Lanrian

• Lanrian 26 Feb 2019 20:59 UTC

  13 points

  on: New ver­sions of posts in “Map and Ter­ri­tory” and “How To Ac­tu­ally Change Your Mind” are up (also, new re­vi­sion sys­tem)

  Neat! Two ques­tions:

  Is it pos­si­ble to cre­ate a link that always lead to the newest ver­sion of an ar­ti­cle, or will all links lead to the ver­sion that ex­isted when the link was cre­ated?

  Will all ed­its be visi­ble in the his­tory, or can the au­thor make minor ed­its with­out trig­ger­ing a new ver­sion?

• Lanrian 17 Feb 2019 18:00 UTC

  1 point

  in reply to: avturchin's comment on: Quan­tify­ing an­thropic effects on the Fermi paradox

  Hm, I still can’t find a way to in­ter­pret this that doesn’t re­duce it to prior prob­a­bil­ity.

  Den­sity cor­re­sponds to how com­mon life is (?), which is pro­por­tional to $f_l$. Then the “size” of an area with a cer­tain den­sity cor­resonds to the prior prob­a­bil­ity of a cer­tain $f_l$? Thus, “the to­tal num­ber of peo­ple in low den­sity ar­eas is greater than the to­tal num­ber of peo­ple in high den­sity ar­eas, be­cause the size of the low den­sity area is so much greater” cor­re­sponds to “$p(f_l=low)\ast low>p(f_l=high)\ast high$, be­cause the prior prob­a­bil­ity (de­noted by p()) of $f_l=low$ is so much greater”.

• Lanrian 16 Feb 2019 13:05 UTC

  2 points

  in reply to: avturchin's comment on: Quan­tify­ing an­thropic effects on the Fermi paradox

  I ex­pected to find here a link on the Grace SIA Dooms­day ar­gu­ment. She uses the same logic as you, but then turns to the es­ti­ma­tion of the prob­a­bil­ity that Great filter is ahead. It looks like you ig­nore pos­si­ble high ex­tinc­tion rate im­plied by SIA (?). Also, Univer­sal DA by Vilenkin could be men­tioned.

  Yup, I talk about this in the sec­tion Prior prob­a­bil­ity dis­tri­bu­tion. SIA definitely pre­dicts dooms­day (or any­thing that pre­vents space colon­i­sa­tion), so this post only ap­plies to the frac­tion of pos­si­ble Earths where the prob­a­bil­ity of doom isn’t that high. De­spite be­ing small, that frac­tion is in­ter­est­ing to a to­tal con­se­quen­tial­ist, since it’s the one where we have the best chance at af­fect­ing a large part of the Uni­verse (as­sum­ing that our abil­ity to re­duce x-risk gets pro­por­tion­ally smaller as the prob­a­bil­ity of spread­ing to space goes be­low 0.01 % or so).

  Another ques­tion, which is in­ter­est­ing for me, is how all this af­fects the pos­si­bil­ity of SETI-at­tack—send­ing mal­i­cious mes­sages with the speed of light on the in­ter­galac­tic dis­tance.

  There was a bunch of dis­cus­sion in the com­ments of this post about whether SETI would even be nec­es­sary to find any ETI that wanted to be seen, given that the ETI would have a lot of re­sources available to look ob­vi­ous. At least Paul con­cluded that it was pretty un­likely that we could have missed any civil­i­sa­tion that wanted to be seen. I think that anal­y­sis still stands.

  In­clud­ing the pos­si­bil­ity of SETI-at­tacks in my anal­y­sis would mean that no early civili­a­tion could ever de­velop in an ad­vanced civil­i­sa­tion’s light cone, but the bor­ders be­tween ad­vanced civil­i­sa­tions would still be calcu­lated with the civil­i­sa­tions’ ac­tual ex­pan­sion speed (with the ad­di­tional com­pli­ca­tion that ad­vanced civil­i­sa­tions could ‘jump’ to any early civil­i­sa­tion that ap­pears in their light cone). If we as­sume that the time left un­til we be­come in­vuln­er­a­ble to SETI-at­tacks is neg­ligible (a dan­ger­ous as­sump­tion?), I think that’s roughly equiv­a­lent to the sce­nario un­der Visi­bil­ity of civil­i­sa­tions in Ap­pendix C, from Earth’s per­spec­tive.

