Simple cost-benefit: distributed computing considered harmful as scientific lemon projects run at high resource cost.
philosophy, sociology, charity
2011-09-152015-09-12 in progress certainty: possible

When con­sid­er­ing “help­ing out”, also con­sider the costs of your help­ing out and what alter­na­tives there are.

Charitable supercomputing

(FAH for short) boasts of being the largest/most pow­er­ful dis­trib­uted com­put­ing project in the world with >5 of capac­i­ty, focused on the NP-hard prob­lem of . It is pow­ered by vol­un­teers run­ning its client on their com­put­ers, or more specif­i­cal­ly, their and . (Their archi­tec­tures are more spe­cial­ized than nor­mal CPUs and, if a prob­lem hap­pens to match their archi­tec­ture, can run an order of mag­ni­tude or two faster & more pow­er-­ef­fi­ciently than a CPU would.)

The researchers solve their prob­lems, the vol­un­teers know their idle com­put­ing capac­ity is going to good use—ev­ery­one wins. Right?

But where is this free lunch com­ing from? Why are the researchers rely­ing on vol­un­teers? What does it cost the vol­un­teers to run the clients? Their time in set­ting this all up, of course, but more impor­tant­ly: elec­tric­i­ty. Lots and lots of elec­tric­i­ty.

Wikipedi­ans have already done the leg­work on how much elec­tric­ity con­sump­tion FAH caus­es. Each petaflop costs ~3 megawatts, so at >5 petaflops (6 as of Novem­ber 2011), it uses >15 megawatts. To put it another way, every hour FAH uses up 15 megawat­t-hours1.

Sins of commission

Elec­tric­ity does­n’t come from nowhere. If we are to do even the most sim­plis­tic cost-ben­e­fit analy­sis, we can’t sim­ply assume the cost is 15 megawatts con­jured out of nowhere or that the elec­tric­ity would have been con­sumed any­way. What a silly defense that would be—what are the power com­pa­nies doing, gen­er­at­ing a fixed amount of elec­tric­ity and if the FAHers shut down, the plants dump their elec­tric­ity into the air? Of course the mar­ginal demand from FAH causes the gen­er­a­tion of megawat­t-hours that would not oth­er­wise have been gen­er­at­ed. (Oh, but FAH users are so altru­is­tic that they engage in con­ser­va­tion to off­set their increased elec­tric­ity con­sump­tion! Yeah, right.) And it does­n’t mat­ter what other wastes of elec­tric­ity may be going on—those wastes stand con­demned as well (“two wrongs don’t make a right”). No more dis­trac­tions or excus­es: what are the costs of that cost?

The most obvi­ous cost is air pol­lu­tion. It is major enough that we don’t even need to con­sider any other costs, because air pol­lu­tion kills. The edi­tors of Next Big Future have listed a num­ber of inter­est­ing sta­tis­tics from the WHO and other stud­ies of elec­tric­ity gen­er­a­tion and the dead­li­ness of air pol­lu­tion (see also ).

For exam­ple, half the world’s elec­tric­ity pro­duc­tion is done by that filth­i­est and most fatal of sources, coal. (In com­par­ison, nuclear power is a stroll in the park, which means that French users can run FAH with a 75% cleaner con­science than the rest of us, while Ger­many’s reac­tion to Fukushima demon­strates how fatal knee-­jerk reac­tions can be, in the sad­dest and most utterly pre­dictable way pos­si­ble2.) We will—­gen­er­ous­ly—pre­tend that every­one con­tribut­ing to FAH is located in the US, with its rel­a­tively clean coal plants. Nev­er­the­less, each ter­awat­t-hour of coal kills 15 peo­ple. How often does FAH burn through a ter­awatt? Well, 1000 megawat­t-hours is one gigawat­t-hour, and 1000 gigawat­t-hours is one ter­awat­t-hour, and , so how fast does FAH burn through 1 mil­lion megawat­t-hours? Each year, FAH uses megawat­t-hours, or 131.5 gigawat­t-hours, or 13.15% of a ter­awat­t-hour. Each ter­awat­t-hour means, remem­ber, 15 dead peo­ple; in our grim cal­cu­lus, what is 13.15% of 15 dead peo­ple? 1.8 dead peo­ple.

So if the power is entirely derived from coal, FAH kills 2 peo­ple a year.

What if the power is from an oil plant? That’s worse! The listed deaths per ter­awat­t-hour for oil is 36 dead peo­ple, for an annual death toll of 4.4 peo­ple. Hey, it could be much worse: if FAH had been invented and pop­u­lar­ized in Chi­na, with a ter­awat­t-hour death toll of 278, that’d be 34 deaths every year.

