NunoSempere

• NunoSempere 12 Feb 2020 15:38 UTC

  5 points

  on: AI Im­pacts: His­toric trends in tech­nolog­i­cal progress

  I’ve been fol­low­ing this with in­ter­est.

  Re: Telecom­mu­ni­ca­tions perfor­mance, the red tele­phone might also be a a dis­con­ti­nu­ity in prac­ti­cal terms.

  The 1962 Cuban Mis­sile Cri­sis made the hotline a pri­or­ity. Dur­ing the stand­off, offi­cial diplo­matic mes­sages typ­i­cally took six hours to de­liver; un­offi­cial chan­nels, such as via tele­vi­sion net­work cor­re­spon­dents, had to be used too as they were quicker

  That is, even though faster sys­tems ex­isted, they hadn’t been im­ple­mented in the area of com­mu­ni­ca­tions be­tween the Soviet Union and the USA (pretty huge blindspot), but could be im­ple­mented more or less im­me­di­ately, once both regimes ac­tu­ally both­ered.

  Also of in­ter­est to read­ers might be: some other dis­con­ti­nu­ities, one in pas­sen­ger ship length and the other one on time needed to cir­cum­nav­i­gate the Earth. AI im­pacts also has a cou­ple of other dis­con­ti­nu­ities on their web­page, not men­tioned/​linked above:

  • Ship size (also pretty re­cently added)

  • From nu­clear weapons

  • Altitude

  • ...

• NunoSempere 31 Jan 2020 8:52 UTC

  1 point

  on: EA Men­tal Health Sur­vey: Re­sults and Anal­y­sis.

  Run­ning the re­sults again on the SlateS­tarCodex Sur­vey re­sults of this year. EAs are less men­tally ill than non-EAs, who are less men­tally ill than re­spon­dents who iden­tify as “sort of” EA. From this we should con­clude that all three groups are all ba­si­cally the same, be­cause the differ­ence is not sig­nifi­cant at all.

  The code used to find that out is, in R:

  setwd(“my/​di­rec­tory”) ## di­rec­tory in which the 2020ssc_pub­lic.xlsx file re­sides, down­load­able from https://​​slat­estar­codex.com/​​2020/​​01/​​20/​​ssc-sur­vey-re­sults-2020/​​
  in­stall.pack­ages(“openxlsx”, de­pen­den­cies = TRUE)
  library(openxlsx)
  
  D ← read.xlsx(“2020ssc_pub­lic.xlsx”)
  
  men­tally_ill_strict =  c(“I have a for­mal di­ag­no­sis of this con­di­tion”)
  men­tally_ill_loose =  c(“I have a for­mal di­ag­no­sis of this con­di­tion”, “I think I might have this con­di­tion, al­though I have never been for­mally di­ag­nosed”)
  
  m ← men­tally_ill_strict
  D$Men­tal­lyIll_1 = D$De­pres­sion %in% m | D$Anx­iety %in% m | D$OCD %in% m  | D$Eat­ingdi­s­or­der %in% m  | D$PTSD %in% m  | D$Al­co­holism %in% m  | D$Dru­gad­dic­tion %in% m  | D$Border­line %in% m  | D$Bipo­lar %in% m
  
  m ← men­tally_ill_loose
  D$Men­tal­lyIll_2 = D$De­pres­sion %in% m | D$Anx­iety %in% m | D$OCD %in% m  | D$Eat­ingdi­s­or­der %in% m  | D$PTSD %in% m  | D$Al­co­holism %in% m  | D$Dru­gad­dic­tion %in% m  | D$Border­line %in% m  | D$Bipo­lar %in% m
  
  sum­mary(lm(D$Men­tal­lyIll_1 ~ D$EAID))
  sum­mary(glm(D$Men­tal­lyIll_1 ~ D$EAID, fam­ily = “bino­mial”)) ## This is a lo­gis­tic re­gres­sion
  
  sum­mary(lm(D$Men­tal­lyIll_1 ~ D$EAID))
  sum­mary(glm(D$Men­tal­lyIll_1 ~ D$EAID, fam­ily = “bino­mial”)) ## This is a lo­gis­tic re­gres­sion
  

  And the re­sults are

  Call:
  lm(for­mula = D$Men­tal­lyIll_1 ~ D$EAID)
  
