DanArmak

• DanArmak 25 Mar 2020 16:49 UTC

  7 points

  in reply to: steve2152’s comment on: Could we use cur­rent AI meth­ods to un­der­stand dolphins?

  I think the dis­par­ity in num­ber of words is pro­por­tion­ally so large that this method won’t work. The (small) hy­po­thet­i­cal set of dolphin words wouldn’t match to a small sub­set of English words, be­cause what’s be­ing matched is re­ally the (em­bed­ded) struc­ture of the re­la­tion­ship be­tween the words, and any suffi­ciently small sub­set of English words loses most of its in­ter­est­ing struc­ture be­cause its ‘real’ struc­ture re­lates it to many words out­side that sub­set.

  Sup­port that dolphins (hy­po­thet­i­cally! coun­ter­fac­tu­ally! not re­al­is­ti­cally!) use only 10 words to talk about fish, but hu­mans use 100 words to do the same. I ex­pect you can’t match the re­la­tion­ship struc­ture of the 10 dolphin words to the much more com­plex struc­ture of the 100 hu­man words. But no sub­set of ~10 English words out of the 100 is a mean­ingful sub­set that hu­mans could use to talk about fish.

• DanArmak 22 Mar 2020 15:12 UTC

  13 points

  on: Could we use cur­rent AI meth­ods to un­der­stand dolphins?

  The ap­proach of the linked ar­ti­cle tries to match words mean­ing the same thing across lan­guages by sep­a­rately build­ing a vec­tor em­bed­ding of each lan­guage cor­pus and then look­ing for struc­tural (neigh­bor­hood) similar­ity be­tween the em­bed­dings, with an ex­tra global ‘ro­ta­tion’ step map­ping the two vec­tor spaces on one an­other.

  So if both lan­guages have a word for “cat”, and many other words re­lated to cats, and the re­la­tion­ship be­tween these words is the same in both lan­guages (e.g. ‘cat’ is close to ‘dog’ in a differ­ent way than it is close to ‘food’), then these words can be suc­cess­fully trans­lated.

  But if one lan­guage has a tiny vo­cab­u­lary com­pared to the other one, and the vo­cab­u­lary isn’t even a sub­set of the other lan­guage’s (dolphins don’t talk about cats), then you can’t get far. Un­less you have an English train­ing dataset that only uses words that do have trans­la­tions in Dolphin. But we don’t know what dolphins talk about, so we can’t build this dataset.

  Also, this is ma­chine learn­ing on text with dis­tinct words; do we even have a ‘sep­a­rate words’ parser for dolphin sig­nals?

• DanArmak 17 Mar 2020 15:22 UTC

  2 points

  in reply to: Petter Wingren’s comment on: Good News: the Con­tain­ment Mea­sures are Working

  I think you are ab­solutely right.

  I’m not quite that sure I’m right. (I was gen­uinely ask­ing about the mechanism, not claiming there isn’t one!) I am not an ex­pert and there are other epi­demics that die out with­out hav­ing in­fected most of the pop­u­la­tion, like in­deed sea­sonal flu and cold, and I don’t know all the causes of that that might ap­ply here.

  End re­sult will be less ca­su­alties but longer pan­demic.

  It could be worse; ‘less over­all ca­su­alties’ re­lies on the rea­son­able but un­proven as­sump­tions of:

  • Reli­able nat­u­ral im­mu­niza­tion, i.e. peo­ple won’t (of­ten) catch it twice

  • Few or no mu­ta­tions that act as a ‘sec­ond wave’ or in the ex­treme case like the sea­sonal flu that hap­pens ev­ery year

  • (Most) peo­ple with light/​no symp­toms don’t end up with long term com­pli­ca­tions, or a per­sis­tent virus that can re­ac­ti­vate later

• DanArmak 17 Mar 2020 10:32 UTC

  7 points

  on: Good News: the Con­tain­ment Mea­sures are Working

  More­over, the worst of the pan­demic, or at least of this first wave, will be over in 2-3 months, as long as the con­tain­ment mea­sures are in place,

  What’s the mechanism be­hind the slow­ing growth? When we let up on our mea­sures (quaran­tine etc) why will the growth not speed up again, un­til most of the pop­u­la­tion have been ex­posed?

