roystgnr

• roystgnr 9 Aug 2019 19:42 UTC

  7 points

  on: Can we re­ally pre­vent all warm­ing for less than 10B$ with the mostly side-effect free geo­eng­ineer­ing tech­nique of Marine Cloud Bright­en­ing?

  we still need to ad­dress ocean acidification

  And changes in pre­cip­i­ta­tion pat­terns (I’ve seen ev­i­dence that re­duc­ing so­lar in­ci­dence is go­ing to re­duce ocean evap­o­ra­tion, in­de­pen­dent of tem­per­a­ture).

  There’s also the “dou­ble catas­tro­phe” prob­lem to worry about. Even if the me­dian ex­pected out­come of a geo­eng­ineer­ing pro­cess is de­cent, the down­side var­i­ance be­comes much worse.

  I still sus­pect MCB is our least bad near- to medium-term op­tion, and even in the long term the pos­si­bil­ity of tar­geted geo­eng­ineer­ing to im­prove lo­cal cli­mates is awfully tempt­ing, but it’s not a panacea.

  As an un­re­lated aside, that CCC link rates “Methane Re­duc­tion Port­fo­lio” as “Poor”; I’d have marked it “Coun­ter­pro­duc­tive” for the mo­ment. The biggest long-term global warm­ing prob­lem is CO2 (thanks to the short half-life of methane), and the biggest ob­sta­cle to CO2 emis­sions re­duc­tion is vot­ers who think global warm­ing is over­sold. Let the prob­lem get big­ger un­til it can’t be ig­nored, and then pick the sin­gle-use-only low hang­ing fruit.

• roystgnr 17 Dec 2018 16:25 UTC

  4 points

  in reply to: moridinamael’s comment on: Preschool: Much Less Than You Wanted To Know

  Alex has not skipped a grade or put in an some se­cret fast-track pro­gram for kids who went to preschool, be­cause this does not ex­ist.

  Even more con­found­ing: my kids have been skip­ping kinder­garten in part be­cause they didn’t go to preschool. My wife works from home, and has spent a lot of time teach­ing them things and dou­ble-check­ing things they teach them­selves.

  Preschools don’t do track­ing any more than grade schools, so even if in the­ory they might provide bet­ter in­struc­tion than the av­er­age over­worked par­ent(s), the out­put will be 100% to­tally-ready-for-kinder­garten (who will be stuck as you de­scribe), which in the long term won’t look as good as a mix of 95% not-quite-as-ready-for-kinder­garten (who will catch up as you de­scribe) and 5% ready-for-first-grade (who will per­ma­nently be a year ahead).

• roystgnr 20 Aug 2018 20:13 UTC

  6 points

  in reply to: Sniffnoy’s comment on: Dis­solv­ing the Fermi Para­dox, and what re­flec­tion it provides

  Gah, of course you’re cor­rect. I can’t imag­ine how I got so con­fused but thank you for the cor­rec­tion.

• roystgnr 20 Jul 2018 18:52 UTC

  1 point

  in reply to: Sniffnoy’s comment on: Dis­solv­ing the Fermi Para­dox, and what re­flec­tion it provides

  You don’t need any cor­re­la­tion be­tween $X$ and $Y$ to have $E\left[XY\right]\ne E\left[X\right]E\left[Y\right]$ . Sup­pose both vari­ables are 1 with prob­a­bil­ity .5 and 2 with prob­a­bil­ity .5; then their mean is 1.5, but the mean of their prod­ucts is 2.25.

• roystgnr 5 Jul 2018 22:00 UTC

  3 points

  in reply to: Patrick Cruce’s comment on: Dis­solv­ing the Fermi Para­dox, and what re­flec­tion it provides

  Not quite. Ex­pected value is lin­ear but doesn’t com­mute with mul­ti­pli­ca­tion. Since the Drake equa­tion is pure mul­ti­pli­ca­tion then you could use point es­ti­mates of the means in log space and sum those to get the mean in log space of the re­sult, but even then you’d *only* have the mean of the re­sult, whereas what would re­ally be a “para­dox” is if $P(N=1|N\ne 0)$ turned out to be tiny.