  Resi­d­u­als:
      Min      1Q  Me­dian      3Q     Max 
  −0.5256 −0.5151  0.4744  0.4849  0.5325 
  
  Coeffi­cients:
              Es­ti­mate Std. Er­ror t value Pr(>|t|)    
  (In­ter­cept)  0.46753    0.04027  11.609   <2e-16 ***
  D$EAIDNo     0.04757    0.04109   1.158    0.247    
  D$EAIDSorta  0.05811    0.04154   1.399    0.162    
  D$EAIDYes    0.03499    0.04328   0.808    0.419    
  ---
  Sig­nif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
  
  Resi­d­ual stan­dard er­ror: 0.4998 on 7335 de­grees of free­dom
  Mul­ti­ple R-squared:  0.0004201,	Ad­justed R-squared:  1.124e-05 
  F-statis­tic: 1.027 on 3 and 7335 DF,  p-value: 0.3792
  
  > sum­mary(glm(D$Men­tal­lyIll_1 ~ D$EAID, fam­ily = “bino­mial”)) ## This is a lo­gis­tic re­gres­sion
  
  Call:
  glm(for­mula = D$Men­tal­lyIll_1 ~ D$EAID, fam­ily = “bino­mial”)
  
  De­viance Resi­d­u­als: 
     Min      1Q  Me­dian      3Q     Max  
  −1.221  −1.203   1.134   1.152   1.233  
  
  Coeffi­cients:
              Es­ti­mate Std. Er­ror z value Pr(>|z|)
  (In­ter­cept)  −0.1301     0.1615  −0.805    0.421
  D$EAIDNo      0.1905     0.1648   1.156    0.248
  D$EAIDSorta   0.2327     0.1666   1.397    0.162
  D$EAIDYes     0.1401     0.1735   0.807    0.419
  
  (Disper­sion pa­ram­e­ter for bino­mial fam­ily taken to be 1)
  
      Null de­viance: 10167  on 7338  de­grees of free­dom
  Resi­d­ual de­viance: 10164  on 7335  de­grees of free­dom
  AIC: 10172
  
  Num­ber of Fisher Scor­ing iter­a­tions: 3
  
  > sum­mary(lm(D$Men­tal­lyIll_1 ~ D$EAID))
  
  Call:
  lm(for­mula = D$Men­tal­lyIll_1 ~ D$EAID)
  
  Resi­d­u­als:
      Min      1Q  Me­dian      3Q     Max 
  −0.5256 −0.5151  0.4744  0.4849  0.5325 
  
  Coeffi­cients:
              Es­ti­mate Std. Er­ror t value Pr(>|t|)    
  (In­ter­cept)  0.46753    0.04027  11.609   <2e-16 ***
  D$EAIDNo     0.04757    0.04109   1.158    0.247    
  D$EAIDSorta  0.05811    0.04154   1.399    0.162    
  D$EAIDYes    0.03499    0.04328   0.808    0.419    
  ---
  Sig­nif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
  
  Resi­d­ual stan­dard er­ror: 0.4998 on 7335 de­grees of free­dom
  Mul­ti­ple R-squared:  0.0004201,	Ad­justed R-squared:  1.124e-05 
  F-statis­tic: 1.027 on 3 and 7335 DF,  p-value: 0.3792
  
  > sum­mary(glm(D$Men­tal­lyIll_1 ~ D$EAID, fam­ily = “bino­mial”)) ## This is a lo­gis­tic re­gres­sion
  
  Call:
  glm(for­mula = D$Men­tal­lyIll_1 ~ D$EAID, fam­ily = “bino­mial”)
  
  De­viance Resi­d­u­als: 
     Min      1Q  Me­dian      3Q     Max  
  −1.221  −1.203   1.134   1.152   1.233  
  
  Coeffi­cients:
              Es­ti­mate Std. Er­ror z value Pr(>|z|)
  (In­ter­cept)  −0.1301     0.1615  −0.805    0.421
  D$EAIDNo      0.1905     0.1648   1.156    0.248
  D$EAIDSorta   0.2327     0.1666   1.397    0.162
  D$EAIDYes     0.1401     0.1735   0.807    0.419
  
  (Disper­sion pa­ram­e­ter for bino­mial fam­ily taken to be 1)
  
      Null de­viance: 10167  on 7338  de­grees of free­dom
  Resi­d­ual de­viance: 10164  on 7335  de­grees of free­dom
  AIC: 10172
  
  Num­ber of Fisher Scor­ing iter­a­tions: 3
  
  

• NunoSempere 30 Jan 2020 10:04 UTC

  4 points

  in reply to: CEEALAR's comment on: The EA Ho­tel is now the Cen­tre for En­abling EA Learn­ing & Re­search (CEEALAR)

  Cen­ter for As­sist­ing EA Study And Re­search: CAESAR.

