PeterMcCluskey

• PeterMcCluskey 21 Jun 2019 15:21 UTC

  7 points

  on: Should ra­tio­nal­ity be a move­ment?

  Note that the ra­tio­nal­ity com­mu­nity used to have the Sin­gu­lar­ity Sum­mit, which was fairly similar to EAGlobal in its at­ti­tude to­ward high-sta­tus speak­ers.

• PeterMcCluskey 11 Jun 2019 15:04 UTC

  3 points

  in reply to: Dagon's comment on: Get Rich Real Slowly

  The risk/​re­turn curve hides some im­por­tant prob­lems.

  Many in­vestors have more in­come when the stock mar­ket is high than when it is low.

  For ex­am­ple, dona­tions to MIRI are likely heav­ily in­fluenced by the price of tech stocks and by cryp­tocur­rency prices. MIRI needs to pay salaries that fluc­tu­ate much less than those prices. If MIRI in­vests in as­sets that have much cor­re­la­tion with those prices, MIRI will likely end up buy­ing dis­pro­por­tionate amounts of those as­sets within a year af­ter they peak, and buy­ing less (some­times even sel­l­ing) within a year af­ter they reach a low.

  Maybe that works well for short-term mar­ket fluc­tu­a­tions, but when­ever there are sig­nifi­cant mar­ket fluc­tu­a­tions that last for sev­eral years, that would mean that MIRI would be buy­ing mainly at the worst time. I ex­pect that would cause them to un­der­perform the rele­vant bench­mark by at least a few per­centage points.

  So I ex­pect MIRI will be bet­ter off min­i­miz­ing the volatility of their in­vest­ments.

  Note that this con­clu­sion holds even if MIRI is risk-neu­tral.

  I ex­pect that a fair num­ber of in­di­vi­d­ual in­vestors have in­come fluc­tu­a­tions that im­ply the same con­clu­sion.

• PeterMcCluskey 10 Jun 2019 2:25 UTC

  3 points

  on: Ma­jor Up­date on Cost Disease

  They have a good ar­gu­ment that the Bau­mol effect ex­plains a sig­nifi­cant frac­tion of cost dis­ease.

  Much of their point is that it’s not sur­pris­ing that some in­dus­tries have bet­ter pro­duc­tivity growth than oth­ers.

  But they don’t have a full ex­pla­na­tion of the ex­tent to which pro­duc­tivity growth varies by in­dus­try. Why do med­i­cal pro­duc­tivity trends ap­pear much worse than auto re­pair pro­duc­tivity trends? What caused the phar­ma­ceu­ti­cal in­dus­try to be­come less pro­duc­tive?

  Their par­tial an­swer seems to be Han­so­nian:

  A key dis­tinc­tion is when la­bor is an in­put into the good or ser­vice and when la­bor, in essence, is the ser­vice. Con­sumers do not care how much la­bor is used as an in­put into the pro­duc­tion of a car, but they do care how much la­bor in­put is used in a mas­sage, artis­tic perfor­mance, or doc­tor’s visit.

  That’s point­ing in the right gen­eral di­rec­tion, but I ex­pect a full ex­pla­na­tion would also in­volve con­sumers be­ing un­will­ing or un­able to switch to com­pa­nies that offer cheaper ser­vices.

• PeterMcCluskey 10 May 2019 15:26 UTC

  12 points

  on: Tales From the Amer­i­can Med­i­cal System

  This doesn’t sound like some­thing that would hap­pen at Kaiser. It’s some­times un­clear whether peo­ple at Kaiser care about my health, but they seem to have clear in­cen­tives to be po­lite and to not waste my time with un­nec­es­sary vis­its.

• PeterMcCluskey 24 Apr 2019 19:45 UTC

  20 points

  on: [An­swer] Why wasn’t sci­ence in­vented in China?

  This is im­pres­sive given the amount of time you put into it.

  There’s more ev­i­dence in the book State, Econ­omy, and the Great Diver­gence: Great Bri­tain and China, 1680s − 1850s, by Peer Vries, which I re­viewed here. In par­tic­u­lar, Vries dis­putes the claim that prop­erty rights were se­cure in Bri­tain be­fore the in­dus­trial rev­olu­tion.

• PeterMcCluskey 9 Apr 2019 19:10 UTC

  11 points

  on: What are CAIS’ bold­est near/​medium-term pre­dic­tions?