• roystgnr 31 May 2018 21:58 UTC

  19 points

  on: Ex­pres­sive Vocabulary

  the thing you know perfectly well peo­ple mean when they say “chem­i­cals”

  I hon­estly don’t un­der­stand what that thing is, ac­tu­ally.

  To use an ex­am­ple from a Face­book post I saw this week:

  Is P-Men­thane-3,8-diol (PMD) a chem­i­cal? What about oil from the lemon eu­ca­lyp­tus tree? Oil of Le­mon Eu­ca­lyp­tus is typ­i­cally re­fined un­til it’s 70% PMD in­stead of 2%; does that turn it into a chem­i­cal? What if we were to re­fine it all the way to 100%? What if, now that we’ve got 100% PMD, we just start us­ing PMD syn­the­sized at a chem­i­cal plant in­stead?

  I do like the idea from an­other com­ment here that

  the motte is “tech­ni­cally, ev­ery­thing is a chem­i­cal,” and the bailey is “No need to worry about the con­tent of the food you buy.”

  But even that can be in­verted. Peo­ple buy “no ni­trites added” ba­con to try to avoid a dan­ger­ous chem­i­cal, and they end up get­ting the re­sult of (all nat­u­ral! or­ganic!) cel­ery-juice and cel­ery-pow­der pro­cesses. At best, the ni­trates in the cel­ery still end up get­ting con­verted to ni­trites dur­ing cur­ing, ex­cept that now there’s an­other loose vari­able in the pro­duc­tion so a bal­ance be­tween “car­diac/​can­cer risk from high ni­trite lev­els” ver­sus “bo­tulism risk from low ni­trite lev­els” is even harder to achieve. At worst, the con­sumer falsely be­lieves they’ve no fur­ther need to worry about the con­tent of the food they buy, and so they don’t mod­er­ate con­sump­tion of it they way they would have mod­er­ated con­sump­tion of “chem­i­cals”.

• roystgnr 29 Jan 2018 18:36 UTC

  6 points

  on: The differ­ent types (not sizes!) of infinity

  So, does 1+ω make sense? It does, for the or­di­nals and hy­per­re­als only.

  It make sense for car­di­nals (the size of “a set of some in­finite car­di­nal­ity” unioned with “a set of car­di­nal­ity 1″ is “a set with the same in­finite car­di­nal­ity as the first set”) and in real anal­y­sis (if lim f(x) = in­finity, then lim f(x)+1 = in­finity) too.

  What about −1+ω? That only makes sense for the hy­per­re­als.

  And for car­di­nals (the size of the set differ­ence be­tween “a set of some in­finite car­di­nal­ity” and “a sub­set of one el­e­ment” is the same in­finite car­di­nal­ity) and in real anal­y­sis (if lim f(x) = in­finity, then lim −1+f(x) = in­finity) too.

• roystgnr 6 Dec 2017 20:42 UTC

  6 points

  in reply to: yagudin’s comment on: Mas­ter­ing Chess and Shogi by Self-Play with a Gen­eral Re­in­force­ment Learn­ing Algorithm

  I be­lieve the an­swer to your sec­ond ques­tion is prob­a­bly tech­ni­cally “yes”; if there’s any way in which AZ mis­pre­dicts rel­a­tive to a hu­man, then there’s some Ensem­ble Learn­ing clas­sifier that weights AZ move choices with hu­man move choices and performs bet­ter than AZ alone. And be­cause Go has so many pos­si­ble states and moves at each state, hu­mans would have to be much, much, much worse at play over­all for us to con­clude that hu­mans were worse along ev­ery di­men­sion.

  How­ever, I’d bet the an­swer is prac­ti­cally “no”. If AlphaZero vs the top hu­mans is now an Elo differ­ence of 1200, that gives a pre­dicted hu­man vic­tory rate of about 1/​1000. We’d never be able to play enough games to get enough sys­tem­atic data to iden­tify a di­men­sion along which hu­mans chose bet­ter moves. And even if we did, it’s en­tirely pos­si­ble that the best re­sponse to that iden­ti­fi­ca­tion would be “give AlphaZero more train­ing time on those cases”, not “give AlphaZero a hu­man part­ner in those cases”. And even if we did de­cide to give AlphaZero a hu­man part­ner, how of­ten would the part­ner’s in­put end up over­rid­ing the move AlphaZero alone would have cho­sen? Would the hu­man even be able to use­lessly pay at­ten­tion for game af­ter game, just so they could be ready to con­tribute to a win­ning move on game 100?