[Question] And was it for ra­tio­nal, ob­ject-level rea­sons?

• DanArmak 7 Mar 2020 17:47 UTC

  5 points

  on: Com­ing Back to Biking

  If you don’t mind the ex­pense, you might want to con­sider an elec­tric bike; I found them to be just as fun to ride as reg­u­lar ones. You can set the e-as­sist level low enough to ex­ert se­ri­ous effort when you want to, or just turn off the en­g­ine en­tirely, and con­versely set it high if you’re tired or your knees hurt.

• DanArmak 7 Mar 2020 17:43 UTC

  2 points

  in reply to: renato’s comment on: Com­ing Back to Biking

  I found a drop­per post to be a great help with that. It’s much eas­ier to figure out the right height while rid­ing and not hav­ing to dis­mount to ad­just it. And anec­do­tally, it some­times feels bet­ter to ad­just it up or down by 5-10 mil­lime­ters, maybe due to differ­ent clothes or shoes or pos­ture or sur­face grade.

  Note: even the cheap­est drop­per posts costs around 100 euro (from a cur­sory Google search). Peo­ple who aim for cheap bikes of­ten don’t con­sider them. If you can af­ford it, con­sider if it would be a small in­vest­ment into your com­fort and longer-term health.

• DanArmak 27 Jan 2020 17:46 UTC

  2 points

  in reply to: maximkazhenkov’s comment on: Mo­ral pub­lic goods

  I don’t see how that is ap­pli­ca­ble.

  In the first case, to avoid the penalty of be­ing fined, you pay taxes.

  In the sec­ond case, to avoid the penalty of be­ing taxed, you don’t donate.

  If I al­low you to donate with­out be­ing taxed, it doesn’t fol­low that you will donate. Maybe you don’t want to donate to be­gin with, or not un­less ev­ery­one else does as well. That’s the model the OP as­sumes.

  Tax rates on non-dona­tion gifts (= marginal in­come taxes of the non-rich) are “only” a few tens of per­cents. For the OP’s model to work, he had to as­sume a ra­tio of 1:1,000,000 be­tween the value to a no­ble of keep­ing or donat­ing money. That’s as if there was a 99.999,999,9% tax rate on dona­tions! If there was such a tax rate, then mak­ing dona­tions tax-free would cer­tainly stim­u­late a lot of dona­tions. But as it is, un­der the OP’s gen­eral as­sump­tions, tax rates of ~~ 30% should not much mat­ter.

• DanArmak 27 Jan 2020 11:48 UTC

  2 points

  in reply to: maximkazhenkov’s comment on: Mo­ral pub­lic goods

  I don’t un­der­stand your point. Pay­ing taxes or not is not re­lated to whether and how much other peo­ple also make char­i­ta­ble de­duc­tions. Be­zos donat­ing less or more doesn’t in­fluence Gates donat­ing less or more. What is the co­or­di­na­tion mechanism?

• DanArmak 27 Jan 2020 10:45 UTC

  2 points

  in reply to: paulfchristiano’s comment on: Mo­ral pub­lic goods

  Thanks, that’s in­for­ma­tive.

  One thing I would like to figure out is whether this can be ex­plained by busi­nesses re­struc­tur­ing so that some of the rich peo­ple who used to be own­ers get­ting div­i­dends are now com­pany ex­ec­u­tives get­ting salaries—but the salaries are still set mostly by them­selves to benefit them­selves, out of pro­por­tion to the value of their work to the com­pany. Direc­tors or board mem­bers of­ten also get salaries, again for very lit­tle work in most cases.

