics se­quence, which I find su­perflu­ous, un­nec­es­sary, and, as some­one with a Physics PhD, not very ac­cu­rate. I have raised the ire of Eliezer here when I ex­pressed this opinion quite vo­cally. I still think that it can be sum­ma­rized in one sen­tence like “con­sider all pos­si­bil­ities and as­sign your best guess of prob­a­bil­ities to them”. There is no need for quan­tum at all, es­pe­cially given that QM can be simu­lated on a clas­si­cal com­puter, at worst with ex­po­nen­tial slow­down.

  My fa­vorite se­quence is, by far, Hu­man guide to words. I am not so fond of the Map/​Ter­ri­tory on­tol­ogy though, be­ing a “it’s maps all the way down” anti-re­al­ist (or post-ra­tio­nal­ist, as some tend to define them­selves). But ev­ery­thing that is re­lated to cog­ni­tive bi­ases and how to no­tice them in one­self is a must-read.

• shminux 18 Mar 2019 3:54 UTC

  4 points

  on: In­tu­itions on Nega­tive Utilitarianism

  Would a util­ity func­tion like U=hap­piness-tan(suffer­ing) de­scribe your in­tu­ition? Small amounts of hap­piness and suffer­ing con­tribute lin­early to the util­ity, but suffer­ing be­comes pro­gres­sively more im­por­tant as the amounts in­crease, un­til at some point (Pi/​2 in this ex­am­ple, but the units are ar­bi­trary, of course), the con­tri­bu­tion of suffer­ing to your util­ity func­tion goes to nega­tive in­finity, so no amount of hap­piness can out­weigh it?

• shminux 17 Mar 2019 21:43 UTC

  2 points

  on: More re­al­is­tic tales of doom

  Paul, is any­one at MIRI or el­se­where do­ing nu­mer­i­cal simu­la­tion of your ideas? Or are those just open-loop thoughts?

Boe­ing 737 MAX MCAS as an agent cor­rigi­bil­ity failure

• shminux 13 Mar 2019 20:44 UTC

  4 points

  on: For­mal­is­ing con­tin­u­ous info cas­cades? [Info-cas­cade se­ries]

  The Wikipe­dia en­try on the sub­ject is also quite good.

  The link is not to The Wikipe­dia, but to some pushy and con­fus­ing wiki reader app. Con­sider fix­ing.

  On a differ­ent note, have you tried to do nu­mer­i­cal simu­la­tions of the phe­nomenon you are de­scribing? Mul­ti­ple agents in­ter­act­ing un­der var­i­ous con­di­tions, watch if an equil­ibrium emerges, and what kind.

• shminux 13 Mar 2019 15:26 UTC

  12 points

  on: Un­der­stand­ing in­for­ma­tion cascades

  Have you read the back­re­ac­tion blog where Sabine Hossen­felder de­tails much the same phe­nomenon in high-en­ergy physics? She claims that the pre­vailing group­think ended up be­liev­ing into String The­ory with­out a shred of ev­i­dence for (only some vague hints), and so far with ev­ery sin­gle pre­dic­tion of it re­futed?

• shminux 12 Mar 2019 4:50 UTC

  2 points

  on: Meta-Dooms­day Ar­gu­ment: Uncer­tainty About the Val­idity of the Prob­a­bil­is­tic Pre­dic­tion of the End of the World

  I am one of those who feel like the DA is ut­ter BS. To test it fur­ther, I’d run a bunch of simu­la­tions and see how pre­dic­tive the past his­tory is of fu­ture de­vel­op­ments. I don’t know if the simu­la­tions or their math­e­mat­i­cal equiv­a­lents have been done, but here is what I would do:

  • Take a fixed-length fixed-pop­u­la­tion-size ver­sion, calcu­late the ac­cu­racy of the pre­dic­tions based on the time of the pre­dic­tion (it would be the best for those ex­actly mid­way-through the run). The rest of the ob­servers would be bi­ased in one of the two di­rec­tions.

  • Take a fixed length ex­po­nen­tially grow­ing pop­u­la­tion that ends abruptly, es­ti­mate the re­main­ing fu­ture time based on the num­ber of ob­servers and on the growth rate. The un­bi­ased pre­dic­tion hap­pens at the time of A-ln(2), which is ba­si­cally at the very end of a run of any du­ra­tion.