• roystgnr 8 Nov 2017 16:04 UTC

  2 points

  in reply to: magfrump’s comment on: LDL 2: Non­con­vex Optimization

  Oh, well in that case the point isn’t sub­tlely lack­ing, it’s just eas­ily dis­proven. Given any func­tion from I^N to R, I can take the ten­sor product with cos(k pi x) and get a new func­tion from I^{N+1} to R which has k times as many non-globally-op­ti­mal lo­cal op­tima. Pick a de­cent k and iter­ate, and you can see the num­ber grow­ing ex­po­nen­tially with higher di­men­sion, not ap­proach­ing 0.

  Per­haps there’s some­thing spe­cial about the func­tions we try to op­ti­mize in deep learn­ing, a prop­erty that rules out such cases? That could be. But you’ve said noth­ing spe­cial about (or even defined) a par­tic­u­lar class of deep learn­ing prob­lems, rather you’ve made a claim about all higher di­men­sional op­ti­miza­tion prob­lems, a claim which has an in­finite num­ber of coun­terex­am­ples.

• roystgnr 7 Nov 2017 17:48 UTC

  12 points

  in reply to: Allen Rice’s comment on: Moloch’s Toolbox (1/​2)

  a Le­viathan could try to tran­scend/​strad­dle these fruits/​niches and force them up­ward into a more Pareto op­ti­mal con­di­tion, maybe even into the non-Nash E. states if we’re ex­tra lucky.

  Re­mem­ber that old Yud­kowsky post about Oc­cam’s Ra­zor, wherein he points out how “a witch did it” sounds su­per-sim­ple, but the word “witch” hides a ton of hid­den com­plex­ity? I’m pretty sure you’re do­ing the same thing here with the word “could”. In­stead of try­ing to pic­ture what an imag­i­nary all-pow­er­ful leader could do, imag­ine what a typ­i­cal leader would do. You’ve just given a speech to Hillary Clin­ton about why it’s im­per­a­tive to dec­i­mate the uni­ver­sity sys­tem for eco­nomic rea­sons, or to Don­ald Trump about how pro­tec­tion­ism is always Pareto-sub­op­ti­mal. Do they re­act the way you wish them to act? Does their re­ac­tion make you want to give them more force­ful power, or less?

  For that mat­ter, you don’t have to imag­ine any speeches—ev­ery poli­ti­cian and their staff have ac­cess to the same in­ter­net you do, and can learn ev­ery­thing about co­or­di­na­tion prob­lems and their solu­tions that you can, and even with­out to­tal­i­tar­ian offi­cial power they still have enough of a “bully pulpit” and enough cred­i­bil­ity, se­ri­ous­ness, and so on to push the world out of at least the shal­low­est non-globally-op­ti­mal lo­cal op­tima. How many of them are do­ing so? By con­trast, how many of them are de­liber­ately set­ting up more bar­ri­ers to such changes?

• roystgnr 7 Nov 2017 17:32 UTC

  6 points

  in reply to: ChristianKl’s comment on: Moloch’s Toolbox (1/​2)

  In the BFR an­nounc­ment Musk promized that tick­ets from in­ter­con­ti­nen­tal rocket travel for the price of a cur­rent econ­omy ticket.

  “Full-fare” econ­omy, which is much more ex­pen­sive than even the “typ­i­cal” in­ter­na­tional econ­omy seat tick­ets you’re think­ing of, but yes, and even out­siders don’t think it’s im­pos­si­ble. It is very sen­si­tive to a lot of as­sump­tions—third party spread­sheets I’ve seen say low thou­sands of dol­lars per ticket is pos­si­ble, but it wouldn’t take many as­sump­tions to fall short be­fore prices jumped into the low tens of thou­sands.

  If that’s the price to travel by rocket it would also be the price to go to a space sta­tion.