  One clear differ­ence be­tween Drexler’s wor­ld­view and MIRI’s is that Drexler ex­pects progress to con­tinue along the path that re­cent ML re­search has out­lined, whereas MIRI sees more need for fun­da­men­tal in­sights.

  So I’ll guess that Drexler would pre­dict maybe a 15% chance that AI re­search will shift away from deep learn­ing and re­in­force­ment learn­ing within a decade, whereas MIRI might say some­thing more like 25%.

  I’ll guess that MIRI would also pre­dict a higher chance of an AI win­ter than Drexler would, at least for some defi­ni­tion of win­ter that fo­cused more on diminish­ing IQ-like re­turns to in­vest­ment, than on over­all spend­ing.

• PeterMcCluskey 6 Apr 2019 2:09 UTC

  24 points

  on: How do peo­ple be­come am­bi­tious?

  The liter­a­ture on learned hel­pless­ness de­scribes how to de­stroy am­bi­tion. That sug­gests that any good an­swer should re­sem­ble mov­ing away from those situ­a­tions.

• PeterMcCluskey 4 Apr 2019 16:51 UTC

  4 points

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

  See this more rigor­ous anal­y­sis by Robert A. Fre­itas. Yes, it will cause some prob­lems.

• PeterMcCluskey 1 Apr 2019 15:13 UTC

  7 points

  on: What are effec­tive strate­gies for miti­gat­ing the im­pact of acute sleep de­pri­va­tion on cog­ni­tion?

  I used to think I didn’t get any­thing from try­ing to nap, since I never fell asleep dur­ing one. Then some­one re­ported that I had been snor­ing dur­ing one of those naps in which I hadn’t no­ticed any sign that I’d fallen asleep. Now I’m un­cer­tain whether naps are worth­while.

• PeterMcCluskey 19 Mar 2019 22:07 UTC

  1 point

  on: Im­pli­ca­tions of liv­ing within a Simulation

  Robin Han­son has, of course, writ­ten about how the simu­la­tion hy­poth­e­sis should af­fect our be­hav­ior.

• PeterMcCluskey 9 Mar 2019 20:50 UTC

  3 points

  in reply to: ChristianKl's comment on: How dan­ger­ous is it to ride a bi­cy­cle with­out a helmet?

  In­ter­est­ing. The abil­ity to fold it enough to fit in a back­pack should re­duce the has­sle of stor­ing it at my des­ti­na­tion, which has been part of why I’ve been re­luc­tant to use one.

• PeterMcCluskey 9 Mar 2019 17:29 UTC

  8 points

  on: How dan­ger­ous is it to ride a bi­cy­cle with­out a helmet?

  I also ride my bike in Berkeley with­out a helmet.

  Some other con­sid­er­a­tions which in­fluence me:

  • helmet use seems to make crashes more likely (by mak­ing bik­ers and/​or drivers less cau­tious), so it’s mis­lead­ing to use data about harm that’s con­di­tioned on there be­ing a re­ported ac­ci­dent.

  • I’m fairly care­ful to avoid roads with heavy traf­fic, or with cars driv­ing more than about 30 mph. I ex­pect that fatal­ity rates vary a lot by these road fac­tors.

  • I use a cheap bike that doesn’t go as fast as the av­er­age bike.

  Alas, I don’t have good ev­i­dence about how to quan­tify these con­sid­er­a­tions.

• PeterMcCluskey 27 Feb 2019 16:36 UTC

  6 points

  on: Un­con­scious Economies

  Economists usu­ally treat minds as black boxes. That seems to help them de­velop their mod­els, maybe via helping them to ig­nore is­sues such as “I’d feel em­bar­rassed if my mind worked that way”.

  There doesn’t seem to be much in­cen­tive for text­books to im­prove their effec­tive­ness at per­suad­ing the marginal stu­dent. The in­cen­tives might even be back­wards, as be­com­ing a good economist al­most re­quires think­ing in ways that seem odd to the av­er­age stu­dent.

• PeterMcCluskey 17 Feb 2019 17:40 UTC

  2 points

  in reply to: Kaj_Sotala's comment on: Why didn’t Agoric Com­put­ing be­come pop­u­lar?