With coal and oil out of the way, we can look at the minor­ity fuels which make up a small slice of the US power sup­ply. Bio­fuel is pretty bad with a death toll of 12/TWh; hydro isn’t too bad at 1.4/TWh; solar, wind, and nuclear power have <1/TWh death tolls. But of course, we don’t live in an envi­ron­men­tal fan­tasy where all our power is gen­er­ated by those cleaner sources, and even if we did, that would­n’t help the peo­ple in the past who have already been killed by FAH’s pol­lu­tion.

The actual power mix of the USA in 2009 was 45% coal, 24% nat­ural gas, 20% nuclear, and 7% hydro, so bal­anc­ing our num­bers that gives us 1.01 annual deaths for a USA power mix. Phew! Only one dead per­son. Does­n’t that make you feel bet­ter?

We already saw how much elec­tric­ity FAH con­sumes: 15 megawat­t-hours. But how much does each megawat­t-hour cost? The EIA says the aver­age US rate for 1 kilo­wat­t-hour in Novem­ber 2010 was $0.0962. A megawat­t-hour is 1000 kilo­wat­t-hours, so 1 megawat­t-hour is , or$96.2, so = $1443/hr. And its annual bill is , or ~12,650,000 dol­lars per year3.$12.65 mil­lion is a lot of mon­ey. Money is both fun­gi­ble and lim­it­ed; by spend­ing that money on FAH power bills, that was not spent on other things, although these points seem lost on a lot of peo­ple.4 What could one have done with that? Meta-char­ity esti­mates that <$1000 could save one life; another source says “Cost-­ef­fec­tive­ness esti­mates per death-averted are$64–294 for a range of coun­tries”5. (One mod­est pro­posal is to use this $1000 fig­ure as the base unit of a new coinage: the DC or ‘dead child’; it has the merit over the dol­lar of pos­si­bly ingrain­ing an under­stand­ing of s.) And these inter­ven­tions are the kind of things that can absorb a lot of mon­ey. (There are a lot of peo­ple out there who could use some help.) If <$1000 will buy 1 life, then $12.65m would buy ~12,650 lives. Quite a few, that. One won­ders whether FAH was the best form of char­ity the 300k or so vol­un­teers could have cho­sen to engage in. (FAH is lucky to be an aca­d­e­mic pro­ject; oth­er­wise we would see deeply ironic out­comes like shut­ting down & beg­ging for fund­ing—while its vol­un­teers burn 0.5 petaflops annu­al­ly. Per­haps if all those petaflop­s—spent on check­ing and recheck­ing for sig­nal­s—had ever done any­thing use­ful in those decades, they would not have to. It would be inter­est­ing to hear from Seti@home at what point they would con­sider it was time to throw in the towel and finally admit the obvi­ous: that there is not a scrap of astro­nom­i­cal evi­dence of any kind indi­cat­ing intel­li­gent alien life and so there just .) Maybe they would have done bet­ter to donate a few dol­lars to a reg­u­lar char­i­ty, and not run up their elec­tric­ity bill. One might won­der, though, about the case where one isn’t pay­ing for one’s elec­tric­i­ty. So, either you are pay­ing for all of your elec­tri­cal bill or you’re not: • If you are pay­ing for all of it, then yes, you can donate your elec­tri­cal costs! Just don’t run Fold­ing@Home and send Oxfam a Pay­pal dona­tion at the end of the year. • If you are not pay­ing for all of it, if some­one is shar­ing the bill or foot­ing the bill entire­ly, then donat­ing directly may harm your pock­et­book, yes. But in such a sit­u­a­tion, does it really still make sense to force the other to pay for all the elec­tric­ity you are using? The over­all eco­nom­ics are bad per the orig­i­nal note, it’s an inef­fi­cient way to turn some­one else’s money into char­i­ty. What right do you have to burn the elec­tric­ity like there’s no tomor­row, for that mat­ter? If you weren’t going to use a year’s worth of elec­tric­i­ty, then whomever is pay­ing for your elec­tric­ity is poorer by that$10 as surely as if you had pick­-pock­eted him