  Th­ese are things that might be col­lo­quially called ‘cap­i­tal’. Jeff Be­zos has a to­tal com­pen­sa­tion of 1.6 mil­lion; that is in­deed a tiny part of his net worth, but I still think of it as “Jeff Be­zos is a cap­i­tal­ist who is mak­ing money from the suc­cess­ful busi­ness he owns”, not as “Jeff Be­zos is be­ing paid for his tal­ents as a CEO”. I don’t care about the dis­tinc­tion from the in­come he gets from Ama­zon div­i­dends, shares, or his salary as a CEO. But then I’m not an economist; per­haps these are re­ally sig­nifi­cant differ­ences that I should care about.

• DanArmak 27 Jan 2020 10:02 UTC

  2 points

  in reply to: maximkazhenkov’s comment on: Mo­ral pub­lic goods

  That doesn’t seem to be enough to ex­plain the rich not vot­ing in the past to in­crease the marginal tax (like a few of them are now call­ing to do). Many differ­ent tax bills have been pro­posed over his­tory; this doesn’t seem like an idea no­body thought of un­til now.

  It’s likely that the rich (or peo­ple in gen­eral) don’t trust the gov­ern­ment with their money, don’t be­lieve it would be spent nearly as effec­tively or benefi­cially as pure re­dis­tri­bu­tion, and may en­tirely op­pose some of the gov­ern­ment’s uses of tax money and not want to fund them.

  In that case, what we need is a bill that pro­poses a spe­cial tax on the rich whose pro­ceeds go only and di­rectly to­wards re­dis­tri­bu­tion, or some sort of uni­ver­sal in­come which not suffi­cient to live on but also doesn’t count as ‘in­come’ for or­di­nary tax­a­tion and dis­qual­ifi­ca­tions for so­cial ser­vices for the poor. And such a thing is plau­si­bly hard to think of, draft, and get enough sup­port be­hind.

  Also, I would claim char­ity tax de­duc­tion already is such a co­or­di­na­tion mechanism

  It’s not a co­or­di­na­tion mechanism; it doesn’t al­low peo­ple to com­mit to giv­ing money if and only if ev­ery­one else also gives money, as a tax does. Even if giv­ing money was free (un­taxed), the OP’s co­or­di­na­tion prob­lem would re­main.

• DanArmak 26 Jan 2020 21:23 UTC

  5 points

  in reply to: maximkazhenkov’s comment on: Mo­ral pub­lic goods

  That’s a good point. Then Wei_Dai’s ques­tion be­comes more im­por­tant: why don’t we see other co­or­di­na­tion mechanisms in this space, be­sides forcible tax­a­tion? And why don’t rich peo­ple dis­pro­por­tionately vote in fa­vor of more pro­gres­sive taxes on them­selves?

  After all, if all we can learn from this post is that rich peo­ple don’t in fact have the posited prefer­ence, so this model doesn’t ap­ply, then it’s not very in­ter­est­ing.

• DanArmak 26 Jan 2020 18:53 UTC

  3 points

  in reply to: paulfchristiano’s comment on: Mo­ral pub­lic goods

  I think this prob­a­bly isn’t right—e.g. cap­i­tal in­come is a minor­ity for the top 1% of earn­ers in the US to­day, and the situ­a­tion is even starker for global in­equal­ity.

  That’s sur­pris­ing to me. Where does most of the in­come of rich peo­ple come from, then? Can you point me to some rele­vant re­source?

• DanArmak 26 Jan 2020 18:50 UTC

  4 points

  in reply to: maximkazhenkov’s comment on: He­donic asymmetries

  Why would you sup­pose that?

  Ex­actly for the rea­son you give your­self—we now change our be­hav­ior and our en­vi­ron­ment on much shorter timescales than evolu­tion op­er­ates on, due in large part to mod­ern tech­nol­ogy. We have a goal of cir­cum­vent­ing evolu­tion (see: this post) and we mod­ify our goals to suit our­selves. Evolu­tion is no longer fast enough to be the main de­ter­mi­nant of pre­vailing be­hav­ior.

