Roko

• Roko 24 Oct 2020 16:33 UTC

  4 points

  in reply to: Liam Donovan’s comment on: Bet On Biden

  it doesn’t help that Trump un­ex­pect­edly won the 2016 elec­tion, trans­fer­ring a lot of money to bi­ased bet­tors.

  I mean they got the an­swer right so it seems a lit­tle ar­ro­gant to call them bi­ased. Maybe they just had a good heuris­tic to do with prefer­ence falsifi­ca­tion?

• Roko 24 Oct 2020 15:46 UTC

  2 points

  on: What is our true life ex­pec­tancy?

  The ex­pec­ta­tion is prob­a­bly around 1 billion:

  10% × 10 billion years (live roughly as long as the uni­verse has ex­isted for already) +

  90% × die within 1000 years (likely within 70 for most peo­ple here!)

  To­tal: 1 billion

  I am ig­nor­ing the in­finite pos­si­bil­ities since any finite sys­tem will start re­peat­ing con­figu­ra­tions even­tu­ally, so I don’t think that in­finite life for a hu­man even makes sense (you’d just be re­run­ning the life you already have and I don’t think that counts).

• Roko 18 Oct 2020 22:35 UTC

  6 points

  on: Have the lock­downs been worth it?

  A spread­sheet model I made in­ves­ti­gates the trade-off be­tween lost life-years from covid-19 deaths ver­sus lost life-years from re­duced qual­ity of life be­ing locked down. You can make a copy of it and play with the var­i­ous pa­ram­e­ters.

  The spread­sheet uses real data about the mor­tal­ity risk from covid-19 and the pop­u­la­tion struc­ture (life ex­pec­tancy, pop­u­la­tion pyra­mid) for the USA.

  With the pa­ram­e­ters that I chose, lock­downs of 1.25 years de­stroy about 0.25 life-years per per­son on net.

  https://​​docs.google.com/​​spread­sheets/​​d/​​1wBcHkt9i_4hGSXRrqurr82j-R_X03nyr_Eds6YLjkzc/​​

• Roko 15 Oct 2020 10:21 UTC

  7 points

  in reply to: Mark Xu’s comment on: The Solomonoff Prior is Malign

  adding ex­e­cu­tion/​mem­ory con­straints pe­nal­izes all hypothesis

  In re­al­ity these con­straints do ex­ist, so the ques­tion of “what hap­pens if you don’t care about effi­ciency at all?” is re­ally not im­por­tant. In prac­tice, effi­ciency is ab­solutely crit­i­cal and ev­ery­thing that hap­pens in AI is dom­i­nated by effi­ciency con­sid­er­a­tions.

  I think that mesa-op­ti­miza­tion will be a prob­lem. It prob­a­bly won’t look like aliens liv­ing in the Game of Life though.

  It’ll look like an in­ter­nal op­ti­mizer that just “de­cides” that the minds of the hu­mans who cre­ated it are an­other part of the en­vi­ron­ment to be op­ti­mized for its not-cor­rectly-al­igned goal.

• Roko 14 Oct 2020 18:32 UTC

  4 points

  on: Fermi Challenge: Trains and Air Cargo

  Re­lated to this ques­tion, I dis­cov­ered a new rule for Fermi es­ti­mates.

  If you want to es­ti­mate the mean value of a log­nor­mally dis­tributed ran­dom vari­able, giv­ing the mid­dle or­der of mag­ni­tude will be wrong as these are skewed.

  There is a sim­ple rule for get­ting this right that I dis­cov­ered: take your mid­dle or­der of mag­ni­tude (i.e. if you think it’s be­tween 10 and 100, the mid­dle is 10^1.5) and add 1.15 times the square of your es­ti­mate of the stan­dard de­vi­a­tion in log-space. So in this case that’s some­thing 1.5 +1.15 × 0.5^2. Then do 10 to the power that. This gives you about 60 - twice the an­swer you would have got­ten with the mid­dle or­der of mag­ni­tude.

  This ap­plies to things like “how big is a coun­try” and “how many miles of track is there for a coun­try of a given size” which might both be log­nor­mally dis­tributed.

  https://​​www.less­wrong.com/​​posts/​​LEn­tkjvDSxSt­dGN39/​​short­form?com­men­tId=soxf4Ynakr7aSjM3m

• Roko 14 Oct 2020 17:49 UTC

  6 points

  AF

  on: The Solomonoff Prior is Malign

  It seems to me that us­ing a com­bi­na­tion of ex­e­cu­tion time, mem­ory use and pro­gram length mostly kills this set of ar­gu­ments.