  No; a space sta­tion trip could even be an or­der of mag­ni­tude more ex­pen­sive. (Which I ad­mit sur­prised me too).

  An in­ter­con­ti­nen­tal bal­lis­tic arc can re­quire nearly as much fuel as Low Earth Or­bit, true, but some hops are shorter than oth­ers, and in ev­ery case the differ­ence then goes through an ex­po­nen­tial func­tion in the rocket equa­tion. BFR might end up need­ing half as much pro­pel­lant on a mere transat­lantic flight, and the whole “su­per cheap flights” plan is pred­i­cated on them re­duc­ing non-pro­pel­lant costs to a small mul­ti­ple of pro­pel­lant costs, so that’s not a triv­ial change.

  Also, an in­ter­con­ti­nen­tal flight can pack peo­ple in like sar­dines, with enough life sup­port to go up and down again. A flight to a space sta­tion im­plic­itly in­cludes the cost of lift­ing and op­er­at­ing (and if it’s go­ing to be big enough to be use­ful for a full BFR pas­sen­ger flight, as­sem­bling) a space sta­tion which is roomy enough and well equipped enough for ev­ery pas­sen­ger to stay a while.

  Maybe there are some ill­nesses where you can benefit from be­ing in zero g

  This is a main­stay of clas­sic Sci-Fi, but so far IIRC ev­ery effect we’ve found has been in the op­po­site di­rec­tion. Zero G is like anti-ex­er­cise for the hu­man body, like be­ing bedrid­den.

  and you can have your hos­pi­tal out­side of any ju­ris­dic­tion.

  This makes more sense, and (un­like the rest of my huge di­gres­sion) gets straight at the heart of the post here, but I don’t think or­bit adds any ex­tra sense to the idea. There are a hun­dred lit­tle ju­ris­dic­tions around the world who would hap­pily is­sue the med­i­cal equiv­a­lent of a “Flag of Con­ve­nience” and waive what­ever reg­u­la­tions you think are stupid in re­turn for get­ting heavy in­vest­ment and med­i­cal tourism. Even if there weren’t, and you had to go out­side ev­ery ju­ris­dic­tion, in­ter­na­tional wa­ters a hun­dred miles offshore would have the (ques­tion­able!) le­gal sta­tus you’re look­ing for and would still be cheaper and more con­ve­nient than in­ter­na­tional space a hun­dred miles up.

• roystgnr 23 Oct 2017 16:02 UTC

  22 points

  on: Four Scopes Of Advice

  Failing to fol­low good strate­gic ad­vice isn’t even the worst failure mode here; un­less you’re lucky you may not be given any strate­gic ad­vice at all in re­sponse to a tac­ti­cal ques­tion. If no­body no­tices that you’re com­mit­ting the XY Prob­lem, then you may be given good ad­vice for the tac­ti­cal prob­lem you asked about, fol­low it, and end up worse off than you were be­fore with re­spect to the strate­gic prob­lem you should have been ask­ing about in­stead.

• roystgnr 23 Oct 2017 13:46 UTC

  6 points

  on: LDL 2: Non­con­vex Optimization

  This ar­gu­ment doesn’t seem to take into ac­count se­lec­tion bias.

  We don’t get into a lo­cal op­ti­mum be­cuase we picked a ran­dom point and wow, it’s a lo­cal op­ti­mum, what are the odds!?

  We get into a lo­cal op­ti­mum be­cause we used an al­gorithm that speci­fi­cally *finds* lo­cal op­tima. If they’re still there in higher di­men­sions then we’re still li­able to fall into them rather than into the global op­ti­mum.

• roystgnr 16 Oct 2017 15:52 UTC

  2 points

  on: Why no to­tal win­ner?

  Is there some more gen­eral limit to power beget­ting power that would also af­fect AGI?

  The only one which im­me­di­ately comes to mind is in­flex­i­bil­ity. Often com­pa­nies shrink or fail en­tirely be­cause they’re sud­denly sub­ject to com­pe­ti­tion armed with a new idea. Why do the new ideas end up im­ple­mented by smaller com­peti­tors? The pos­i­tive feed­back of “larger com­pa­nies have more peo­ple who can think of new ideas” is dom­i­nated by the nega­tive feed­backs of “even the largest com­pany is tiny com­pared to its com­ple­ment” and “com­pa­nies de­velop mono­cul­tures where ev­ery­one thinks the same way” and “com­pa­nies tend to in­ter­nally sup­press new ideas which would de­value the com­pany’s ex­ist­ing as­sets”.