  The third idea I had re­lated to this is the pos­si­bil­ity that “bad fine tun­ing” of the uni­verse will over­weight the ex­pected gain of the civil­i­sa­tion den­sity from SIA. For ex­am­ple, if a uni­verse will be perfectly fine-tuned, ev­ery star will have a planet with life; how­ever, it re­quires al­most un­be­liev­able fidelity of its pa­ram­e­ters tun­ing. The more prob­a­ble is the set of the uni­verses there fine tun­ing is not so good, and the hab­it­able planets are very rare.

  If I un­der­stand you cor­rectly, this is an ar­gu­ment that our prior prob­a­bil­ity of $f_l$ should be heav­ily weighted to­wards life be­ing very un­likely? That could change the con­clu­sion if the prior prob­a­bil­ity of $f_l$ was in­versely pro­por­tional to $f_l$, or even more ex­tremely tilted to­wards lower num­bers. I don’t see any par­tic­u­lar rea­son why we would be that con­fi­dent that life is un­likely, though, es­pe­cially since the rele­vant prob­a­bil­ity mass in my anal­y­sis already puts $f_l$ be­neath ${10}^{-10}$. Hav­ing a prior that puts ${10}^{10}$ times more prob­a­bil­ity mass on $f_l={10}^{-20}$ than $f_l={10}^{-10}$ is very ex­treme, given the un­cer­tainty about this area.

Quan­tify­ing an­thropic effects on the Fermi paradox

• Lanrian 10 Feb 2019 22:01 UTC

  3 points

  on: Prob­a­bil­ity space has 2 metrics

  $P(p1,p2)=|p1-p2|=\mid \frac{1}{e^{p_1}+1}-\frac{1}{e^{p_2}+1}|$

  I think this should have b in­stead of p: $P(p1,p2)=|p1-p2|=\mid \frac{1}{e^{b_1}+1}-\frac{1}{e^{b_2}+1}|$

• Lanrian 6 Feb 2019 22:20 UTC

  8 points

  on: Thoughts on Ben Garfinkel’s “How sure are we about this AI stuff?”

  What does OTTMH mean?

• Lanrian 1 Feb 2019 18:34 UTC

  1 point

  in reply to: Vaniver's comment on: Why is this util­i­tar­ian calcu­lus wrong? Or is it?

  Wait, are you claiming that hu­mans have moral in­tu­itions be­cause it max­i­mizes global util­ity? Surely moral in­tu­itions have been pro­duced by evolu­tion. Why would evolu­tion se­lect for agents with be­havi­our that max­i­mize global util­ity?

• Lanrian 25 Jan 2019 22:15 UTC

  6 points

  in reply to: Andrew Quinn's comment on: The 3 Books Tech­nique for Learn­ing a New Skilll

  3blue1brown has a se­ries on the essence of lin­ear alge­bra as well. It’s pretty great, and and could do well as the Why.

  I also like Lin­ear Alge­bra Done Right a lot, but it doesn’t fit neatly into this frame­work. It’s a bit too rigor­ous to be Why, not prac­ti­cal enough to be How, and it’s ap­proach differs enough from other books to make it difficult to look things up in.

• Lanrian 21 Jan 2019 8:48 UTC

  7 points

  in reply to: ChristianKl's comment on: Life can be bet­ter than you think

  The last per­son that I re­mem­ber writ­ing some­thing along the lines of You don’t have to ex­pe­rience nega­tive emo­tion, on LessWrong didn’t turn out well.

  I’m not sure what this means.

  Either you’re say­ing that the LW com­mu­nity dis­ap­proved of this per­son. In this case, see the front­page guidelines: try to pre­sent ar­gu­ments for your own view in­stead of stat­ing opinions of oth­ers.

  Or you’re say­ing that this per­son hacked around with they’re own emo­tions un­til they lost mean­ing. In this case, I am very cu­ri­ous about what they did!

  (Or you’re say­ing some­thing com­pletely differ­ent.)

• Lanrian 20 Jan 2019 8:58 UTC

  3 points

  in reply to: Benito's comment on: Why not tool AI?

  Eric Drexler’s re­port on com­pre­hen­sive AI ser­vices also con­tains rele­vant read­ings. Here is Ro­hin’s sum­mary of it.

• Lanrian 6 Jan 2019 0:49 UTC

  26 points

  on: Does anti-malaria char­ity de­stroy the lo­cal anti-malaria in­dus­try?