  As­sist­ing ~ Aid­ing ~ Abetting

• NunoSempere 23 Jan 2020 14:08 UTC

  3 points

  on: AMA: Rob Mather, founder and CEO of the Against Malaria Foundation

  I started some­what by ac­ci­dent...

  Would you give a guess as to the prob­a­bil­ity you would have as­signed to end­ing up do­ing some­thing like this be­fore you in fact did? How un­likely was this class of out­comes?

• NunoSempere 17 Jan 2020 10:06 UTC

  14 points

  on: EA in­ter­est: Google Trends

  [The images don’t load for me. What I usu­ally do in these cases is to up­load the image to github, and then add it to the EA fo­rum with ![](imageURL)]

A glow­ing re­view of two free on­line MIT Global Poverty courses

• NunoSempere 19 Dec 2019 21:22 UTC

  4 points

  in reply to: jacobjacob's comment on: Why do so­cial move­ments fail: Two con­crete ex­am­ples.

  Cu­ri­ously enough, ’twas me who fulfilled the bounty.

[Part 1] Am­plify­ing gen­er­al­ist re­search via fore­cast­ing – mod­els of im­pact and challenges

[Part 2] Am­plify­ing gen­er­al­ist re­search via fore­cast­ing – re­sults from a pre­limi­nary exploration

• NunoSempere 13 Dec 2019 10:56 UTC

  13 points

  on: Opinion: Es­ti­mat­ing In­ver­te­brate Sentience

  You could also con­sider pro­vid­ing your prob­a­bil­ities in the form of dis­tri­bu­tions, for ex­am­ple by an­swer­ing a ques­tion like “What prob­a­bil­ity would you as­sign to trib­bles [1] be­ing sen­tient af­ter 1000 more hours of re­search?” This would per­haps solve part of the prob­lem of com­mu­ni­cat­ing the un­cer­tainty which you want to com­mu­ni­cate.

  Some ex­am­ple an­swers might be:

  

  

  

  In the first one your un­cer­tainty is sig­nifi­cant. You have a prob­a­bil­ity of ~45%, but you con­sider it likely that you will up­date a lot. You just don’t know in which di­rec­tion: You con­sider it equally likely that you will end up at 30% or at 70%.

  In the sec­ond one, re­search has mostly been done and you’ve already mostly made up your mind. You have a prob­a­bil­ity of ~45%, and you be­lieve that fur­ther re­search is most likely to move you to 42%, or to 47%, but not that much fur­ther. You’d be shocked if you ended up with a prob­a­bil­ity of more than 60%.

  The third one is the dis­tri­bu­tional equiv­a­lent of you mak­ing a shrug. Your prob­a­bil­ity is 45%, but re­ally, trib­bles be­ing so un­der­ex­plored means that your dis­tri­bu­tion looks pretty much uniform.

  Note that it’s even pos­si­ble that all three dis­tri­bu­tions have the same mean (~45%). This would mean that in all three cases you’d think that a bet of 45:55 would be a fair bet (that is, that it has an ex­pected value of 0).

  You could also pre­dict the stan­dard de­vi­a­tion of your dis­tri­bu­tion (how broad it is) af­ter 1000h of re­search, 2000h, 5000h, etc., ag­gre­gate the dis­tri­bu­tions of all your re­searchers, and do other nice things.

  [1] Fic­tional species.

  What links here?

  • MichaelA's comment on MichaelA’s Shortform by MichaelA (LessWrong; 16 Jan 2020 11:33 UTC; 7 points)

• NunoSempere 9 Dec 2019 14:00 UTC

  4 points

  on: EA Lead­ers Fo­rum: Sur­vey on EA pri­ori­ties (data and anal­y­sis)

  Hey, re­gard­ing your ques­tion “What types of tal­ent do you cur­rently think [your or­ga­ni­za­tion //​ EA as a whole] will need more of over the next 5 years? (Pick up to 6)”, I think you might want to word it some­what differ­ently, and per­haps dis­am­biguate be­tween:

  • On a scale 0-10, how much [X] tal­ent will [EA/​your or­ga­ni­za­tion] need over the next 5 years?”.

