Ozzie Gooen(Ozzie Gooen)

[Question] How to es­ti­mate the EV of gen­eral in­tel­lec­tual progress

• Ozzie Gooen 23 Jan 2020 10:36 UTC

  17 points

  on: Do­ing good is as good as it ever was

  I think that per­son­ally, I’d mostly ad­vo­cate for at­tempts to de­cou­ple mo­ti­va­tion from to­tal im­pact mag­ni­tude, rather for at­tempts to ar­gue that high im­pact mag­ni­tude is achiev­able, so to speak, when try­ing to im­prove mo­ti­va­tion.

  If you at­tach your mo­ti­va­tions to a spe­cific mag­ni­tude of “$2,000 per life saved”, then you can ex­pect them to fluc­tu­ate heav­ily when es­ti­mates change. But ideally, you would want your mo­ti­va­tions to stay pretty op­ti­mal and thus con­sis­tent for your goals. I think this ideal is some­what pos­si­ble and can be worked to­wards.

  The main goal of a con­se­quen­tial­ist should be to op­ti­mize a util­ity func­tion, it re­ally shouldn’t mat­ter what the spe­cific mag­ni­tudes are. If the great­est thing I could do with my life is to keep a small room clean, then I should spend my great­est effort on that thing (my own wellbe­ing aside).

  I think that most peo­ple aren’t ini­tially com­fortable with re-cal­ibrat­ing their goals for ar­bi­trary util­ity func­tion mag­ni­tudes, but I think that learn­ing to do so is a grad­ual pro­cess that could be learned, similar to learn­ing stoic philos­o­phy.

  It’s similar to learn­ing how to be con­tent no mat­ter one’s con­di­tions (aside from ex­treme phys­i­cal ones), as dis­cussed in The Myth of Sisy­phus.

  https://​​en.wikipe­dia.org/​​wiki/​​The_Myth_of_Sisyphus

• Ozzie Gooen 23 Jan 2020 9:35 UTC

  2 points

  in reply to: MichaelA's comment on: Po­ten­tial down­sides of us­ing ex­plicit probabilities

  I think that makes sense. Some of it is a mat­ter of in­ter­pre­ta­tion.

  From one per­spec­tive, the op­ti­mizer’s curse is a dra­matic and challeng­ing dilemma fac­ing mod­ern anal­y­sis. From an­other per­spec­tive, it’s a rather ob­vi­ous and sim­ple ar­ti­fact from poorly-done es­ti­mates.

  I.E. they some­times say that if mathamat­i­ci­ans re­al­ize some­thing is pos­si­ble, they con­sider the prob­lem triv­ial. Here the op­ti­mizer’s curse is con­sid­ered a rea­son­ably-well-un­der­stood phe­nom­ena, un­like some other es­ti­ma­tion-the­ory ques­tions cur­rently be­ing faced.

• Ozzie Gooen 21 Jan 2020 22:55 UTC

  3 points

  on: Po­ten­tial down­sides of us­ing ex­plicit probabilities

  And I think that, even when one is ex­tremely un­cer­tain, the op­ti­mizer’s curse doesn’t mean you should change your prefer­ence or­der­ing (just that you should be far less cer­tain about it, as you’re prob­a­bly greatly­over­es­ti­mat­ing the value of best-seem­ing op­tion).

  Ok, I’ll flag this too. I’m sure there are statis­ti­cal situ­a­tions where an ex­treme out­come im­plies that an ad­just­ment for cor­re­la­tion good­hart­ing would make it seem worse than other op­tions; i.e. change or­der.

  That said, I’d guess this isn’t likely to hap­pen that of­ten for re­al­is­tic cases, es­pe­cially when there aren’t highly ex­treme out­liers (which, to be fair, we do have with EA).