  • Re­peat with: ex­po­nen­tial to­tal du­ra­tion dis­tri­bu­tion with a given rate, with a va­ri­ety of lin­ear, ex­po­nen­tial and other growth rates, Gaus­sian dis­tri­bu­tion with the same, lo­gis­tic growth rates with all of the above to­tal du­ra­tion dis­tri­bu­tions, and any­thing else that makes sense to try.

  My guess would be that the re­sult of these calcu­la­tions would show that there is ab­solutely no cor­re­la­tion in gen­eral be­tween pre­dic­tions and ac­tual out­comes, and that the ac­cu­racy would vary wildly based on the as­sump­tions, and there is lit­tle or no a pri­ori rea­son to pick one spe­cific dis­tri­bu­tion over an­other, ex­cept the im­pe­tus to spec­u­late based on the lack of in­for­ma­tion, a nat­u­ral hu­man trait.

• shminux 8 Mar 2019 15:41 UTC

  5 points

  on: Ideas for an ac­tion co­or­di­na­tion website

  Con­sider first what worked and what didn’t work his­tor­i­cally for co­or­di­na­tion as a means to move be­tween equil­ibria. Revolu­tions/​up­ris­ings/​unioniz­ing/​boy­cotting some­times worked and some­times didn’t. What de­ter­mines the odds of suc­cess?

• shminux 7 Mar 2019 2:58 UTC

  6 points

  on: Ask­ing for help teach­ing a crit­i­cal think­ing class.

  One of the bet­ter skills you can im­part on them is in­ter­nal­iz­ing their own fal­li­bil­ity. And not a one-time ex­er­cise that leaves them feel­ing that it was all a trick that does not ap­ply to their day-to-day life. A per­sonal re­flec­tion on how you came to re­al­ize your own limits, and how you grap­pled with be­ing over­con­fi­dent in some­thing you be­lieved. Maybe have them de­scribe, to the class or in small groups, a case where they were 100% sure of some­thing, and then re­al­ized they were wrong, and dis­cuss the rea­sons they ended up be­ing wrong, and what les­son they learned, or didn’t learn. Then an­other ex­er­cise talk­ing about what they are cur­rently 100% sure about, and why, and which of their im­plicit as­sump­tions would have to be in­valid for this cer­tainty to de­crease to some­thing rea­son­able. Bonus points to those who, af­ter do­ing the ex­er­cise, ac­tu­ally changed their mind about some­thing im­por­tant and dear to them. Ex­tra bonus points to those who con­cluded that they need to learn more and ask for more in­for­ma­tion on how to do that. Just off the top of my head, any­way.

• shminux 4 Mar 2019 8:14 UTC

  4 points

  in reply to: Charlie Steiner's comment on: To un­der­stand, study edge cases

  Not sure if it’s “con­tra” your post. In physics there are effec­tive the­o­ries and the In­ten­tional Stance is an ex­am­ple of one, which is, in­deed, a “low-en­ergy” ap­prox­i­ma­tion. My point is not about en­ergy level, but more about find­ing the limits of ap­pli­ca­bil­ity to see the con­stituents bet­ter. For ex­am­ple in a con­trol sys­tem it is hard to figure out the in­nards un­til you, say, open the feed­back loop in some way, or find a regime where the the sys­tem goes into res­o­nance, or sat­u­rates, etc.

• shminux 3 Mar 2019 19:06 UTC

  2 points

  in reply to: Aiyen's comment on: To un­der­stand, study edge cases

  I meant it more in a sense that peo­ple like Michelan­gelo had to un­der­stand the in­ner work­ings of the hu­man body, well, the anatomy of it, to paint it faith­fully when it’s in one piece.

• shminux 3 Mar 2019 5:12 UTC

  5 points

  in reply to: quanticle's comment on: To un­der­stand, study edge cases

  I agree that edge cases are not very good for syn­the­sis. They are es­sen­tial for anal­y­sis, how­ever. Notic­ing pat­terns in a well func­tion­ing sys­tem is use­ful for build­ing phe­nomenolog­i­cal mod­els of the sys­tem, but not nec­es­sar­ily for figur­ing out its con­stituents. And yeah, some­times it is pos­si­ble to figure out a lot from a nom­i­nally func­tion­ing sys­tem.

To un­der­stand, study edge cases

• shminux 22 Feb 2019 2:27 UTC

  4 points

  on: If a “Kick­starter for Inad­e­quate Equlibria” was built, do you have a con­crete in­ad­e­quate equil­ibrium to fix?