  Some­thing like a game-of-life ini­tial con­figu­ra­tion that leads to the even­tual evolu­tion of in­tel­li­gent game-of-life aliens who then strate­gi­cally feed out­puts into GoL in or­der to ma­nipu­late you may have very good com­plex­ity perfor­mance, but both the speed and mem­ory are go­ing to be pretty awful. The fixed cost in mem­ory and ex­e­cu­tion steps of es­sen­tially simu­lat­ing an en­tire uni­verse is huge.

  But yes, the pure com­plex­ity prior cer­tainly has some per­verse and un­set­tling prop­er­ties.

  EDIT: This is re­ally a spe­cial case of Mesa-Op­ti­miz­ers be­ing dan­ger­ous. (See, e.g. https://​​www.less­wrong.com/​​posts/​​XWPJfgBymBbL3jdFd/​​an-58-mesa-op­ti­miza­tion-what-it-is-and-why-we-should-care). The set of dan­ger­ous Mesa-Op­ti­miz­ers is ob­vi­ously big­ger than just “simu­lated aliens” and even time- and space-effi­cient al­gorithms might run into them.

• Roko 14 Oct 2020 17:30 UTC

  4 points

  on: Roko’s Shortform

  One weird trick for es­ti­mat­ing the ex­pec­ta­tion of Log­nor­mally dis­tributed ran­dom vari­ables:

  If you have a vari­able X that you think is some­where be­tween 1 and 100 and is Log­nor­mally dis­tributed, you can model it as be­ing a ran­dom vari­able with dis­tri­bu­tion ~ Log­nor­mal(1,1) - that is, the log­a­r­ithm has a dis­tri­bu­tion ~ Nor­mal(1,1).

  What is the ex­pec­ta­tion of X?

  Naively, you might say that since the ex­pec­ta­tion of log(X) is 1, the ex­pec­ta­tion of X is 10^1, or 10. That makes sense, 10 is at the mid­point of 1 and 100 on a log scale.

  This is wrong though. The chances of larger val­ues dom­i­nate the ex­pec­ta­tion or av­er­age of X.

  But how can you es­ti­mate that cor­rec­tion? It turns out that the rule you need is 10^(1 + 1.15*1^2) ≈ 141.

  In gen­eral, if X ~ Log­nor­mal(a, b) where we are work­ing to base 10 rather than base e, this is the rule you need:

  E(X) = 10^(a + 1.15*b^2)

  The 1.15 is ac­tu­ally ln(10)/​2.

  For a product of sev­eral in­de­pen­dent log­nor­mals, you can just mul­ti­ply these to­gether, which means adding in the ex­po­nent. If you have 2 or 3 things which are all log­nor­mal, the var­i­ance-as­so­ci­ated cor­rec­tions can eas­ily add up to quite a lot.

  Re­mem­ber: add 1.15 times the sum of squares of log-var­i­ances!

Roko’s Shortform

• Roko 14 Oct 2020 10:26 UTC

  2 points

  on: Fermi Challenge: Trains and Air Cargo

  A coun­try with trains might have the equiv­a­lent of 3-10 times the length of the coun­try worth of train tracks, and coun­tries are roughly 600-3000 miles across, with per­haps 10-100 coun­tries with a lot of trains that are also not far be­low the 600 mile size. The ge­o­met­ric means are ap­prox­i­mately 5.5vlength , 1350 miles, 30 coun­tries. Mul­ti­ply all of these and you get 222,750 miles.

  Q1 An­swer: 222,750 miles.

  Post-hoc edit: If you take these num­bers and do the right math on them, you get a more ac­cu­rate an­swer. The right math for this kind of thing is not just to mul­ti­ply the ge­o­met­ric means. You want to take 50 draws from a ran­dom vari­able that’s the product of size S times length mul­ti­plier L, which are log­nor­mally dis­tributed. If you do that you get about 600,000. If you do it for 30, you get 346,000. So, if you use the num­bers I gave with some ac­tual math, it’s 346,000. But I looked at the cor­rect an­swer be­fore de­cid­ing to bother to do the math.

  Air freight is per­haps 10%-30% of pas­sen­ger travel since I see more pas­sen­ger planes at air­ports than freight. Pas­sen­ger travel is eas­ier to es­ti­mate. Say that each of 1,000,000,000 peo­ple travel twice per year by air. 15 peo­ple plus their bag­gage is one met­ric ton. 2 × 17% × 1,000,000,000 /​ 15 = 22,000,000 met­ric tons.

  Q2 An­swer: 22,000,000 met­ric tons per year for both years.

• Roko 7 Jul 2020 8:57 UTC

  2 points

  in reply to: Davidmanheim’s comment on: A Per­sonal (In­terim) COVID-19 Postmortem

  But what about the ~3 months of lock­down and mas­sive Eco­nomic dis­rup­tion that we had to go through? Don’t you that that could have been avoided by clos­ing our bor­ders tightly in Jan­uary? Do we have ev­i­dence to ei­ther con­firm or ex­clude that now?

