ryan_b

• ryan_b 11 Dec 2019 15:53 UTC

  4 points

  in reply to: Tenoke's comment on: Open & Wel­come Thread—De­cem­ber 2019

  The com­ment­ing guidelines al­lows users to set their own norms of com­mu­ni­ca­tion for their own pri­vate posts. This lets us ex­per­i­ment with differ­ent norms to see which work bet­ter, and also al­lows the LessWrong com­mu­nity to di­ver­sify into differ­ent sub­com­mu­ni­ties should there be in­ter­est. It says habryka’s guidelines be­cause that’s who posted this post; if you go back through the other open threads, you will see other peo­ple posted many of them, and differ­ent com­ment­ing guidelines here and there. I think the posts that speak to this the most are:

  [Meta] New mod­er­a­tion tools and mod­er­a­tion guidelines (by habryka)

  Meta-tations on Moder­a­tion: Towards Public Archipelago (by Rae­mon)

• ryan_b 9 Dec 2019 22:29 UTC

  4 points

  on: Open & Wel­come Thread—De­cem­ber 2019

  There’s a post some­where in the ra­tio­nal­sphere that I can’t re­lo­cate for the life of me. Can any­body help?

  The point was com­mu­ni­ca­tion. The ex­am­ple given was the differ­ence be­tween a lec­ture and a ser­mon. The dis­tinc­tion the au­thor made was some­thing like a pro­fes­sor talk­ing to stu­dents in class, each of whom then goes home and does home­work by them­selves, ver­sus a preacher who gives his ser­mon to the con­gre­ga­tion, with the ex­pec­ta­tion that they will break off into groups and dis­cuss the ser­mon among them­selves.

  I have a vague mem­ory that there were graph­ics in­volved.

  I have tried lo­cal search on LessWrong, site search of LessWrong, and brows­ing a few post his­to­ries that seemed like they might be the au­thor based on a vague sense of aes­thetic similar­ity. I was sure it was here, but now I fear it may have been el­se­where or it is hid­den in some other kind of post.

• ryan_b 8 Dec 2019 20:00 UTC

  6 points

  on: The Les­son To Unlearn

  I re­ally liked this es­say.

  And as hack­ing bad tests shrinks in im­por­tance, ed­u­ca­tion will evolve to stop train­ing us to do it.

  This, how­ever, is en­tirely ex­ces­sive op­ti­mism.

• ryan_b 7 Dec 2019 16:43 UTC

  2 points

  in reply to: leggi's comment on: III. Con­scious Pro­pri­o­cep­tion—Your Sense of Po­si­tion, Move­ment & Balance.

  I get all the nor­mal pain/​tem­per­a­ture/​pres­sure/​fric­tion feed­back just fine. It is only the prob­lem of know­ing where they are in space with­out look­ing at them.

• ryan_b 6 Dec 2019 16:54 UTC

  4 points

  in reply to: ryan_b's comment on: What are some non-purely-sam­pling ways to do deep RL?

  I don’t know what the pro­ce­dure for this is, but it oc­curs to me that if we can spec­ify in­for­ma­tion about an en­vi­ron­ment via differ­en­tial equa­tions in­side the neu­ral net­work, then we can also com­pare this net­work’s out­put to one that doesn’t have the same in­for­ma­tion.

  In the name of learn­ing more about how to in­ter­pret the mod­els, we could try some­thing like:

  1) Con­struct an ar­tifi­cial en­vi­ron­ment which we can com­pletely spec­ify via a set of differ­en­tial equa­tions.

  2) Run a neu­ral net­work to learn that en­vi­ron­ment with ev­ery com­bi­na­tion of those differ­en­tial equa­tions.

  3) Com­pare all of these to sev­eral con­trol cases of not pro­vid­ing any differ­en­tial equa­tions.

  It seems like how the con­trol case differs from each of the cases-with-struc­tural-in­for­ma­tion should give us some in­for­ma­tion about how the net­work learns the en­vi­ron­men­tal struc­ture.

• ryan_b 6 Dec 2019 16:09 UTC

  4 points

  on: Good Pos­ture & Body Align­ment. Your Base-Line for Self-Assess­ment & Im­prove­ment.