  I think one mis­take some­one could make here would be to say that be­cause the or­der­ing may be pre­served, the prob­lem wouldn’t be “fixed” at all. But, the un­cer­tain­ties and re­la­tion­ships them­selves are of­ten use­ful in­for­ma­tion out­side of or­der­ing. So a nat­u­ral con­clu­sion in the case of in­tense noise (which leads to the op­ti­mizer’s curse) would be to ac­cept a large amount of un­cer­tainty, and maybe use that knowl­edge to be more con­ser­va­tive; for in­stance, try­ing to get more data be­fore go­ing all-in on any­thing in par­tic­u­lar.

• Ozzie Gooen 21 Jan 2020 20:37 UTC

  7 points

  on: What are words, phrases, or top­ics that you think most EAs don’t know about but should?

  The Co­op­er­a­tive Principle

  The co­op­er­a­tive prin­ci­ple de­scribes how peo­ple achieve effec­tive con­ver­sa­tional com­mu­ni­ca­tion in com­mon so­cial situ­a­tions—that is, how listen­ers and speak­ers act co­op­er­a­tively and mu­tu­ally ac­cept one an­other to be un­der­stood in a par­tic­u­lar way.

  There are 4 cor­re­spond­ing max­ims. I think the main non-ob­vi­ous ones are:

  Maxim of quan­tity:

  1. Make your con­tri­bu­tion as in­for­ma­tive as is re­quired (for the cur­rent pur­poses of the ex­change).

  2. Do not make your con­tri­bu­tion more in­for­ma­tive than is re­quired.

  Maxim of relevance

  1. Be rele­vant to the dis­cus­sion. (For in­stance, when re­spond­ing to, “What would you like for lunch” and you re­spond “I would like a sand­which”; you are ex­pected to be re­spond­ing to that very ques­tion, not to be mak­ing an un­re­lated state­ment.)

  I think this video ex­plains this well.

  Why this is in­ter­est­ing
  I’ve definitely been in con­ver­sa­tions where bring­ing up max­ims of quan­tity and rele­vance would have been use­ful to bring up. Con­ver­sa­tion and dis­cus­sion can be quite difficult. We do a lot of that.

• Ozzie Gooen 21 Jan 2020 20:30 UTC

  16 points

  on: What are words, phrases, or top­ics that you think most EAs don’t know about but should?

  Consilience

  The prin­ci­ple that ev­i­dence from in­de­pen­dent, un­re­lated sources can “con­verge” on strong conclusions

  This word can ar­guably be used to de­scribe the “Many Weak Ar­gu­ments” as­pect of the “Many Weak Ar­gu­ments vs. One Rel­a­tively Strong Ar­gu­ment” post. Jon­ahSinick pointed that out in that post.

  Why this is in­ter­est­ing
  Con­silience is im­por­tant for eval­u­at­ing claims. There’s a fair bit of his­toric dis­cus­sion and ev­i­dence now that shows how use­ful it can be to get a va­ri­ety of ev­i­dence from many differ­ent sources.

• Ozzie Gooen 21 Jan 2020 20:22 UTC

  5 points

  on: What are words, phrases, or top­ics that you think most EAs don’t know about but should?

  Deepity

  The term refers to a state­ment that is ap­par­ently profound but ac­tu­ally as­serts a triv­ial­ity on one level and some­thing mean­ingless on an­other. Gen­er­ally, a deep­ity has (at least) two mean­ings: one that is true but triv­ial, and an­other that sounds profound, but is es­sen­tially false or mean­ingless and would be “earth-shat­ter­ing” if true.

  Why this is in­ter­est­ing
  I mostly think this is just a great phrase to de­scribe a lot of difficult lan­guage I oc­ca­sion­ally see get used in moral dis­cus­sions.

• Ozzie Gooen 21 Jan 2020 20:18 UTC

  7 points

  on: What are words, phrases, or top­ics that you think most EAs don’t know about but should?