• Roko 6 Jul 2020 15:36 UTC

  2 points

  in reply to: Davidmanheim’s comment on: A Per­sonal (In­terim) COVID-19 Postmortem

  So do you think that the ac­tual travel re­stric­tions that hap­pened were just a waste of time, and we should have had fully open bor­ders?

  Or do you think that the re­stric­tions that we had (late and par­tial) were the op­ti­mal dis­ease-fight­ing policy (again, ne­glect­ing poli­ti­cal con­sid­er­a­tions)?

• Roko 5 Jul 2020 14:41 UTC

  2 points

  in reply to: Davidmanheim’s comment on: A Per­sonal (In­terim) COVID-19 Postmortem

  Yes it has always been clear to me that shut­ting down travel from one spe­cific place where you think the dis­ease is is not go­ing to work.

  Here in Europe, I have had to post­pone per­sonal va­ca­tions be­cause the bor­ders have in­deed been shut and are only now re­open­ing. Given that Europe ac­tu­ally did close its bor­ders, both ex­ter­nal and in­ter­nal, do you think that it would have been a good idea to go ahead and do that in Jan­uary, as­sum­ing away the poli­ti­cal prob­lem?

• Roko 4 Jul 2020 18:26 UTC

  11 points

  on: A Per­sonal (In­terim) COVID-19 Postmortem

  Good post David!

  I ap­plaud posts like this. It’s a great tra­di­tion that we do them.

  I be­lieve the con­ver­sa­tion about bor­der clo­sures was with me on Twit­ter?

  Are we at a stage yet where we can an­swer coun­ter­fac­tual ques­tions about what would have hap­pened if all in­ter­na­tional flights had been shut off in Jan­uary (As­sum­ing away the poli­ti­cal prob­lem)?

• Roko 24 May 2020 22:17 UTC

  4 points

  in reply to: Sønderjye’s comment on: What is your in­ter­net search method­ol­ogy ?

  It’s prob­a­bly only worth us­ing ecosia if you think your time is vir­tu­ally worth­less. Tree plant­ing is mostly a use­less ac­tivity in the grand scheme of things, though one gets into a some­what com­pli­cated calcu­la­tion to show this.

  Google claims there are 3*10^12 trees in the world, ecosia claims to have planted about 10^8. Trees are in an equil­ibrium with a bunch of stuff like at­mo­spheric CO2 as well. Your per­sonal con­tri­bu­tion to the num­ber of trees via ecosia is likely ~ a few trees (not clear whether they reach ma­tu­rity though).

  Do any­thing else use­ful with your time be­fore this.

• Roko 22 May 2020 19:58 UTC

  4 points

  on: The EMH Aten’t Dead

  There is a sim­ple ex­pla­na­tion to this: It turns out that be­ing a good ra­tio­nal­ist with ac­cess to an op­tion trad­ing ac­count is ac­tu­ally an edge.

  I also re­al­ized that the mar­ket was wrong and told friends to pull out of equities in mid-Fe­bru­ary. They didn’t listen.

• Roko 22 May 2020 16:39 UTC

  2 points

  in reply to: johnswentworth’s comment on: Pro­ject Pro­posal: Gears of Aging

  which an­i­mals?

• Roko 17 May 2020 8:56 UTC

  3 points

  in reply to: Davidmanheim’s comment on: Covid-19 Points of Lev­er­age, Travel Bans and Eradication

  alarmist. Good Judge­ment’s Dash­board has less than a 20% chance of over 350k deaths -

  It’s changed sub­stan­tially since you wrote this BTW. It’s now 54%

• Roko 24 Apr 2020 13:15 UTC

  2 points

  in reply to: Jalex Stark’s comment on: Far-Ul­travi­o­let Light in Public Spaces to Fight Pan­demic is a Good Idea but Premature

  yes!

  https://​​www.biorxiv.org/​​

• Roko 21 Apr 2020 17:13 UTC

  2 points

  in reply to: gilch’s comment on: Far-Ul­travi­o­let Light in Public Spaces to Fight Pan­demic is a Good Idea but Premature

  There are definitely some ma­te­ri­als that will re­sist it. There would be a one-off cost to re­place ma­te­ri­als, then mostly not much change.

• Roko 19 Apr 2020 8:04 UTC

  2 points

  on: Alarm bell for the next pan­demic, V.2

  It would be very valuable to take a par­tic­u­lar “Alarm” and see how many true pos­i­tives, false pos­i­tives, true nega­tives and false nega­tives it would have pro­duced over the past 20 years.