ryan_b

• ryan_b 24 Apr 2019 14:53 UTC

  2 points

  in reply to: rossry's comment on: On the Na­ture of Pro­gram­ming Languages

  The usual ex­am­ple here is mem­ory con­trol. The point of the higher-level lan­guages is to ab­stract away the de­tails of mem­ory and reg­isters, so there is no mal­loc/​free equiv­a­lent when writ­ing in them; for this pur­pose they use garbage col­lec­tion.

  Of course, even­tu­ally peo­ple found a need for ad­dress­ing these kinds of prob­lems, and so fea­tures to al­low for it were added later. C reigns supreme in em­bed­ded ap­pli­ca­tions be­cause of the pre­cise mem­ory and I/​O ca­pa­bil­ities, but there is stuff for em­bed­ded Haskell and em­bed­ded LISP now. But note that in these sources they are talk­ing about stuff like spe­cial com­pilers and strate­gies for keep­ing the au­to­matic garbage col­lec­tion from blow­ing ev­ery­thing up, whereas with C, you just mostly write reg­u­lar C. Also in­ter­rupts.

• ryan_b 23 Apr 2019 17:01 UTC

  13 points

  on: De­gree of du­pli­ca­tion and co­or­di­na­tion in pro­jects that ex­am­ine com­put­ing prices, AI progress, and re­lated top­ics?

  I pro­pose that the mo­ti­va­tion for all of these pro­jects is not to find the an­swer, but rather to build the in­tu­itions of the pro­ject mem­bers. If you were to com­pare the effects on in­tu­ition of read­ing re­search vs. perform­ing re­search, I strongly ex­pect that perform­ing re­search would be greater.

  Be­cause of this, I ex­pect that a sig­nifi­cant chunk of all the peo­ple who are work­ing in any ca­pac­ity on the AI risk prob­lem will take a di­rect shot at similar pro­jects them­selves, even if they don’t write it up. I would also be sur­prised to find an org with­out any such peo­ple in it.

• ryan_b 23 Apr 2019 15:07 UTC

  4 points

  on: 1960: The Year The Sin­gu­lar­ity Was Cancelled

  This seems highly plau­si­ble to me. It would be in­ter­est­ing to see some­thing that more closely tracks differ­ent kinds of in­no­va­tion—for ex­am­ple, how do hy­giene and vac­ci­na­tions (which stop dy­ing) com­pare to do­mes­ti­cated crops and ir­ri­ga­tion (which in­crease pro­duc­tivity di­rectly) in terms of pop­u­la­tion growth?

  Faster growth rates means more money means more AIs re­search­ing new tech­nol­ogy means even faster growth rates, and so on to in­finity.

  My prior is that this will not re­quire any­where near hu­man-level in­tel­li­gence. I firmly ex­pect this can be ac­com­plished with the kind of AI we already pos­sess, in tan­dem with cer­tain im­mi­nent kinds of au­toma­tion.

• ryan_b 23 Apr 2019 14:59 UTC

  6 points

  in reply to: moses's comment on: 1960: The Year The Sin­gu­lar­ity Was Cancelled

  If 1 in 1,000,000 peo­ple is an ir­re­press­ible ge­nius and pro­duces a tech­nolog­i­cal in­ven­tion, then we should see as many tech­nolog­i­cal in­ven­tions as there are mil­lions of peo­ple.

  Alter­na­tively, each per­son has some chance of mak­ing such an in­ven­tion, and the more per­sons there are the more chances for in­ven­tion(s) hap­pen­ing.

  If the ques­tion is ‘what is the model of in­no­va­tion that jus­tifies the as­sump­tion’ then I don’t know, but I would guess some var­i­ant of the Great Men the­ory of his­tory. We might model it as an IQ dis­tri­bu­tion.

• ryan_b 22 Apr 2019 18:24 UTC

  4 points

  in reply to: ryan_b's comment on: On the Na­ture of Pro­gram­ming Languages

  On the other-other hand, an ex­am­ple was star­ing me in the face that points more closely to your old in­tu­itions: I just started read­ing The Struc­ture and In­ter­pre­ta­tion of Clas­si­cal Me­chan­ics, which is the text­book used for clas­si­cal me­chan­ics at MIT. Of par­tic­u­lar note is that the book uses Scheme, a LISP di­alect, in or­der to en­force clar­ity and cor­rect­ness of un­der­stand­ing of me­chan­ics. The pro­gram­ming lan­guage is only cov­ered in the ap­pendix; they spend an hour or two on it in the course.

  The goal here is to raise the stan­dard of un­der­stand­ing the world to ‘can you ex­plain it to the com­puter.’