  No­bel Cause Corruption

  From Wikipe­dia:

  Noble cause cor­rup­tion is cor­rup­tion caused by the ad­her­ence to a tele­olog­i­cal eth­i­cal sys­tem, sug­gest­ing that peo­ple will use un­eth­i­cal or ille­gal means to at­tain de­sir­able goals, a re­sult which ap­pears to benefit the greater good. Where tra­di­tional cor­rup­tion is defined by per­sonal gain, no­ble cause cor­rup­tion forms when some­one is con­vinced of their righ­teous­ness, and will do any­thing within their pow­ers to achieve the de­sired re­sult. An ex­am­ple of no­ble cause cor­rup­tion is po­lice mis­con­duct “com­mit­ted in the name of good ends” or ne­glect of due pro­cess through “a moral com­mit­ment to make the world a safer place to live.”

  Why this is in­ter­est­ing
  I think one se­ri­ous con­cern around con­se­quen­tial­ist thought is that it can be used in dan­ger­ous ways. I think this term de­scribes some of this, and the cor­re­spond­ing liter­a­ture pro­vides ex­am­ples that seem similar to what I can ex­pect fu­ture peo­ple to fol­low who mi­suse EA con­tent.

[Question] What are words, phrases, or top­ics that you think most EAs don’t know about but should?

• Ozzie Gooen 21 Jan 2020 11:49 UTC

  3 points

  in reply to: MichaelA's comment on: Po­ten­tial down­sides of us­ing ex­plicit probabilities

  Hm… Some of this would take a lot more writ­ing than would make sense in a blog post.

  On over­con­fi­dence in prob­a­bil­ities vs. in­tu­itions: I think I mostly agree with you. One cool thing about prob­a­bil­ities is that they can be much more straight­for­wardly ver­ified/​falsified and mea­sured us­ing met­rics for cal­ibra­tion. If we had much larger sys­tems, I be­lieve we could do a great deal of work to bet­ter en­sure cal­ibra­tion with defined prob­a­bil­ities.

  “should even­tu­ally be used for most things”

  I’m not say­ing that hu­mans should come up with unique prob­a­bil­ities for most things on most days. One ex­am­ple I’d con­sider “used for most things” is a case where an AI uses prob­a­bil­ities to tell hu­mans which ac­tions seem the best, and hu­mans go with what the AI states. Similar could be said for “a trusted com­mit­tee” that uses prob­a­bil­ities as an in-be­tween.

  “we could learn to get much bet­ter than them later”

  I think there are strong claims that top­ics like Bayes, Causal­ity, Ra­tion­al­ity even, are still rel­a­tively poorly un­der­stood, and may be ad­vanced a lot in the next 30-100 years. As we get bet­ter with them, I pre­dict we would get bet­ter at for­mal mod­el­ing.

  I re­ject ex­pected util­ity the­ory (and re­lated stances such as cost-benefit anal­y­sis), at least if it comes as a for­mal way of spel­ling out a max­i­miz­ing con­se­quen­tial­ist moral stance which does not prop­erly in­cor­po­rate rights.

  This is a com­pli­cated topic. It think a lot of Utili­tar­i­ans/​Con­se­quen­tial­ists wouldn’t deem many in­ter­pre­ta­tions of rights as meta­phys­i­cal or ter­mi­nally-valuable things. Another way to look at it would be to at­tempt to map the rights to a util­ity func­tion. Utility func­tions re­quire very, very few con­di­tions. I’m per­son­ally a bit cyn­i­cal of val­ues that can’t be mapped to util­ity func­tions, if even in a highly-un­cer­tain way.

  But Chris Smith quotes the pro­posed solu­tion, and then writes… It’s clear Chris Smith has thought about some of this topic a fair bit, but my im­pres­sion is that I dis­agree with him. It’s quite pos­si­ble that much of the dis­agree­ment is se­man­tic; where he says ‘this solu­tion is un­work­able’ I may say, ‘the solu­tion re­sults in a very wide amount of un­cer­tainty’. I think it’s clear to ev­ery­one (the main re­searchers any­way) that there’s lit­tle data about many of these top­ics, and that Bayesian or any kind of statis­ti­cal ma­nipu­la­tions can’t fun­da­men­tally con­vert “very lit­tle data” into “a great deal of con­fi­dence”.