  I can vouch for sud­den and sig­nifi­cant gains in com­fort and func­tion­al­ity by fo­cus­ing on im­prov­ing your pos­ture. The method I used was less thor­ough than here—in­stead I just used an ex­er­cise band and a few stretch­ing ex­er­cises to im­prove the shoulder po­si­tion. This pro­vided im­proved com­fort im­me­di­ately, and sig­nifi­cant re­duc­tion in the frag­ility of my back in a mat­ter of days.

• ryan_b 6 Dec 2019 16:05 UTC

  5 points

  on: III. Con­scious Pro­pri­o­cep­tion—Your Sense of Po­si­tion, Move­ment & Balance.

  I just dis­cov­ered this se­quence, and I am pleased and im­pressed. The sub­ject of this post is some­thing I have been look­ing at learn­ing more about a lot re­cently, be­cause I have a prob­lem in the area.

  Speci­fi­cally, I never know where my feet are po­si­tioned. I can in­fer it, and I can con­firm it, but I sim­ply don’t feel the po­si­tion of my feet in re­la­tion to the rest of my body. Even when I am try­ing to fo­cus on it.

  By con­trast, I do feel where my calves are in space. Most of the time when I need to place my feet pre­cisely, I am ac­tu­ally just aiming my calves at that point and rely­ing on the fact that my feet are on the end of my calves.

• ryan_b 5 Dec 2019 17:32 UTC

  LW: 7 AF: 4

  AF

  on: What are some non-purely-sam­pling ways to do deep RL?

  This doesn’t strike di­rectly at the sam­pling ques­tion, but it is re­lated to sev­eral of your ideas about in­cor­po­rat­ing the differ­en­tiable func­tion: Neu­ral Or­di­nary Differ­en­tial Equa­tions.

  This is be­ing ex­ploited most heav­ily in the Ju­lia com­mu­nity. The broader pitch is that they have for­mal­ized the re­la­tion­ship be­tween differ­en­tial equa­tions and neu­ral net­works. This al­lows things like:

  • ap­ply­ing differ­en­tial equa­tion tricks to com­put­ing the out­puts of neu­ral networks

  • us­ing neu­ral net­works to solve pieces of differ­en­tial equations

  • us­ing differ­en­tial equa­tions to spec­ify the weight­ing of information

  The last one is the most in­trigu­ing to me, mostly be­cause it solves the prob­lem of ma­chine learn­ing mod­els hav­ing to start from scratch even in en­vi­ron­ments where in­for­ma­tion about the en­vi­ron­ment’s struc­ture is known. For ex­am­ple, you can provide it with Maxwell’s Equa­tions and then it “knows” elec­tro­mag­netism.

  There is a blog post about the pa­per and us­ing it with the Differ­en­tialEqua­tions.jl and Flux.jl libraries. There is also a good talk by Christo­pher Rack­auckas about the ap­proach.

  It is mostly about us­ing ML in the phys­i­cal sci­ences, which seems to be go­ing by the name Scien­tific ML now.

• ryan_b 5 Dec 2019 16:37 UTC

  LW: 19 AF: 10

  AF

  on: Seek­ing Power is Prov­ably In­stru­men­tally Con­ver­gent in MDPs

  Strong up­vote, this is amaz­ing to me. On the post:

  • Another ex­am­ple of ex­plain­ing the in­tu­itions for for­mal re­sults less for­mally. I strongly sup­port this as a norm.

  • I found the graph­ics helpful, both in style and con­tent.

  Some thoughts on the re­sults:

  • This strikes at the heart of AI risk, and to my in­ex­pert eyes the lack of any­thing rigor­ous to build on or crit­i­cize as a mechanism for the flashiest con­cerns has been a big fac­tor in how difficult it was and is to get en­gage­ment from the rest of the AI field. Even if the for­mal­ism fails due to a crit­i­cal flaw, the abil­ity to spot such a flaw is a big step for­ward.

  • The for­mal­ism of av­er­age at­tain­able util­ity, and the ex­plicit dis­tinc­tion from num­ber of pos­si­bil­ities, pro­vides pow­er­ful in­tu­ition even out­side the field. This in­cludes ar­eas like war­fare and busi­ness. I re­al­ize it isn’t the goal, but I have always con­sid­ered ap­pli­ca­bil­ity out­side the field as an im­por­tant test be­cause it would be deeply con­cern­ing for think­ing about goal-di­rected be­hav­ior to mys­te­ri­ously fail when ap­plied to the only ex­tant things which pur­sue goals.