• ryan_b 22 Apr 2019 15:59 UTC

  9 points

  on: He­len Toner on China, CSET, and AI

  He­len’s com­ments on the as­sumed su­pe­ri­or­ity of China in gath­er­ing data about peo­ple brought to mind the re­cent dis­agree­ment be­tween Rich Sut­ton and Max Wel­ling.

  Sut­ton re­cently ar­gued that the bit­ter les­son of AI is meth­ods which lev­er­age com­pute the best are most effec­tive. Wel­ling re­sponds by ar­gu­ing the case that data is similarly im­por­tant, par­tic­u­larly for do­mains which are not well defined.

  This causes me to sus­pect that the US and China di­rec­tions for AI re­search will sig­nifi­cantly di­verge in the medium term.

• ryan_b 22 Apr 2019 15:19 UTC

  12 points

  on: On the Na­ture of Pro­gram­ming Languages

  My in­tu­ition is strongly op­po­site yours of ten years ago.

  • For ex­am­ple, there are Do­main Spe­cific Lan­guages, which are de­signed ex­actly for one prob­lem do­main.

  • C, the most wide­spread gen­eral-pur­pose pro­gram­ming lan­guage, does things that are ex­tremely difficult or im­pos­si­ble in highly ab­stract lan­guages like Haskell or LISP, which doesn’t seem to match the no­tion of all three be­ing a helpful way to think about the world.

  • Most of what we wind up do­ing with pro­gram­ming lan­guages is build­ing soft­ware tools. We pre­fer pro­grams to be writ­ten such that the think­ing is clear and cor­rect, but this seems to me mo­ti­vated more by con­ve­nience than any­thing else, and it rarely turns out that way be­sides.

  I would go as far as to say that the case of ‘our im­perfect brains deal­ing with a com­plex world’ is in fact a se­ries of spe­cific sub-prob­lems, and we build tools for solv­ing them on that ba­sis.

  On the other hand, it feels like there is a large in­fluence on pro­gram­ming lan­guages that isn’t well cap­tured by the tool-for-prob­lem or crutch-for-psy­chol­ogy di­chotomy: work­ing with other peo­ple. Con­sider the ob­ject-ori­ented lan­guages, like Java. For all that an ob­ject is a con­ve­nient way to rep­re­sent the world, and for all that it is meant to provide ab­strac­tions like in­her­i­tance, what ac­tu­ally seems to have driven the pop­u­lar­ity of ob­ject ori­en­ta­tion is that it pro­vides a way for the next pro­gram­mer not to know ex­actly what is hap­pen­ing in the code, but in­stead to take the cur­rent crop of ob­jects as given and then do what­ever ad­di­tional thing they need done.

  Should we con­sider a group of peo­ple sep­a­rated in time work­ing on the same prob­lem, an in­de­pen­dent prob­lem? Or should we con­sider that work­ing with peo­ple-in-the-fu­ture is some­thing we are psy­cholog­i­cally bad at, and we need a bet­ter way to or­ga­nize our think­ing about it? While the former seems more rea­son­able to me, I don’t ac­tu­ally know the an­swer here. One way to tell might be if the peo­ple who wrote Java said speci­fi­cally some­where that they wanted a lan­guage that would make it eas­ier for mul­ti­ple peo­ple to write large pro­grams to­gether over time. Another way might be if ev­ery­one who learned Java chose it be­cause they liked not hav­ing to worry that much about what the last guy did, so long as the ob­jects work.

• ryan_b 19 Apr 2019 14:33 UTC

  2 points

  on: Hu­man perfor­mance, psy­chom­e­try, and base­ball statistics

  This is very old, but if I am eye­bal­ling the timeline cor­rectly we should be ap­proach­ing the point where you are de­cid­ing whether to cut your losses or en­dorse the les­sons. So if I may, how did it go?

• ryan_b 17 Apr 2019 20:32 UTC

  2 points

  on: Could waste heat be­come an en­vi­ron­ment prob­lem in the fu­ture (cen­turies)?

  There’s a post about this at the blog DoTheMath, which calcu­lates we boil our­selves with waste heat in ~400 years, as­sum­ing GDP dou­bles ev­ery 100 years and per cap­ita en­ergy con­sump­tion in­creases at the same rate it has been for the pre­vi­ous ~400 years.

  The usual eco­nomic re­tort is that the econ­omy could look very differ­ent from the one we are used to, and de­cou­ple from en­ergy con­sump­tion. But the as­sump­tion about waste heat is what is do­ing the work here, and we have re­cently de­vel­oped ther­mal tran­sis­tors. Th­ese tran­sis­tors have been de­signed out of quan­tum ob­jects. And it turns out we might be able to beat the Planck limit in the far field. Which is to say, we can build heat com­put­ers, and then waste heat could be con­verted into com­pu­ta­tion.