  Ku­dos for iden­ti­fing that post. The main solu­tion I was refer­ring to was the one de­scribed in the sec­ond com­ment:

  In statis­tics the solu­tion you de­scribe is called Hier­ar­chi­cal or Mul­tilevel Model­ing. You as­sume that you data is drawn from a set of dis­tri­bu­tions which have their pa­ram­e­ters drawn from an­other dis­tri­bu­tion. This au­to­mat­i­cally shrinks your es­ti­mates of the dis­tri­bu­tions to­wards the mean. I think it’s a pretty use­ful trick to know and I think it would be good to do a writeup but I think you might need to have a de­cent grasp of bayesian statis­tics first.

  The op­ti­mizer’s curse ar­guably is ba­si­cally within the class of Good­hart-like prob­lems https://​​www.less­wrong.com/​​posts/​​5bd75cc58225bf06703754b2/​​the-three-lev­els-of-good­hart-s-curse

  I’m not say­ing that these are easy to solve, but rather, there is a math­e­mat­i­cal strat­egy to gen­er­ally fix them in ways that would make sense in­tu­itively. There’s no bet­ter ap­proach than to try to ap­prox­i­mate the math­e­mat­i­cal ap­proach, or go with an ap­proach that in-ex­pec­ta­tion does a de­cent job at ap­prox­i­mat­ing the math­e­mat­i­cal ap­proach.

• Ozzie Gooen 21 Jan 2020 10:58 UTC

  3 points

  in reply to: RomeoStevens's comment on: Love seems like a high priority

  Just want to sec­ond this, I think it’s a pretty well-known is­sue in the in­dus­try. Dat­ing apps that do in­cred­ibly well at set­ting up peo­ple on dates will get lit­tle use (be­cause they typ­i­cally charge monthly rates, and will get to charge fewer months if the cus­tomer is happy and leaves quickly). It’s pos­si­bly a very large mar­ket failure.

  I could imag­ine a hy­po­thet­i­cal app that was able to charge a lot more ini­tially (Like, $2,000) but did a much bet­ter job. Of course, one is­sue with this is that these apps need a lot of users, so this could be re­ally difficult.

  I think it could be pos­si­ble to figure out a solu­tion here, but I imag­ine the solu­tion may be 2/​3rds pay­ment/​eco­nomic-in­no­va­tion.

  Per­haps an idea solu­tion would look some­thing like a mix be­tween per­sonal guidance and on­line sup­port.

• Ozzie Gooen 20 Jan 2020 14:22 UTC

  7 points

  on: Po­ten­tial down­sides of us­ing ex­plicit probabilities

  Ku­dos for this write-up, and for your many other posts (both here and on LessWrong, it seems) on un­cer­tainty.

  Over­all, I’m very much in the “Prob­a­bil­ities are pretty great and should even­tu­ally be used for most things” camp. That said, I think the “Scout vs. Soldier” mind­set is use­ful, so to speak; in­ves­ti­gat­ing both sides is pretty use­ful. I’d definitely as­sign some prob­a­bil­ity to be­ing wrong here.

  My im­pres­sion is that we’re prob­a­bly in broad agree­ment here.

  Some quick points that come to mind:

  1. The de­bate on “are ex­plicit prob­a­bil­ities use­ful” is very similar to those of “are met­rics use­ful”, “are cost-benefit analy­ses use­ful”, and “is con­se­quen­tial­ist rea­son­ing use­ful.” I ex­pect that there’s broad cor­re­la­tion be­tween those who agree/​dis­agree with these.