  • I find the re­sult aes­thet­i­cally pleas­ing. This is not im­por­tant, but I thought I would men­tion it.

• ryan_b 4 Dec 2019 21:02 UTC

  2 points

  on: Sym­biotic Wars

  I feel like this was ren­dered its own ex­plicit meme in the form of The Game.

• ryan_b 4 Dec 2019 15:02 UTC

  11 points

  on: Is the rate of sci­en­tific progress slow­ing down? (by Tyler Cowen and Ben South­wood)

  They ask whether TFP and re­lated mea­sures un­der­value the tech sec­tor. They con­clude no:

  • Coun­tries with smaller tech sec­tors than the US see a similar pro­duc­tivity slow­down.

  • Even if un­der­val­ued, the tech sec­tor is not big enough to ex­plain the whole slow­down in the US.

  • The slow­down be­gins in 1973, pre­dat­ing the tech sec­tor.

• ryan_b 4 Dec 2019 14:52 UTC

  11 points

  in reply to: jmh's comment on: Is the rate of sci­en­tific progress slow­ing down? (by Tyler Cowen and Ben South­wood)

  They agree, and even raise ap­prox­i­mately the same point:

  To con­sider a sim­ple ex­am­ple, imag­ine that an Amer­i­can com­pany is in­effi­cient, and then a man­age­ment con­sul­tant comes along and teaches that com­pany bet­ter per­son­nel man­age­ment prac­tices, thereby boost­ing pro­duc­tivity. Does this count as an im­prove­ment in TFP or not? Or is it sim­ply an in­crease in la­bor sup­ply, namely that of the con­sul­tant? On one hand, some hith­erto-ne­glected idea is in­tro­duced into the pro­duc­tion pro­cess. That might mil­i­tate in fa­vor of count­ing it as TFP. On the other hand, the in­tro­duced idea is not a new one, and ar­guably the busi­ness firm in ques­tion is sim­ply en­gaged in “catch up” eco­nomic growth, rel­a­tive to more tech­nolog­i­cally so­phis­ti­cated firms.

  I am con­fused by their dis­tinc­tion be­tween “catch up” growth and reg­u­lar growth; it seems to me it should not mat­ter how long it takes for an idea to diffuse when count­ing its value. Con­sider if each idea were like a cor­po­ra­tion: it’s not like any­one dis­misses the growth that hap­pened af­ter the iPod came out as “catch up” value, and only gains dur­ing Jobs’ origi­nal tenure count as “real” value.

  It does seem clear to me that the timing prob­lem makes it very difficult to dis­en­tan­gle from other ideas at this high level.

• ryan_b 3 Dec 2019 16:28 UTC

  11 points

  in reply to: ryan_b's comment on: Is the rate of sci­en­tific progress slow­ing down? (by Tyler Cowen and Ben South­wood)

  Prob­lems with the TFP:

  Timing:

  For in­stance, all at­tempts to mea­sure sci­en­tific progress through pro­duc­tivity come up against a timing prob­lem: the in­no­va­tion does not hap­pen at the same time as it is adopted.

  Rele­vant foot­note:

  Comin, Diego, Bart Ho­bijn, and Em­i­lie Rovito. Five facts you need to know about tech­nol­ogy diffu­sion. No. w11928. Na­tional Bureau of Eco­nomic Re­search, 2006.

  Un­der­val­ues en­hance­ments to la­bor, cap­i­tal, or land:

  many sci­en­tific ad­vances work through en­abling a greater sup­ply of la­bor, cap­i­tal, and land, and those ad­vances will be un­der­val­ued by a TFP met­ric. Let’s say some­one in­vents a use­ful painkil­ler, and that makes it eas­ier for many peo­ple to show up to work and be pro­duc­tive. Out­put will rise, yet that ad­vance will show up as an in­crease in la­bor sup­ply, rather than as an in­crease in tech­nol­ogy or sci­en­tific knowl­edge.

  Some ideas are counted as cap­i­tal:

  The more gen­eral prob­lem is that many sci­en­tific and tech­nolog­i­cal ad­vances are em­bod­ied in con­crete cap­i­tal goods.

  Some ideas are counted as la­bor:

  If a worker gen­er­ates and car­ries for­ward a new sci­en­tific idea for pro­duc­ing more with a given amount of la­bor, that mea­sures the same way as the worker be­ing taught greater con­scien­tious­ness and pro­duc­ing more.