  While not com­pre­hen­sively cov­ered, GiveWell men­tions this in a few places. The sec­ond point here links to a re­port with this sec­tion dis­cussing whether peo­ple are will­ing to pay for nets, as well as a link to this old blog post which briefly makes the ar­gu­ment that peo­ple won’t buy their own nets, since pre­vi­ous hand-outs (from other char­i­ties) have re­sulted in a lack of lo­cal pro­duc­ers and an ex­pec­ta­tion of free nets. They also men­tion that nets have some pos­i­tive ex­ter­nal­ities, and mostly benefits chil­dren, who aren’t the ones pay­ing, which gives some rea­son to sub­si­dize them.

• Lanrian 21 Dec 2018 21:20 UTC

  3 points

  in reply to: TAG's comment on: You can be wrong about what you like, and you of­ten are

  Blind spots and bi­ases can be harm­ful to your goals with­out be­ing harm­ful to your re­pro­duc­tive fit­ness. Be­ing wrong about which fu­ture situ­a­tions will make you (per­ma­nently) hap­pier is an ex­cel­lent ex­am­ple of such a blind spot.

• Lanrian 19 Dec 2018 16:49 UTC

  6 points

  on: 2018 AI Align­ment Liter­a­ture Re­view and Char­ity Comparison

  Shah et al.’s Value Learn­ing Se­quence is a short se­quence of blog posts out­lin­ing the speci­fi­ca­tion prob­lem.

  The link goes to the Embed­ded Agency se­quence, not the value learn­ing se­quence (https://​www.less­wrong.com/​s/​4dHMdK5TLN6xcq­tyc)

• Lanrian 16 Dec 2018 16:29 UTC

  3 points

  in reply to: AlexMennen's comment on: An Ex­ten­sive Cat­e­gori­sa­tion of In­finite Paradoxes

  “In­deed Pas­cal’s Mug­ging type is­sues are already pre­sent with the more stan­dard in­fini­ties.”

  Right, in­finity of any kind (sur­real or oth­er­wise) doesn’t be­long in de­ci­sion the­ory.

  But Pas­cal’s Mug­ging type is­sues are pre­sent with large finite num­bers, as well. Do you bite the bul­let in the finite case, or do you think that un­bounded util­ity func­tions don’t be­long in de­ci­sion the­ory, ei­ther?

• Lanrian 14 Dec 2018 9:18 UTC

  1 point

  on: An Ex­ten­sive Cat­e­gori­sa­tion of In­finite Paradoxes

  Great post!

  Satan’s Ap­ple: Satan has cut a deli­cious ap­ple into in­finitely many pieces. Eve can take as many pieces as she likes, but if she takes in­finitely many pieces she will be kicked out of par­adise and this will out­weigh the ap­ple. For any finite num­ber i, it seems like she should take that ap­ple piece, but then she will end up tak­ing in­finitely many pieces.

  Pro­posed solu­tion for finite Eves (also a solu­tion to Trumped, for finite Trumps who can’t count to sur­real num­bers):

  After hav­ing eaten n pieces, Eve’s de­ci­sion isn’t be­tween eat­ing n pieces and eat­ing n+1 pieces, it’s be­tween eat­ing n pieces and what­ever will hap­pen if she eats the n+1st piece. If Eve knows that the fu­ture Eve will be fol­low­ing the strat­egy “always eat the next ap­ple piece”, then it’s a bad de­ci­sion to eat the n+1st piece (since it will lead to get­ting kicked out of par­adise).

  So what strat­egy should Eve fol­low? Con­sider the prob­lem of pro­gram­ming a strat­egy that an Eve-bot will fol­low. In this case, the best strat­egy is the strat­egy that will lead to the largest amount of finite pieces be­ing eaten. What this strat­egy is de­pends on the hard­ware, but if the hard­ware is finite, then there ex­ists such a strat­egy (per­haps count the num­ber of pieces and stop when you reach N, for the largest N you can store and com­pare with). Gen­er­al­is­ing to (finite) hu­mans, the best strat­egy is the strat­egy that re­sults in the largest amount of finite pieces eaten, among all strate­gies that a hu­man can pre­com­mit to.

  Of course, if we al­low in­finite hard­ware, then the prob­lem is back again. But that’s at least not a prob­lem that I’ll ever en­counter, since I’m run­ning on finite hard­ware.