of $10; few would agree that you are Robin Hood who may steal from the rich and give to the char­i­ta­ble. He accepted that risk when he gave you access to his elec­tric­ity and deserves how­ever you can con­trive to screw him over? What an atti­tude! Other points can be dealt with sim­i­lar­ly: • per­haps one wor­ries that over­head on a small dona­tion of$10 will elim­i­nate the val­ue. But the over­head on finan­cial trans­ac­tions is usu­ally only a few per­cent, and the dif­fer­ence between FAH and the best char­i­ties is a dif­fer­ence of far more than a few per­cent. A dol­lar is a dol­lar, no mat­ter where it comes from. If you and 99 other peo­ple each donate $100 to 100 char­i­ties, then it’s the same as if each per­son donated$10,000 to just 1 char­i­ty. The only dif­fer­ence is what­ever over­head there might be; and even if we say that our Fold­ing@Home con­trib­u­tors lose 50% to over­head and the char­i­ties wind up get­ting only $6 mil­lion in their bank accounts to use, that’s still thou­sands more lives saved than by run­ning Fold­ing@Home and wast­ing the same amount of mon­ey!6 • per­haps one wor­ries that it’s eas­ier to run a FAH client than to donate reg­u­lar­ly; leav­ing aside the basic fact that a one-­time dona­tion through Pay­pal is a lot eas­ier than installing FAH, check­ing that it works, and per­pet­u­ally sysad­min­ning one’s com­puter for it, there are many ways to make donat­ing eas­i­er. One could set up a recur­ring dona­tion. One could annu­ally flip a coin to decide whether to donate twice the usual amount. (Or one could roll a 1d10 dice every year, agree­ing to donate 10× one’s annual dona­tion.) One could ask for oth­ers to donate in lieu of a birth­day or Christ­mas pre­sent; one could take advan­tage of employer match­ing plans. And so on. (If ease of con­tribut­ing is one’s true rea­son, then good news—it is very eas­ily dealt with!) Benefits But hey, per­haps it’s done good research that will save even more lives. Biol­o­gy, hell yeah! Wikipedia has a of 75 papers pub­lished draw­ing in some way on FAH. That is an aver­age of 7.5 papers per year. The skep­tic will notice that not a few (espe­cially early papers, nat­u­ral­ly) seem more con­cerned with FAH per se than with actual new results gen­er­ated by it, and that project lead Vijay Pande seems to be author or co-au­thor on almost all of the papers, which does­n’t indi­cate a large research com­mu­nity around the large invest­ment of FAH. None of them seem impor­tant, and the num­ber of pub­li­ca­tions seems to have peaked back in 2005–2006. The few actual com­pounds seem stalled in their test tubes. And a reminder, the wasted money amounts to many thou­sands of lives; for these sort of stakes, one would hope that one had good evi­dence, not mere pos­si­bil­i­ty. But let’s lower stan­dards and ask for ordi­nary evi­dence. What rea­son do we have to think Fold­ing@Home has poten­tial to save mil­lions of lives? It has been oper­at­ing for nearly 11 years. 11! And noth­ing that has yet saved a sin­gle life. (Read­ers are invited to pro­vide a coun­ter-ex­am­ple.) At what point do we stop talk­ing about its poten­tial to save mil­lions of lives and about the pos­si­bil­ity of around Mer­cury? ‘Basic research’ is great and all, but at some point enough is enough. Would­n’t it be bet­ter to redi­rect efforts to, say, —which besides not being pol­lut­ing, is a fun puz­zle game that can solve long-­s­tand­ing prob­lems?7 As I am not a biol­o­gist nor omni­scient, I can’t say for sure that the FAH work was­n’t use­ful, and I cer­tainly could­n’t say it looks pretty worth­less. But I feel much more com­fort­able assert­ing that the$12.65m could have been bet­ter spent on sav­ing those 12,000 peo­ple (and the 12,000 peo­ple the year before, and the 12,000 the year before that, and…).