  2. In cases where prob­a­bil­ities are ex­pected to be harm­fully, hope­fully prob­a­bil­ities could be used to tell us as such. Like, we could pre­dict that ex­plicit and pub­lic use would be harm­ful.

  3. I’d definitely agree that it’s very pos­si­ble to use prob­a­bil­ities poorly. I think a lot of Holden’s crit­i­cisms here would fall into this camp. Neu­ral Nets for a while were hon­estly quite poor, but thank­fully that didn’t lead to sci­en­tists aban­don­ing those. I think prob­a­bil­ities are a lot bet­ter now, but we could learn to get much bet­ter than them later. I’m not sure how we can get much bet­ter with­out them.

  4. The op­ti­mizer’s curse can be ad­justed for with rea­son­able use of Bayes. Bayesian hi­er­ar­chi­cal mod­els should deal with it quite well. There’s been some dis­cus­sion of this around “Good­hart” on LessWrong.

• Ozzie Gooen 19 Jan 2020 22:26 UTC

  6 points

  on: Love seems like a high priority

  I think one pos­si­ble point against tractabil­ity is that many coun­tries have been try­ing to work on similar things as a means to in­crease pop­u­la­tion.

  How­ever, the flip side to this is that if you were to come up with a clever in­ter­ven­tion, these coun­tries could act as large non-EA buy­ers.

  I re­mem­ber see­ing some other writ­ing on dat­ing apps in gen­eral. I think among smart tech peo­ple I know there is some agree­ment that dat­ing apps could be im­proved a whole lot. That said, it is difficult to make money from them, so it’s not in­cred­ibly ex­cit­ing as a busi­ness ven­ture.

  Some pos­si­bly rele­vant links:

  https://​​www.quora.com/​​Are-there-coun­tries-which-en­courage-their-pop­u­la­tion-to-increase

  https://​​www.economist.com/​​asia/​​2019/​​10/​​03/​​japans-state-owned-ver­sion-of-tinder

  https://​​www.busi­ness­in­sider.com/​​japanese-gov­ern­ment-dat­ing-ser­vices-2016-10?r=US&IR=T

  https://​​www.vox.com/​​2016/​​5/​​23/​​11440638/​​child-benefit-child-allowance

  https://​​www.men­talfloss.com/​​ar­ti­cle/​​33485/​​6-cre­ative-ways-coun­tries-have-tried-their-birth-rates

  https://​​www.dw.com/​​en/​​france-moves-to-en­courage-large-fam­i­lies/​​a-1720921

• Ozzie Gooen 19 Jan 2020 20:31 UTC

  19 points

  on: Love seems like a high priority

  Ku­dos for se­ri­ously recom­mend­ing a novel-to-EA-for-me in­ter­ven­tion! It seems there have been a few rel­a­tively novel ideas posted re­cently, it’s been quite nice. I think the base rate for these sorts of in­ter­ven­tions mak­ing it through all the next filters (much more EA con­sid­er­a­tion, ac­tual im­ple­men­ta­tion) is quite low, and I don’t have par­tic­u­lar rea­son to cur­rently be­lieve this is much differ­ent than other recom­men­da­tions. That said, even if there were a 0.1% chance this were cost-effec­tive, the EV of in­tro­duc­ing it could be quite high.

• Ozzie Gooen 17 Jan 2020 21:24 UTC

  4 points

  in reply to: aarongertler's comment on: The ‘sprint­ing be­tween oases’ strategy

  I think I like the idea of more ab­stract posts be­ing on the EA Fo­rum, es­pe­cially if the main in­tended even­tual use is straight­for­ward EA causes. Ar­guably, a whole lot of the in­ter­est­ing work to be done is kind of ab­stract.

  This spe­cific post seems to be some­what re­lated to global sta­bil­ity, from what I can tell?

  I’m not sure what the ideal split is be­tween this and LessWrong. I imag­ine that as time goes one we could do a bet­ter cluster anal­y­sis.