  An im­por­tant met­ric is an in­de­pen­dent ex­pert es­ti­mate of how likely a par­tic­u­lar equil­ibrium will be shaken if there is a large-scale co­or­di­na­tion effort, and what po­ten­tial costs might be. Strikes, lo­cal­ized (e.g. at a Wal­mart store) or coun­try-wide (an air-traf­fic con­trol­ler strike dur­ing Rea­gan years) of­ten back­fire badly. An in­side es­ti­mate is guaran­teed to be heav­ily bi­ased, and so is largely use­less. So I would run any sug­ges­tion for a kick­starter like that through a pay­off es­ti­mate filter.

• shminux 21 Feb 2019 16:03 UTC

  2 points

  in reply to: philh's comment on: Grace­ful Shutdown

  Yes, a few sim­ple shut­down no­tifi­ca­tions with­out wait­ing for feed­back is a rea­son­able thing to do. No wait­ing for replies, no grace pe­ri­ods. Maybe wait a sec­ond or two be­fore cut­ting the power, if you are in charge of pow­er­ing the clients and a shut­down means power off.

• shminux 21 Feb 2019 8:13 UTC

  11 points

  on: In-group in­den­ti­fi­ca­tion?

  I sus­pect that the idea of a visi­ble group iden­ti­fi­ca­tion sym­bol would evoke an im­me­di­ate cringe from most reg­u­lar read­ers of this fo­rum. If you feel a strong urge to be­long to a group and show that you be­long to the oth­ers, a loose col­lec­tion of peo­ple who are in­ter­ested in im­prov­ing their think­ing pat­terns is a poor can­di­date for such a group.

• shminux 20 Feb 2019 4:39 UTC

  0 points

  on: Blackmail

  I am not sure why you pick on black­mail speci­fi­cally. Most of your points ap­ply to many kinds of hu­man in­ter­ac­tions. The origi­nal rea­son for some spe­cific ac­tions be­ing made ille­gal is based on con­se­quen­tial­ism, to con­vert util­i­tar­ian rea­sons into de­on­tolog­i­cal ones, and, af­ter a time, into virtue-eth­i­cal. Not all high nega­tive util­ity ac­tions get out­lawed, and many pos­i­tive util­ity ac­tions get out­lawed for other rea­sons, but the gen­eral pat­tern per­sists. Changes in so­ciety even­tu­ally re­sult in changes in laws. What used to be ille­gal be­comes le­gal and vice versa, gen­er­ally based on the amount of real or per­ceived harm it causes. When copy­ing be­came triv­ial, copy­right laws grew teeth. Once same sex is no longer an emo­tional hor­ror, it is no longer out­lawed. If one day hu­man life be­comes cheap again (like it is now in some places), mur­der will even­tu­ally be­come le­gal and ac­cepted. To pro­duc­tively dis­cuss black­mail’s le­gal­ity one would need to eval­u­ate the ac­tual, not imag­ined or edge-cases harm it causes in the con­text of other ac­tivi­ties, le­gal and ille­gal, and see where it fits on the util­ity scat­ter plot. If you find it to be broadly among the cloud of ille­gal ac­tivi­ties, then you have a case for it be­ing made ille­gal. If it is on the mar­gins be­tween le­gal and ille­gal, then you don’t have a case. That’s it.

• shminux 19 Feb 2019 2:39 UTC

  3 points

  on: De­cel­er­at­ing: laser vs gun vs rocket

  Avoid­ing car­ry­ing fuel with you is cer­tainly tempt­ing, so­lar sail and var­i­ous beam-pow­ered types of propul­sion are some of the bet­ter known ones. Var­i­ous types of the so­lar wind sail would give much more thrust if one could figure out how to effi­ciently stop or deflect the so­lar wind par­ti­cles (mostly pro­tons), and var­i­ous ver­sions of a ram­jet when in in­ter­stel­lar space. https://​en.wikipe­dia.org/​wiki/​Mag­netic_sail#In­ter­stel­lar_travel is a de­cent sum­mary. My bet is on that kind of tech­nol­ogy, un­less some­one rev­olu­tionizes space travel by figur­ing out how to bend space­time more effi­ciently than with sheer mass, and makes some­thing like the Alcu­bierre drive fea­si­ble.