  In the an­ces­tral en­vi­ron­ment of pre-agri­cul­tural so­cieties these be­hav­iors you de­scribe line up with max­i­miz­ing in­clu­sive ge­netic fit­ness pretty well

  We don’t know al­most any­thing about most rele­vant hu­man be­hav­ior from be­fore the in­ven­tion of writ­ing. Did they e.g. con­sume a lot of art (mu­sic, sto­ry­tel­ling, the­ater, danc­ing)? How much did such con­sump­tion cor­re­late with sta­tus or other fit­ness benefits, e.g. by con­spicu­ous con­sump­tion or ad­ver­tis­ing wealth? We re­ally don’t know.

• DanArmak 26 Jan 2020 17:40 UTC

  10 points

  on: He­donic asymmetries

  Pes­simistic er­rors are no big deal. The agent will ran­domly avoid be­hav­iors that get pe­nal­ized, but as long as those be­hav­iors are rea­son­ably rare (and aren’t the only way to get a good out­come) then that’s not too costly.

  Also, if an out­come re­ally is very bad, evolu­tion has no rea­son to limit the amount of suffer­ing ex­pe­rienced.

  Get­ting burned is bad for you. Evolu­tion makes it painful so you know to avoid it. But if strong and ex­treme pain re­sult in the same amount of avoidance, evolu­tion has no rea­son to choose “strong” over “ex­treme”. In fact, it might pre­fer “ex­treme” to get a more ro­bust out­come.

  And that’s how we get the abil­ity to in­flict the kind of tor­ture which is ‘worse than death’, and to imag­ine a Hell (and simu­la­tions thereof) with in­finite nega­tive util­ity, even though evolu­tion doesn’t have a con­cept of a fate be­ing “worse than death”—cer­tainly not worse than the death of your ex­tended fam­ily.

• DanArmak 26 Jan 2020 17:17 UTC

  3 points

  on: He­donic asymmetries

  But op­ti­mistic er­rors are catas­trophic. The agent will sys­tem­at­i­cally seek out the be­hav­iors that re­ceive the high re­ward, and will use loop­holes to avoid penalties when some­thing ac­tu­ally bad hap­pens. So even if these er­rors are ex­tremely rare ini­tially, they can to­tally mess up my agent.

  In other words, make sure your agent can’t wire­head! And from evolu­tion’s point of view, not spend­ing your time pro­mot­ing your in­clu­sive ge­netic fit­ness is wire­head­ing.

  We didn’t have the tech­nol­ogy to liter­ally wire­head un­til maybe very re­cently, so we went about it the long way round. But we still spend a lot of time and re­sources on e.g. con­sum­ing art, even ab­sent sig­nal­ling benefits, like watch­ing TV alone at home. And evolu­tion doesn’t seem to be likely to “fix” that given some more time.

  We don’t be­have in a “Malthu­sian” way, in­vest­ing all ex­tra re­sources in in­creas­ing the num­ber or rel­a­tive pro­por­tion of our de­scen­dants in the next gen­er­a­tion. Even though we definitely could, since pop­u­la­tion grows ge­o­met­ri­cally. It’s hard to have more than 10 chil­dren, but if ev­ery de­scen­dant of yours has 10 chil­dren as well, you can spend even the world’s biggest for­tune. And yet such clan­nish be­hav­ior is not a com­mon theme of any his­tory I’ve read; peo­ple pre­fer to get (al­most un­bound­edly) richer in­stead, and spend those riches on lux­u­ries, not chil­dren.

• DanArmak 26 Jan 2020 15:02 UTC

  7 points

  in reply to: cousin_it’s comment on: Have epistemic con­di­tions always been this bad?

  The im­me­di­ate ex­am­ple that comes to mind is when Richard Stal­l­man was can­celed in Oc­to­ber, some peo­ple feared he was or was in dan­ger of be­ing home­less. I re­mem­ber read­ing a post about this on Eric Ray­mond’s per­sonal blog, which he has since ap­par­ently deleted or hid­den. Part of the info was in posts on Stal­l­man’s own blog stal­l­man.org, which seems to be down, also refer­enced in this red­dit thread.