• Ozzie Gooen 16 Jan 2020 7:08 UTC

  20 points

  on: Space gov­er­nance is im­por­tant, tractable and neglected

  +1 for a se­ri­ous sug­ges­tion of an in­ter­ven­tion that to me is pretty novel to EA. I feel like we should be con­sid­er­ing a larger class of op­tions than we typ­i­cally have. Space gov­er­nance seems like a rea­son­able op­tion to con­sider to me.

• Ozzie Gooen 14 Jan 2020 20:16 UTC

  3 points

  on: Moloch and the Pareto op­ti­mal frontier

  I like the first bit, but am a bit con­fused on the Moloch bit. Why ex­actly would we ex­pect that it “max­i­mizes com­pet­i­tive­ness”?

[Question] Best units for com­par­ing per­sonal in­ter­ven­tions?

• Ozzie Gooen 13 Jan 2020 8:37 UTC

  2 points

  on: Disad­van­tages of the measures

  I guess one ques­tion is:

  Can we use meta-met­rics to tell us the ex­pected value of us­ing spe­cific met­rics?

  There are definitely times where spe­cific met­rics are EV-nega­tive, but I’d hope that can be de­ter­mined with other met­rics. (Like a fore­cast of the es­ti­mated value)

• Ozzie Gooen 13 Jan 2020 8:34 UTC

  16 points

  on: Should and do EA orgs con­sider the com­par­a­tive ad­van­tages of ap­pli­cants in hiring de­ci­sions?

  On Rejection

  I’d be quite sur­prised if re­jec­tion were fre­quent. I’ve been part of /​ close to a few or­ga­ni­za­tions and haven’t heard it dis­cussed. A few rea­sons:

  1. I think most orgs are over­con­fi­dent in them­selves com­pared to other or­ga­ni­za­tions (prob­a­bly due to se­lec­tion effects).

  2. I think most or­ga­ni­za­tions would pre­fer to give the ap­pli­cants the op­tion, rather than re­mov­ing it. They may tell the ap­pli­cant in ques­tion that they be­lieve it’s pos­si­ble that their job is less effec­tive, but that it’s still the ap­pli­cant’s de­ci­sion to choose.

  3. “It feels wrong” in similar ways that ly­ing or steal­ing feel wrong.

  4. I haven’t heard about this be­ing a fac­tor, but I would imag­ine there could be le­gal is­sues for re­ject­ing a can­di­date be­cause you want to effec­tively con­trol which group they work for.

  I think (2) is a pretty good rea­son, for hu­mil­ity rea­sons if noth­ing else. Ap­pli­cants may have a lot of good rea­sons for do­ing things that may at first seem sub-op­ti­mal. (3) is also pretty good. I could imag­ine in­stances where there’s one highly-anx­iety-pro­duc­ing but seem­ingly-effec­tive op­tion for a per­son. It seems kind of cruel if all their other job prospects re­fuse them to force them into that po­si­tion. I could imag­ine some pretty nasty de­ci­sion con­se­quences that could re­sult if that were a con­sid­er­a­tion.

  On Recom­men­da­tions to Other Orgs

  I have wit­nessed cases of or­ga­ni­za­tions sug­gest­ing good peo­ple to other or­ga­ni­za­tions, es­pe­cially if they both (1) think they were good, but (2) didn’t fit with their own ap­pli­ca­tion pro­cess. This seems pretty rea­son­able to me.

  Per­son­ally, when I chat to po­ten­tial hires of things I’m work­ing around, I try to be as hon­est as pos­si­ble re­gard­ing to which or­ga­ni­za­tions they would be best for, even some­times con­nect­ing them with peo­ple rep­re­sent­ing other or­ga­ni­za­tions.

  In some sense these or­gani­sa­tions are com­pet­ing, but in a big­ger sense, we’re all try­ing to help make sure the world goes OK.