  Those con­cerns would have slowed adop­tion of agoric com­put­ing, but they seem to ap­ply to mar­kets in gen­eral, so they don’t seem use­ful in ex­plain­ing why agoric com­put­ing is less pop­u­lar than mar­kets in other goods/​ser­vices.

• PeterMcCluskey 17 Feb 2019 17:32 UTC

  1 point

  in reply to: korin43's comment on: Why didn’t Agoric Com­put­ing be­come pop­u­lar?

  The cen­tral plan­ner may know ex­actly what re­sources ex­ist on the sys­tem they own, but they don’t know all the al­gorithms and data that are available some­where on the in­ter­net. Agoric com­put­ing would en­able more op­tions for get­ting pro­gram­mers and database cre­ators to work for you.

• PeterMcCluskey 16 Feb 2019 17:50 UTC

  12 points

  on: Why didn’t Agoric Com­put­ing be­come pop­u­lar?

  One ob­sta­cle has been se­cu­rity. To de­velop any soft­ware that ex­changes ser­vices for money, you need to put sub­stan­tially more thought into the se­cu­rity risks of that soft­ware, and you prob­a­bly can’t trust a large frac­tion of the ex­ist­ing base of stan­dard soft­ware. Coau­thor Mark S. Miller has de­voted lots of effort to re­plac­ing ex­ist­ing op­er­at­ing sys­tems and pro­gram­ming lan­guages with se­cure al­ter­na­tives, with very limited suc­cess.

  One other ex­pla­na­tion that I’ve won­dered about in­volves con­flicts of in­ter­est. Mar­ket in­ter­ac­tions are valuable mainly when they gen­er­ate co­op­er­a­tion among agents who have di­ver­gent goals. Most soft­ware de­vel­op­ment hap­pens in en­vi­ron­ments where there’s enough co­op­er­a­tion that adding mar­ket forces wouldn’t provide much value via im­proved co­op­er­a­tion. I think that’s true even within large com­pa­nies. I’ll guess that the benefits of the agoric ap­proach only be­come in­ter­est­ing when large num­ber of com­pa­nies switch to us­ing it, and there’s lit­tle re­ward to be­ing the first such com­pany.

• PeterMcCluskey 14 Feb 2019 3:49 UTC

  7 points

  on: In­di­vi­d­ual profit-shar­ing?

  Univer­si­ties have tried some­thing like this for tu­ition. See these men­tions from Alex Tabar­rok. They have some trou­ble with peo­ple de­fault­ing.

• PeterMcCluskey 1 Feb 2019 23:09 UTC

  5 points

  in reply to: ESRogs's comment on: Drexler on AI Risk

  Thanks, I’ve fixed those.

Drexler on AI Risk

• PeterMcCluskey 28 Jan 2019 22:32 UTC

  3 points

  in reply to: PeterMcCluskey's comment on: Refram­ing Su­per­in­tel­li­gence: Com­pre­hen­sive AI Ser­vices as Gen­eral Intelligence

  After fur­ther reread­ing, I now think that what Drexler imag­ines is a bit more com­plex: (sec­tion 27.7) “se­nior hu­man de­ci­sion mak­ers” would have ac­cess to a ser­vice with some strate­gic plan­ning abil­ity (which would have enough power to gen­er­ate plans with dan­ger­ously broad goals), and they would likely re­strict ac­cess to those high-level ser­vices.

  I sus­pect Drexler is de­liber­ately vague about the ex­tent to which the strate­gic plan­ning ser­vices will con­tain safe­guards.

  This, of course, de­pends on the con­tro­ver­sial as­sump­tion that rel­a­tively re­spon­si­ble or­ga­ni­za­tions will de­velop CAIS well be­fore other en­tities are able to de­velop any form of equally pow­er­ful AI. I con­sider that plau­si­ble, but it seems to be one of the weak­est parts of his anal­y­sis.

  And pre­sum­ably the pub­li­cly available AI ser­vices won’t be suffi­ciently gen­eral and pow­er­ful to en­able ran­dom peo­ple to as­sem­ble them into an agent AGI? Com­bin­ing a robo­car + Google trans­late + an air­craft de­signer + a the­o­rem prover doesn’t sound dan­ger­ous. But I’d pre­fer to have some­thing more con­vinc­ing than just “I spent a few min­utes look­ing for risks, and didn’t find any”.