  Any AGI that doesn’t already have half the world’s brain­power would be sub­ject to the first limi­ta­tion (which may mean any AGI who hasn’t taken over half the world or just any AGI less than an hour old, de­pend­ing on how much “foom” the real world turns out to al­low, I ad­mit), and an AGI that prop­a­gates by self-copy­ing might even be more af­fected than hu­mans by the sec­ond limi­ta­tion. Whether an AGI was sub­ject to the third limi­ta­tion or not would de­pend on its as­pira­tions, I sup­pose; any­thing with a “con­quer the world” endgame would hardly let it­self get trapped in the “just stretch out cur­rent rev­enues as long as pos­si­ble” mind­set of Block­buster-vs-Net­flix...

• roystgnr 13 Oct 2017 20:06 UTC

  8 points

  in reply to: Stuart_Armstrong’s comment on: Toy model of the AI con­trol prob­lem: an­i­mated version

  You should provide some more ex­plicit li­cense, if you don’t want to risk headaches for oth­ers later. “yes [It’s gen­er­ally in­tended to be Open Source]” may be enough re­as­surance to copy the code once, but “yes, you can have it un­der the new BSD (or LGPL2.1+, or what­ever) li­cense” would be use­ful to have in writ­ing in the repos­i­tory in case oth­ers want to cre­ate de­rived works down the road.

  Thanks very much for cre­at­ing this!

• roystgnr 6 Apr 2017 21:28 UTC

  1 point

  in reply to: Riothamus’s comment on: OpenAI makes hu­man­ity less safe

  The sim­ple illus­tra­tion is ge­om­e­try; defend­ing a ter­ri­tory re­quires 360 de­grees * 90 de­grees of cov­er­age, whereas the at­tacker gets to choose their vec­tor.

  But at­tack­ing a ter­ri­tory re­quires long sup­ply lines, whereas defen­ders are on their home turf.

  But defend­ing a ter­ri­tory re­quires con­stant readi­ness, whereas at­tack­ers can make a sin­gle fo­cused effort on a sur­prise at­tack.

  But at­tack­ing a ter­ri­tory re­quires mo­bil­ity for ev­ery sin­gle weapons sys­tem, whereas defen­ders can plug their weapons straight into huge power plants or in­cor­po­rate moun­tains into their ar­mor.

  But defend­ing against vi­o­lence re­quires you to keep tar­gets in good re­pair, whereas at­tack­ers have en­tropy on their side.

  But at­tack­ers have to break a Schel­ling point, thereby risk­ing re­tri­bu­tion from oth­er­wise neu­tral third par­ties, whereas defen­ders are less likely to face a coal­i­tion.

  But defen­ders have to make enough of their mil­i­tary ca­pac­ity pub­lic for the pub­lic knowl­edge to serve as a de­ter­rent, whereas at­tack­ers can keep much of their ca­pa­bil­ities a se­cret un­til the at­tack be­gins.

  But at­tack­ers have to leave their tar­gets in an eco­nom­i­cally use­ful state and/​or in an im­me­di­ately-mil­i­tar­ily-crip­pled state for a first strike to be prof­itable, whereas defen­ders can cred­ibly pre­com­mit to purely de­struc­tive re­tal­i­a­tion.

  I could prob­a­bly go on for a long time in this vein.

  Over­all I’d still say you’re more likely to be right than wrong, but I have no con­fi­dence in the ac­cu­racy of that.

• roystgnr 11 Feb 2017 2:06 UTC

  2 points

  in reply to: ingive’s comment on: Open thread, Feb. 06 - Feb. 12, 2017

  $2 debt squared does make sense, though, it is $4 and no debt.

  No, it is $$4.