  • On a scale 0-10, how much will the need for [X] tal­ent in [EA/​your or­ga­ni­za­tion] in­crease over the next 5 years?”.

  This mainly has the ad­van­tage of al­low­ing for more gran­u­lar com­par­i­son. For ex­am­ple, maybe man­age­ment is always a solid 10, whereas gov­ern­ment ex­per­tise is a solid 7, but both always fall in the top 6, and the differ­ence might some­times be im­por­tant. I also think that the sec­ond word­ing is some­what eas­ier to read /​ re­quires less cog­ni­tive la­bor.

  Two book recom­men­da­tions are The Power of Sur­vey De­sign: A User’s Guide for Manag­ing Sur­veys, In­ter­pret­ing Re­sults, and In­fluenc­ing Re­spon­dents and Im­prov­ing sur­vey ques­tions—De­sign and Eval­u­a­tion. I should have a short re­view/​sum­mary and some check­lists some­where, if you’re in­ter­ested.

• NunoSempere 2 Dec 2019 9:30 UTC

  1 point

  on: How do we cre­ate in­ter­na­tional sup­ply chain ac­countabil­ity?

  I re­searched this at some point; will try to write some­thing up this week, though I might do so in a sep­a­rate post. Thanks for ask­ing the ques­tion!

• NunoSempere 2 Dec 2019 9:27 UTC

  1 point

  in reply to: Aaron Gertler's comment on: How do we cre­ate in­ter­na­tional sup­ply chain ac­countabil­ity?

  I dis­agree; I see “sup­ply chain ac­countabil­ity” as a very spe­cific, al­most tech­ni­cal term (the word in­ter­na­tional is a lit­tle bit re­dun­dant, though), which con­tains enough in­for­ma­tion for me to figure out what it means.

• NunoSempere 16 Nov 2019 8:12 UTC

  19 points

  on: I’m Buck Sh­legeris, I do re­search and out­reach at MIRI, AMA

  What would you be work­ing on if you were work­ing on some­thing else?

• NunoSempere 16 Nov 2019 8:12 UTC

  14 points

  on: I’m Buck Sh­legeris, I do re­search and out­reach at MIRI, AMA

  Do you have the in­tu­ition that a ran­dom gifted per­son can con­tribute to tech­ni­cal re­search on AI safety?

• NunoSempere 31 Oct 2019 14:06 UTC

  3 points

  in reply to: casebash's comment on: Shap­ley val­ues: Bet­ter than counterfactuals

  Hey Chris! It was nice see­ing you at the EA Ho­tel, and I’m glad we could talk about this. I’m writ­ing down some of my notes from our con­ver­sa­tions. Is there any­thing I’ve for­got­ten, or which you’d like to add?

  a. What are you us­ing Shap­ley val­ues /​ coun­ter­fac­tual val­ues for?

  You might want to use differ­ent tools de­pend­ing on what your goal is; three differ­ent goals migh be: Co­or­di­na­tion /​ Anal­y­sis /​ Re­ward /​ Award.

  For ex­am­ple, you might want a func­tion which is eas­ier to un­der­stand when an­nounc­ing an award. If you’re re­ward­ing a be­hav­ior, you might want to make sure you’re in­cen­tiviz­ing the right thing.

  b. The prob­lem of choos­ing who to count is more com­pli­cated than I origi­nally thought, and you should in fact ex­clude some agents from your calcu­la­tions.

  The ex­am­ple of: “If a bus driver falls off a cliff and Su­per­man res­cues them and brings them safely to their des­ti­na­tion, ear­lier, the bus driver gets half the credit” is silly, but made the thing re­ally crisp for me.

  Hear­ing that, we then thought that:

  • Yes, the driver gets half the credit un­der Shap­ley val­ues, but the same value as Su­per­man un­der coun­ter­fac­tual value.

  • (also, if the driver dis­tracts Su­per­man from sav­ing a differ­ent bus, then the driver gets 0 or nega­tive value in both cases)

  • (if the driver was in­tel­li­gent enough to know that Su­per­man wasn’t do­ing any­thing im­por­tant, he might ac­tu­ally get half the credit, but only of get­ting there ear­lier. In this sce­nario, had there been no Su­per­man, the driver wouldn’t have fallen off the cliff.).