• Lanrian 13 Dec 2018 9:00 UTC

  1 point

  on: The Bat and Ball Prob­lem Revisited

  How­ever, for the other two I ‘just see’ the cor­rect an­swer. Is this com­mon for other peo­ple, or do you have a differ­ent split?

  I think I figured out and ver­ified the an­swer to all 3 ques­tions in 5-10 sec­onds each, when I first heard them (though I was ex­posed to them in the con­text of “Take the cog­ni­tive re­flec­tion test which peo­ple fail be­cause the ob­vi­ous an­swer is wrong”, which always felt like cheat­ing to me).

  If I re­call cor­rectly, the third ques­tion was eas­ier than the sec­ond ques­tion, which was eas­ier than bat & ball: I think I gen­er­ated the cor­rect an­swer as a sug­ges­tion for 2 and 3 pretty much im­me­di­ately (alongside the sup­pos­edly ob­vi­ous an­swers), and I just had to check them. I can’t quite re­mem­ber my strat­egy for bat & ball, but I think I gen­er­ated the $0.1 ball, $1 bat an­swer, saw that the differ­ence was $0.9 in­stead of $1, ad­justed to $0.05, $1.05, and found that that one was cor­rect.

• Lanrian 13 Dec 2018 8:45 UTC

  8 points

  on: The Bat and Ball Prob­lem Revisited

  I sus­pect that this is less true the other two prob­lems—ra­tios and ex­po­nen­tial growth are top­ics that a math­e­mat­i­cal or sci­en­tific ed­u­ca­tion is more likely to build in­tu­ition for.

  This seems to be con­tra­dicted by:

  the bat and ball ques­tion is the most difficult on av­er­age – only 32% of all par­ti­ci­pants get it right, com­pared with 40% for the wid­gets and 48% for the lily­pads. It also has the biggest jump in suc­cess rate when com­par­ing uni­ver­sity stu­dents with non-stu­dents.

• Lanrian 7 Dec 2018 12:18 UTC

  4 points

  on: Worth keeping

  If there are bet­ter re­place­ments in gen­eral, then you will be in­clined to re­place things more read­ily.

  The so­cial ana­log is that in a com­mu­nity where friends are more re­place­able—for in­stance, be­cause ev­ery­one is ex­tremely well se­lected to be similar on im­por­tant axes—it should be harder to be close to any­one, or to feel safe and accepted

  I can come up with a coun­ter­vailing effect here, as well. Re­veal­ing prob­lems is a risk: you might get help and be in a more trust­ing friend­ship, or you might be dumped. If there are lots of good re­place­ments around, then get­ting dumped mat­ters less, since you can find some­one else. This pre­dicts that peo­ple in com­mu­ni­ties that gather similar peo­ple might ex­pose their prob­lems more of­ten, de­spite be­ing re­placed a higher frac­tion of the time.

  Another differ­ence be­tween cars and friends is that you’re go­ing to get equally good use out of your car re­gard­less of how you feel about it, but you’re friend­ship is go­ing to be differ­ent if you can cred­ibly sig­nal that you won’t re­place it (tak­ing the self­ish-ra­tio­nal-in­di­vi­d­ual model to the ex­treme, you prob­a­bly want to sig­nal that you’d re­place it if the friend started treat­ing you worse, but that you wouldn’t leave it just be­cause your friend re­vealed prob­lems). In a close com­mu­nity, that sig­nal might get worse if you re­peat­edly re­place friends, which pre­dicts that you’d be less likely to re­place friends in closer com­mu­ni­ties.

  No em­piri­cal ev­i­dence of any of this.

• Lanrian 2 Dec 2018 22:49 UTC

  16 points

  on: Dou­ble-Dip­ping in Dun­ning—Kruger

  Par­ti­ci­pants scor­ing in the bot­tom quar­tile on our hu­mor test (...) over­es­ti­mated their per­centile ranking

  A less well-known find­ing of Dun­ning—Kruger is that the best perform­ers will sys­tem­at­i­cally un­der­es­ti­mate how good they are, by about 15 per­centile points.

  Isn’t this ex­actly what you’d ex­pect if peo­ple were good bayesi­ans re­ceiv­ing scarce ev­i­dence? Every­one starts out with as­sum­ing that they’re in the mid­dle, and as they find some­thing easy or hard, they grad­u­ally up­date away from their prior. If they don’t have good in­for­ma­tion about how good other peo­ple are, they won’t up­date too much.