  That doesn’t solve the prob­lem of too much en­ergy use be­ing bad, but if waste heat is com­pu­ta­tion then we can hit peak (safe) out­put, stay there, and still add value.

• ryan_b 17 Apr 2019 19:20 UTC

  2 points

  in reply to: habryka's comment on: StrongerByS­cience: a ra­tio­nal strength train­ing website

  Har! I thought that was just a ti­tling con­ven­tion we’d adopted. Oops!

StrongerByS­cience: a ra­tio­nal strength train­ing website

• ryan_b 17 Apr 2019 17:20 UTC

  8 points

  in reply to: Raemon's comment on: Open Prob­lems in Archipelago

  The bit about bundling in per­son and on­line com­mu­ni­ties caused me to think of the Liter­a­ture Re­view: Distributed Teams post.

  It feels to me like the same trust and com­mu­ni­ca­tion mechanisms from dis­tributed teams stand a good chance of ap­ply­ing to dis­tributed com­mu­ni­ties. I’m tempted to take the Liter­a­ture Re­view ar­ti­cle and go back through the Old LW post­mortem post to see how well the pre­dic­tions match up. From this post:

  Longterm, this kills the en­g­ine by which in­tel­lec­tual growth hap­pens. It’s what kil­led old LessWrong – all the in­ter­est­ing pro­jects were hap­pen­ing in pri­vate, (usu­ally in-per­son) spaces, and that meant that:

  new­com­ers couldn’t latch onto them and learn about them incidentally

  at least some im­por­tant con­cepts didn’t en­ter the in­tel­lec­tual com­mons, where they could ac­tu­ally be cri­tiqued or built upon

  From the Distributed Teams post:

  If you must have some team mem­bers not co-lo­cated, bet­ter to be en­tirely re­mote than leave them iso­lated. If most of the team is co-lo­cated, they will not do the things nec­es­sary to keep re­mote in­di­vi­d­u­als in the loop.

  I feel like mod­el­ling LessWrong as a Distributed Team with strange com­pen­sa­tion might be a use­ful lens.

• ryan_b 17 Apr 2019 14:37 UTC

  2 points

  on: Robin Han­son on Sim­ple, Ev­i­dence Backed Models

  I agree strongly with the in­for­ma­tion sym­me­try ar­gu­ment, al­though I am less con­fi­dent of the in­for­ma­tion vol­ume one—this is be­cause trans­ac­tion costs ap­ply to in­for­ma­tion too.

  I par­tic­u­larly like it be­cause in­for­ma­tion sym­me­try feels like an area which has got­ten a neg­ligible amount of at­ten­tion com­pared to ques­tions of sup­ply and de­mand, and there­fore is a good tar­get for mak­ing a lot of progress rel­a­tively quickly for rel­a­tively lit­tle in­vest­ment.

• ryan_b 17 Apr 2019 14:31 UTC

  4 points

  in reply to: Raemon's comment on: Robin Han­son on Sim­ple, Ev­i­dence Backed Models

  I no­tice a dis­tinct lack of ex­am­ples of suc­cess­ful large-scale dis­course. It feels like this is one of many ar­eas where we pul­led a sort of leg­erde­main where we en­vi­sion the thing we wish were true as the nat­u­ral state of the world or the state of the past, and then be­moan its sab­o­tage or de­cline.

  I cur­rently can­not think of a sin­gle in­stance of good pub­lic dis­course ever. This causes me to think it isn’t sad that clever ar­guers are de­stroy­ing good dis­course; it is in­ter­est­ing that they are an ob­sta­cle to good dis­course ever aris­ing in the first place.

  I’m happy to be wrong though. If we were to at­tack this prob­lem by max­i­miz­ing or du­pli­cat­ing the bright spots, what would they be?

• ryan_b 17 Apr 2019 14:15 UTC

  4 points

  in reply to: Theist's comment on: Open Thread April 2019

  Fol­low­ing on that:

  “Math­e­mat­i­cal Rea­son­ing Abil­ities of Neu­ral Models,” https://​​arxiv.org/​​pdf/​​1904.01557.pdf

  They have pro­posed a pro­ce­du­rally-gen­er­ated data set for test­ing whether a model is ca­pa­ble of the same types of math­e­mat­i­cal rea­son­ing as hu­mans.