  • (if the driver was a pa­per­clip max­i­mizer who didn’t know that Su­per­man was go­ing to be around, then Su­per­man should take all the credit).

  So the an­swer would seem to be some­thing like: -Count­ing only over peo­ple who are broadly similar to you?

  • Who are op­ti­miz­ing over the same thing, or whose de­ci­sions can be changed be­cause of yours? It seems like this is more of a case of causal, rather than sub­junc­tive de­pen­dence.

  c. Shap­ley val­ues and uncertainty

  How do SVs deal with un­cer­tainty? Can you do ex­pected value over SVs? [Yes, you can]. For ex­am­ple, if you have a 1% chance of a SV of 100, you can say that the E[SV] = 1. Even thought the SV for­mal­ism is more com­pli­cated than the coun­ter­fac­tual, it still works el­e­gantly /​ is well-defined, etc.

• NunoSempere 28 Oct 2019 23:27 UTC

  2 points

  in reply to: Stefan_Schubert's comment on: At­tempt at un­der­stand­ing the role of moral philos­o­phy in moral progress

  Re­gard­ing “to what ex­tent have in­di­vi­d­u­als in­fluenced his­tory?”, I have found Great Founder The­ory, valuable, but it doesn’t an­swer that ex­act ques­tion.

• NunoSempere 15 Oct 2019 7:37 UTC

  2 points

  in reply to: ishi's comment on: Shap­ley val­ues: Bet­ter than counterfactuals

  1. would sur­prise me; can you think of a source?

• NunoSempere 11 Oct 2019 10:15 UTC

  2 points

  in reply to: casebash's comment on: Shap­ley val­ues: Bet­ter than counterfactuals

  Fair point re: un­cer­tainty. The situ­a­tion seems pretty sym­met­ric, though: if a poli­ti­cian builds roads just to get votes, and an NGO steps in and does some­thing valuable with that, the poli­ti­cian’s coun­ter­fac­tual im­pact is still the same as the NGO’s, so both the Shap­ley value and coun­ter­fac­tu­als have that prob­lem (?). Maybe one can ex­clude agents acord­ing to how close their goals are to yours, e.g., to­tally ex­clude a pa­per­clip max­i­mizer from both coun­ter­fac­tual and Shap­ley value calcu­la­tions, and ap­ply or­der in­differ­ence to al­lies only (?). This is some­thing I haven’t though about; thanks for point­ing it out.

  Fair point re: epistemic sta­tus. Changed my epistemic sta­tus.

• NunoSempere 11 Oct 2019 10:08 UTC

  1 point

  in reply to: Halstead's comment on: Shap­ley val­ues: Bet­ter than counterfactuals

  I don’t ex­actly claiming to have iden­ti­fied a prob­lem with the coun­ter­fac­tual func­tion, in it­self. The coun­ter­fac­tual is perfectly well defined, and I like it, and it has done noth­ing wrong. I un­der­stand this. It is clear to me that it can’t be added just like that. The func­tion, per se, is fine.

  What I’m claiming is that, be­cause it can’t be ag­gre­gated, it is not the right func­tion to think about in terms of as­sign­ing im­pact to peo­ple in the con­text of groups. I am ar­gu­ing about the area of ap­pli­ca­bil­ity of the func­tion, not about the func­tion. I am claiming that, if you are op­ti­miz­ing for coun­ter­fac­tual im­pact in terms of groups, pit­falls may arise.

  It’s like, when you first see for the same time: −1 = sqrt(-1)*sqrt(-1) = sqrt((-1)*(-1)) = sqrt(1) = 1, there­fore −1 = 1, and you can’t see the mis­take. It’s not that the sqrt func­tion is wrong, it’s that you’re us­ing it out­side it’s limited fief­dom, so some­thing breaks. I hope the ex­am­ple proved amus­ing.

  I’m not only mak­ing state­ments about the coun­ter­fac­tual func­tion, I’m also mak­ing state­ments about the con­cept which peo­ple have in your head which is called “im­pact”, and how that con­cept doesn’t map to coun­ter­fac­tual im­pact some of the time, and about how, if you had to map that con­cept to a math­e­mat­i­cal func­tion, the Shap­ley value is a bet­ter can­di­date.