  It is not clear how con­sis­tent this is:

  In defense of TFP mea­sures, these prob­lems are not always so se­ri­ous if these bi­ases are roughly con­stant over time. In that case, changes in TFP still would re­flect changes in the rate of progress of sci­ence and tech­nol­ogy. The ab­solute level of TFP could be bi­ased by cap­i­tal-em­bod­ied and la­bor-em­bod­ied tech­ni­cal change, but over time, for com­par­i­sons, the ex­pected sign of that bias might be close to zero. Still, it is not ob­vi­ous that the rate em­bod­i­ment of new ideas into cap­i­tal and la­bor, in per­centage terms, should be con­stant over time.

• ryan_b 3 Dec 2019 16:20 UTC

  11 points

  in reply to: ryan_b's comment on: Is the rate of sci­en­tific progress slow­ing down? (by Tyler Cowen and Ben South­wood)

  They look at To­tal Fac­tor Pro­duc­tivity in­stead:

  To­tal fac­tor pro­duc­tivity is an at­tempt to mea­sure over­all eco­nomic effec­tive­ness: how much a so­ciety can do with the in­puts it has. TFP, multi-fac­tor pro­duc­tivity, or the Solow resi­d­ual are all differ­ent names for this same con­cept. It refers to the amount of out­put growth left un­ex­plained af­ter ac­count­ing for all in­puts, i.e. it is a resi­d­ual and not some­thing we mea­sure di­rectly. So if out­put grew by ten, and the con­tri­bu­tions of land, la­bor, and cap­i­tal each were judged at 3 (to­tal of 9), TFP would be mea­sured at one. As such, it is vuln­er­a­ble to mea­sure­ment er­rors in any of the main se­ries that go into its calcu­la­tion. Nonethe­less the hope is that this vari­able mea­sures the “left over” con­tri­bu­tion of ideas to the pro­cess of pro­duc­tion, and thereby helps us mea­sure the effi­cacy of sci­ence. It won’t con­fuse progress in sci­ence with a coun­try hav­ing a large stock of oil.

  This seems to match the his­tor­i­cal record bet­ter:

  One ad­van­tage of TFP is that it seems to cor­re­spond to com­mon in­tu­itions as to when sci­en­tific progress was es­pe­cially high. Robert H. Gor­don, in his book The Rise and De­cline of Amer­i­can Eco­nomic Growth, has ar­gued at great length that sci­en­tific and tech­nolog­i­cal progress reached a peak in the early part of the twen­tieth cen­tury. That was a time when fos­sil fuels, elec­trifi­ca­tion, in­dus­tri­al­iza­tion, ni­tro­gen fer­til­izer, cars, ra­dio, tele­phones, clean wa­ter, vac­cines, and an­tibiotics all took on ma­jor roles in hu­man lives in the wealthier coun­tries. Within a mat­ter of decades,, hu­man life was trans­formed, in large part be­cause of the ex­ten­sion and ap­pli­ca­tion of the ear­lier In­dus­trial Revolu­tion. Con­sis­tent with this pic­ture, Amer­i­can TFP typ­i­cally grew quickly in the 1920s and 1930s, rang­ing from be­tween two to slightly over three per­cent per year. In more re­cent times, in con­trast, TFP growth of­ten has ranged be­tween one and one and a half per­cent.

• ryan_b 3 Dec 2019 16:14 UTC

  11 points

  on: Is the rate of sci­en­tific progress slow­ing down? (by Tyler Cowen and Ben South­wood)

  Disen­tan­gling sci­ence and GDP:

  For in­stance, Nor­we­gian GDP per cap­ita is typ­i­cally 20-30% above that of Swe­den and Den­mark, but Nor­way is ac­cess­ing the same gen­eral stock of sci­en­tific knowl­edge. Similarly, US agri­cul­tural pro­duc­tivity per labourer and labour hour out­performed Europe dur­ing the 19th and 20th cen­turies, but some of that gap prob­a­bly sprung from a high ra­tio of land to in­hab­itant, rather than an in­her­ent tech­nolog­i­cal ad­van­tage, and for some of that time Amer­i­can sci­ence may have been be­hind that of Europe.