• ryan_b 16 Apr 2019 21:12 UTC

  5 points

  in reply to: elizabeth's comment on: Liter­a­ture Re­view: Distributed Teams

  That is much bet­ter, but it raises a more spe­cific ques­tion: here you de­scribed the loop as a prop­erty of the task; but then you also wrote

  • Hire peo­ple like me

    • long OODA loop

  Which seems to mean you are the one with the long loop. I can eas­ily imag­ine differ­ent peo­ple hav­ing differ­ent max­i­mum loop-lengths, be­yond which they are likely to fail. Am I cor­rect in in­ter­pret­ing this to mean some­thing like try­ing to en­sure that the re­mote worker can han­dle the longest-loop task you have to give them?

• ryan_b 16 Apr 2019 17:28 UTC

  4 points

  in reply to: ChristianKl's comment on: High­lights from “In­te­gral Spiritu­al­ity”

  Kah­ne­man’s work does an un­usu­ally ex­cel­lent job of com­ing to ob­ject-level recom­men­da­tions. That seems to be what he is do­ing with his time now.

  I don’t think us­ing fast and slow would have hurt those ideas at all, what with it be­ing the ti­tle of the book. Fur­ther, I’ve seen plenty of cases of try­ing to wran­gle the di­chotomy by piling on ad­di­tional words like the elephant-or-rider con­ver­sa­tion.

  I also note we don’t talk about sys­tem 1 and sys­tem II much any­more. Look­ing at the Cu­rated list for the last three months, I see plenty of posts that are aiming squarely at sys­tem 1 or sys­tem II, ap­ply­ing one sys­tem to the other, or de­scribing one spe­cific tech­nique that could be called sys­tem I or II...but vir­tu­ally none of the posts say any­thing about ei­ther of them or men­tion Kah­ne­man. We’ve moved past the point where a bi­nary dis­tinc­tion is use­ful to our dis­cus­sions, and broad fa­mil­iar­ity with the un­der­ly­ing con­cepts is as­sumed with­out the need for ad­di­tional terms.

  This sug­gests some com­bi­na­tion of sys­tem I/​II be­ing easy to ap­ply cor­rectly, or the com­mu­nity be­ing un­usu­ally good at ap­ply­ing it, or both.

  It looks to me like how pow­er­ful a sys­tem is and how difficult it is to ap­ply cor­rectly are very differ­ent ques­tions, and it feels like they are rarely bal­anced well. I think this is prob­a­bly very difficult to do, and have seen peo­ple failing to ap­ply sys­tems cor­rectly way more of­ten than I have seen them suc­ceed, which gives me a very low prior for un­fa­mil­iar sys­tems’ util­ity.

• ryan_b 16 Apr 2019 15:45 UTC

  2 points

  on: Liter­a­ture Re­view: Distributed Teams

  Ex­cel­lent work! I par­tic­u­larly like in­clud­ing your notes in the com­ments.

  I have one ques­tion about OODA (I see long loops men­tioned in the post, but with­out at­tri­bu­tion; I don’t see them men­tioned in the notes ex­plic­itly). Could you talk more about the long-loop con­clu­sion, and how re­mote work benefits from it?

  My naive guess is that the band­width is­sues as­so­ci­ated with re­mote work cause feed­back to take longer, which means longer OODA loops are a de­sir­able trait in the worker, but my con­fi­dence is not par­tic­u­larly high.

• ryan_b 16 Apr 2019 15:06 UTC

  2 points

  in reply to: ChristianKl's comment on: High­lights from “In­te­gral Spiritu­al­ity”

  I’m not ques­tion­ing the qual­ifi­ca­tions of the source or the good­ness of the con­cepts, just the method cho­sen to com­mu­ni­cate them.

  If the point of the sys­tem is to in­tro­duce an in­fer­en­tial gap on pur­pose, with the goal of leav­ing un­in­tended as­so­ci­a­tions be­hind, I can see the rea­son­ing but dis­agree with it. I see a lot of at­tempts to do this, and a lot of dis­cus­sion us­ing such sys­tems, and vir­tu­ally noth­ing in the way of com­ing back down from the ab­strac­tions to ob­ject-level recom­men­da­tions again.

  This is likely the re­sult of ap­ply­ing the sys­tem badly, but the ease with which a tool is mis­ap­plied an im­por­tant fac­tor in the good­ness of the tool.

• ryan_b 16 Apr 2019 14:16 UTC

  2 points

  in reply to: Elo's comment on: High­lights from “In­te­gral Spiritu­al­ity”

  But if we as­sume other peo­ple also know the sys­tem, why would you have to ex­plain a whole load of con­cep­tual frame­work?

  How is this su­pe­rior to ad­dress­ing the ob­ject-level con­cerns di­rectly?

  Edit: I definitely mis­read that last sen­tence. Ig­nore me!