  I would like to think that RMS didn’t end up home­less, or not for more than a few days, since there must be many peo­ple who would give him dona­tions if it came to that (and if he would ac­cept them). But there has been no (very) pub­lic an­nounce­ment of him be­ing alright, for un­der­stand­able rea­sons. The list of peo­ple and or­ga­ni­za­tions who de­nounced him was im­pres­sively long, re­gard­less. (I mean the ones like GNU and MIT, not the pro­fes­sional de­nounc­ers who de­cided to at­tack him.)

• DanArmak 26 Jan 2020 12:06 UTC

  7 points

  on: Mo­ral pub­lic goods

  The fram­ing of “no­bles” and “peas­ants” dis­tracts me from your ques­tion; it im­plies con­no­ta­tions that you might want to clar­ify or en­dorse, or change your ter­minol­ogy.

  Real-life no­bles don’t pro­duce 10,000x value; they ex­tract value from peas­ants, by force of arms and law and cus­tom. It makes no sense to re­dis­tribute wealth by tax­ing ev­ery­one’s in­come if the no­bles get their in­come by tax­ing the peas­ants; just stop the no­bles from ex­tract­ing so much value.

  Some of the mod­ern su­per-rich do gen­er­ate dis­pro­por­tionately high value, e.g. from high-risk bets they made to build in­no­va­tive com­pa­nies. But most of their in­come still comes from cap­i­tal and own­ing the tools of pro­duc­tion and all that (cita­tion re­quired). And this in­fluences the moral calcu­lus for a lot of peo­ple. The rea­son for tak­ing some of their prop­erty (in­come) is not just that most peo­ple want to do it or that some­one else would en­joy it much more, it’s that it shouldn’t be theirs to be­gin with.

  A ter­minol­ogy of “no­bles” and “peas­ants” im­plies to me the idea that most all of the no­bles’ (the mod­ern rich) in­come is ex­tracted from from the peas­ants (ev­ery­one else), en­abled by the same state that then taxes them. Did you in­tend or en­dorse this view? If not, or if you think it’s ir­rele­vant to the thought ex­per­i­ment, do you think the fram­ing of “no­bles” and “peas­ants” dis­tracts from the is­sue? It does for me.

• DanArmak 18 Jan 2020 18:51 UTC

  9 points

  on: Con­ver­sa­tional Cul­tures: Com­bat vs Nur­ture (V2)

  This post is well writ­ten and not over-long. If the con­cepts it de­scribes are un­fa­mil­iar to you, it is a well writ­ten in­tro­duc­tion. If you’re already fa­mil­iar with them, you can skim it quickly for a warm feel­ing of val­i­da­tion.

  I think the post would be even bet­ter with a short in­tro­duc­tion de­scribing its topic and scope, but I’m aware that other peo­ple have differ­ent prefer­ences. In par­tic­u­lar:

  • There are more than two ‘cul­tures’ or styles of dis­cus­sion, per­haps many more. The post calls this out to­wards the end (ap­par­ently this is new in v2).

  • The post gives two real ex­am­ples of Com­bat Cul­ture, and only one made-up sce­nario of Nur­ture Cul­ture. It does not at­tempt to ground the dis­cus­sion in any­thing quan­ti­ta­tive—how com­mon these cul­tures are, what they cor­re­late with, how to rec­og­nize or test for them, how grad­u­ally they may shade into each other or into some­thing else al­to­gether.

  I don’t want to frame these as short­com­ings; the post is still use­ful and in­ter­est­ing with­out them!

• DanArmak 18 Jan 2020 10:21 UTC

  5 points

  on: Is Click­bait De­stroy­ing Our Gen­eral In­tel­li­gence?

  This post raises some rea­son­able-sound­ing and im­por­tant-if-true hy­pothe­ses. There seems to be a vast open space of pos­si­ble pre­dic­tions, rele­vant ob­ser­va­tions, and al­ter­na­tive ex­pla­na­tions. A lot of it has good treat­ment, but not on LW, as far as I know.

  I would recom­mend this post as an in­tro­duc­tion to some ideas and a start­ing point, but not as a good ar­gu­ment or a ba­sis for any firm con­clu­sions. I hope to see more con­tent about this on LW in the fu­ture.