  If that’s what you meant to write, and it’s also ob­vi­ous to you that you could have writ­ten 40000¢¢ in­stead and still been com­pletely ac­cu­rate, then I’d love to know if you have any ideas of how this com­pu­ta­tion could map to any­thing in the real world. I would have thought that “kilo­gram me­ters squared per sec­ond cubed” was ut­ter non­sense if any­one had just tried to show me the ar­ith­metic with­out ex­plain­ing what it re­ally meant.

  If that’s not what you meant to write, or if it takes a sec­ond to figure out why $$4 isn’t 400¢¢ in­stead of 40000¢¢, then you’ve just got the illu­sion of sense go­ing on. And yes, I just no­ticed that pun and it wasn’t in­ten­tional.

• roystgnr 14 Dec 2016 14:50 UTC

  1 point

  in reply to: Lumifer’s comment on: Tra­di­tions and Ra­tion­al­ity.

  Face­book has pri­vacy set­tings, such that any­one who wants to limit their posts’ di­rect visi­bil­ity can.

  Whether you should take some­one else’s set­tings as ex­plicit con­sent should prob­a­bly vary from per­son to per­son, but I think the “if he didn’t want it to be widely seen he wouldn’t have set it up to be widely seen” heuris­tic is prob­a­bly ac­cu­rate when ap­plied to EY, even if it’s not ap­pli­ca­ble to ev­ery Joe Grandpa.

  Even in the Joe Grandpa case, it doesn’t seem like merely avoid­ing cit­ing and mov­ing on is a good solu­tion. If you truly fear that some­one is shar­ing more to the world than they in­tend to, the kind thing to do is in­form them and help them fix it, not ig­nore them and pray that ev­ery­one else who stum­bles upon it shares your sense of deco­rum.

• roystgnr 18 Nov 2016 18:03 UTC

  0 points

  in reply to: Luke_A_Somers’s comment on: Maine passes Ranked Choice Voting

  The form of the patholo­gies makes a differ­ence, no?

  IIRC the worst pathol­ogy with IRV in a 3 way race is ba­si­cally: you can now safely vote for a third party who can’t pos­si­bly win, but if they show real sup­port then it’s time to put one of the two Schel­ling party can­di­dates in front of them. So it’s not worse than plu­ral­ity but it’s not much of an im­prove­ment. Plus, with more can­di­dates IRV be­comes more and more po­ten­tially in­sane.

  With Schulze or Ranked Pairs the pathol­ogy is DH3: If a third party can win, you can of­ten help them win by vot­ing a “dark horse” can­di­date ahead of their two real com­peti­tors… but if enough peo­ple try that then the “dark horse” ac­tu­ally wins. That could ac­tu­ally end up much worse than plu­ral­ity.

  With Score the unique pathol­ogy is: If you know how to vote tac­ti­cally then your vote is al­most cer­tain to look like an ap­proval vote. So the only rea­son to use it rather than Ap­proval is to en­courage non-game-the­ory-geeks to sab­o­tage them­selves and wa­ter down their own votes.

  With Ap­proval (and Score, to a slightly lesser ex­tent) the pathol­ogy is the “Burr dilemma”. Okay, if Ap­proval vot­ers are good at tac­ti­cal vot­ing and there are many iter­a­tions of ac­cu­rate pub­lic polls, then they pick the Con­dorcet win­ner… but if ei­ther of those things aren’t true then peo­ple at­tempt­ing to vote tac­ti­cally can end up hurt­ing front-run­ner can­di­dates and elect­ing a “dark horse”.

  I’d say this all starts to make plu­ral­ity look rea­son­able by com­par­i­son, be­cause the two party sys­tem tends to only fail by elect­ing overly-con­ser­va­tive (in the Burkean sense) can­di­dates, which is a less dan­ger­ous failure than its op­po­site would be, but I ad­mit there is a re­cent very-salient coun­terex­am­ple.

• roystgnr 20 Jun 2016 19:05 UTC

  2 points

  in reply to: James_Miller’s comment on: Crazy Ideas Thread

  There’s a high-stakes vari­a­tional calcu­lus prob­lem. For what sea­sonal tem­per­a­ture pro­file do we get the best long-term min­i­mum for the sum of “deaths due to ex­treme cold” and “deaths due to trop­i­cal dis­eases whose vec­tor in­sects are stymied by ex­treme cold”.