  The rele­vant foot­note:

  On agri­cul­tural pro­duc­tivity com­par­i­sons, see Broad­berry, Stephen, and Mary O’Ma­hony. “Bri­tain’s Twen­tieth-Cen­tury Pro­duc­tivity Perfor­mance in In­ter­na­tional Per­spec­tive.” Work and pay in the twen­tieth cen­tury (2007): 301-329, and Broad­berry, Stephen N., and Dou­glas A. Ir­win. “La­bor pro­duc­tivity in the United States and the United King­dom dur­ing the nine­teenth cen­tury.” Ex­plo­ra­tions in Eco­nomic His­tory 43, no. 2 (2006): 257-279.

• ryan_b 2 Dec 2019 15:46 UTC

  2 points

  on: CO2 Strip­per Post­mortem Thoughts

  The plan­ning fal­lacy for garage pro­jects is an in­ter­est­ing prob­lem, be­cause it doesn’t lend it­self im­me­di­ately to a refer­ence class un­less you have done a lot of pro­jects be­fore.

  Still, next time you want to tackle a garage in­ven­tion that you can just pre­dict it will take as long as this one. It will be in­ter­est­ing to see how the differ­ence be­tween pro­jects com­pares to the plan­ning fal­lacy im­pact on a sin­gle pro­ject; in very big pro­jects, the size of the pro­ject com­pletely dom­i­nates any con­sid­er­a­tions of field or tech­nol­ogy.

  Come to think of it, does any­one know if there is a maker com­mu­nity some­where that records its bud­gets and timelines? Maybe that could be used as a refer­ence class for garage pro­jects.

• ryan_b 30 Nov 2019 19:59 UTC

  2 points

  on: A vot­ing the­ory primer for rationalists

  I nom­i­nate this post be­cause it does a good job of suc­cinctly de­scribing differ­ent ways to ap­proach one of the core prob­lems of civ­i­liza­tion, which is to say choos­ing policy (or choos­ing policy choosers). A lot of our ac­tivity here is about avoid­ing catas­tro­phe some­how; we have spent com­par­a­tively lit­tle time on big-pic­ture things that are in­cre­men­tal im­prove­ments.

  Anec­do­tally, this post did a good job of jog­ging loose a funk I was in re­gard­ing the pro­cess of poli­tics. Poli­tics is a no­to­ri­ously *ugh field* kind of en­deavor in the broader cul­ture, and a par­tic­u­lar taboo in this com­mu­nity. And yet, the very act of se­ri­ously con­sid­er­ing what the op­tions are is a sooth­ing balm when you are oth­er­wise in a state of over­whelming dis­gust. It’s like the rit­ual of evitable dis­may.

• ryan_b 28 Nov 2019 21:46 UTC

  10 points

  on: Embed­ded Agents

  I nom­i­nate this post for two rea­sons.

  One, it is an ex­cel­lent ex­am­ple of pro­vid­ing sup­ple­men­tal writ­ing about ba­sic in­tu­itions and thought pro­cesses, which is ex­tremely helpful to me be­cause I do not have a good enough com­mand of the for­mal work to in­tuit them.

  Two, it is one of the few ex­am­ples of ex­per­i­ment­ing with differ­ent kinds of pre­sen­ta­tion. I feel like this is un­der­ap­pre­ci­ated and un­der-uti­lized; bet­ter ways of com­mu­ni­cat­ing seems like a strong baseline re­quire­ment of the ra­tio­nal­ity pro­ject, and this post pushes in that di­rec­tion.

• ryan_b 28 Nov 2019 2:40 UTC

  9 points

  on: The Costly Co­or­di­na­tion Mechanism of Com­mon Knowledge

  I have definitely linked this more than any other post. The key in­sight for me is that com­mon knowl­edge is some­thing which has spe­cific costs and trade-offs. Pre­vi­ously I im­plic­itly viewed com­mon knowl­edge as an ac­ci­dent or a fea­ture of the en­vi­ron­ment.

  What links here?

  • The Re­view Phase: Helping LessWrongers Eval­u­ate Old Posts by Ben Pace (9 Dec 2019 0:54 UTC; 43 points)

• ryan_b 26 Nov 2019 18:50 UTC

  11 points

  in reply to: Uriel Fiori's comment on: A The­ory of Per­va­sive Error

  Men­cius Mold­bug is the pen name of Cur­tis Yarvin.