  If you then look at the ex­tremes, the very best and the very worst peo­ple, of course you’re go­ing to see that they should ex­trem­ify their be­liefs. But if ev­ery­one fol­lowed that ad­vice, you’d ruin the ac­cu­racy of the peo­ple more to­wards the mid­dle, since they haven’t re­ceived enough ev­i­dence to dis­t­in­guish them­selves from the ex­tremes.

  (Similarly, I’ve heard that peo­ple of­ten over­es­ti­mate their abil­ity on easy tasks and un­der­es­ti­mate their abil­ity on difficult tasks, which is ex­actly what you’d ex­pect if they had good epistemics but limited ev­i­dence. If task perfor­mance is a func­tion of task difficulty and tal­ent for a task, and the only things you can ob­serve is your perfor­mance, then be­liev­ing that you’re good at tasks you do well at and bad at tasks you fail at is the cor­rect thing to do. As a con­se­quence, say­ing that peo­ple over­es­ti­mate their driv­ing abil­ity doesn’t tell you that much about the qual­ity of their epistemics, in iso­la­tion, be­cause they might be fol­low­ing a strat­egy that op­ti­mises perfor­mance across all tasks.)

  The find­ing that peo­ple at the bot­tom over­es­ti­mate their po­si­tion with 46 per­centile points is some­what more ex­treme than this naïve model would sug­gest. As you say, how­ever, it’s eas­ily ex­plained when you take into ac­count that your abil­ity to judge your perfor­mance on a task is cor­re­lated with your perfor­mance on that task. Thus, the peo­ple at the bot­tom are just re­ceiv­ing noise, so on av­er­age they stick with their prior and judge that they’re about av­er­age.

  Of course, just be­cause some of the ev­i­dence is con­sis­tent with peo­ple hav­ing good epistemics doesn’t mean that they ac­tu­ally do have good epistemics. I haven’t read the origi­nal pa­per, but it seems like peo­ple at the bot­tom ac­tu­ally thinks that they’re a bit above av­er­age, which seems like a gen­uine failure, and I wouldn’t be sur­prised if there are more ex­am­ples of such failures which we can learn to cor­rect. The im­pos­tor syn­drome also seems like a case where peo­ple pre­dictably fail in fix­able ways (since they’d do bet­ter by es­ti­mat­ing that they’re of av­er­age abil­ity, in their group, rather than even try­ing to up­date on ev­i­dence).

  But I do think that peo­ple of­ten are too quick to draw con­clu­sions from look­ing at a spe­cific sub­set of peo­ple es­ti­mat­ing their perfor­mance on a spe­cific task, with­out tak­ing into ac­count how well their strat­egy would do if they were bet­ter or worse, or were do­ing a differ­ent task. This post fixes some of those prob­lems, by re­mind­ing us that ev­ery­one low­er­ing the es­ti­mate of their perfor­mance would hurt the peo­ple at the top, but I’m not sure if it cor­rectly takes into ac­count how the peo­ple in the mid­dle of the dis­tri­bu­tion would be af­fected.

  (The counter-ar­gu­ment might be that peo­ple who know about Dun­ning-Kruger is likely to be at the top of any dis­tri­bu­tion they find them­selves in, but this seems false to me. I’d ex­pect a lot of peo­ple to know about Dun­ning-Kruger (though I may be in a bub­ble) and there are lots of tasks where abil­ity doesn’t cor­re­late a lot with know­ing about Dun­ning-Kruger. Per­haps hu­mor is an ex­am­ple of this.)

• Lanrian 11 Jul 2018 21:51 UTC

  1 point

  in reply to: Gurkenglas's comment on: The Fermi Para­dox: What did Sand­berg, Drexler and Ord Really Dis­solve?

  Sure, there are lots of ways to break calcu­la­tions. That’s true for any the­ory that’s try­ing to calcu­late ex­pected value, though, so I can’t see how that’s par­tic­u­larly rele­vant for an­throp­ics, un­less we have rea­son to be­lieve that any of these situ­a­tions should war­rant some spe­cial ac­tion. Us­ing an­thropic de­ci­sion the­ory you’re not even up­dat­ing your prob­a­bil­ities based on num­ber of copies, so it re­ally is only calcu­lat­ing ex­pected value.