Updating on evidence

The hope function

No one has yet shown me any­thing valu­able done by FAH; but, they argue, it may yet pro­duce some­thing next year or per­haps the year after that. This is a sta­tis­ti­cal argu­ment. There is an inter­est­ing per­spec­tive on this sug­ges­tion, known as the . The idea is that you are look­ing at each time period (each year) for some sort of sin­gu­lar event (a dis­cov­ery jus­ti­fy­ing FAH’s air-pol­lu­tion and fatal­i­ties), but you also have a cer­tain prob­a­bil­ity for the event never hap­pen­ing, for exam­ple because it is impos­si­ble. As each time period passes with­out the event, unsur­pris­ing­ly, one’s prob­a­bil­ity that the event will hap­pen the next period increas­es—as does the belief the event is impos­si­ble! (This isn’t con­tra­dic­to­ry.)

Example: AI

Inter­est­ing­ly, depend­ing on one’s belief as to the prob­a­bil­ity dis­tri­b­u­tion, the increases can be very small. One exam­ple I’ve run con­cerns the cre­ation of Arti­fi­cial Intel­li­gence; AI enthu­si­asts have sug­gested AI could be invented any­where in a very broad time span (Alan Tur­ing, for exam­ple, putting it at post-1990). Many of those dates (like 1990) have already passed, and AI has clearly not hap­pened. This has lead to a skep­ti­cism and belief that AI enthu­si­asts are irra­tional dream­ers who refuse to update on the evi­dence; but what does the hope func­tion say? I and oth­ers ran some esti­mates and found that for rea­son­able val­ues, the fail­ure of AI to hap­pen only min­i­mally sup­ports the belief that AI is impos­si­ble!

For exam­ple, sup­pose one thought AI if pos­si­ble would show up in the 100 years between 2000 and 2100 (a super­set of many expert fore­casts), that its pos­si­bil­ity was 90% (0.9), but it was an essen­tially ran­dom break­through which could as eas­ily hap­pen in 2000 as 2099 (a ). In x years, what is our new belief about its pos­si­bil­i­ty? The equa­tion goes:

We sub­sti­tute in our 100 years and 90% belief:

We plug the year of inter­est into the hope func­tion and find that in 2050, or in 50 years (), our orig­i­nal 90% belief AI is pos­si­ble has fallen only 9%, to 81%! (In­tu­itively8, one might think that a half-­cen­tury of fail­ure would count for more, but it does­n’t.) Even in 2090, after 90 years of fail­ure (), our view that AI is impos­si­ble has fallen only to 47%.

If we adopt a more real­is­tic dis­tri­b­u­tion like a cen­tered in 2050, we get sim­i­lar result­s—very small decreases each year until 2050, fol­lowed by a sud­den plum­met (as most of the chances get used up each year at or near the peak of the bell curve), and then a grad­ual curve down­wards to 0% as 2100 draws near.

Waiting for Foldot

Set­ting up a hope func­tion equa­tion for FAH is con­cep­tu­ally clear: what is the full life­time of FAH, how many years has it run with­out find­ing such a thing, what is the prob­a­bil­ity it will ever find any­thing use­ful, and what is the dis­tri­b­u­tion of said prob­a­bil­i­ty?

1. Inter­net dis­trib­ut­ed-­com­put­ing projects are young enough that there’s no clear life­cy­cle. looks like it will never shut down (hope for E.T. springs eter­nal), other projects shut down after a year or two or fin­ished their goals, like a num­ber of the cryp­to­graphic brute-­forc­ing projects set up to resolve the to defin­i­tively prove DES inse­cure and force the world to move onto more secure algo­rithms. ( has com­pleted 9 such chal­lenges but con­tin­ues 2 long-term tasks, whose value I am dubi­ous about.) Let’s guess and say FAH has another decade to go, for a total n = 20
2. FAH began in 2000 or so, so we can put x at 11
3. This is the con­tentious one. Let’s see what 90% does—that is surely a favor­able prob­a­bil­ity
4. We could choose either the uni­form or bell curve dis­tri­b­u­tion; both are attrac­tive as mod­els (FAH as serendip­ity machine! FAH as project gear­ing up in the early years for its matu­rity and even­tual decline!) But as it hap­pens, we already put n = 20 and x = 11, so we’d be in the mid­dle of the bell curve and that’s almost the same as the uni­form dis­tri­b­u­tion. So we’ll use the sim­pler uni­form dis­tri­b­u­tion.

We sub­sti­tute in: or a fall of 10% in our belief that FAH will ever pro­duce any­thing, which is a lit­tle sober­ing. Switch­ing to the bell curve dis­tri­b­u­tion will only make mat­ters worse; recall that after the mid­point, the bell curve plum­met­ed. With a bell curve cen­tered on 2010 or 2011 and an end date of 2020, the hope func­tion drops the 90% belief down to the <30% range by 2012 or 2013. (Fur­ther exam­ples would not be enlight­en­ing; the reader can cal­cu­late out the vari­a­tions him­self.)

I leave some ques­tions for the FAH enthu­si­ast to pon­der: Are these unrea­son­able assump­tions? (Aren’t they favor­able?) Do you have any real rea­son to believe that FAH’s dis­cov­er­ies ought to be heav­ily back­-load­ed, that one would expect it to take, say, 15 years (con­sum­ing ~2 ter­awat­t-hours)? Would you have argued for this back­-load­ing before read­ing about the hope func­tion? If you do not dis­pute the assump­tions, have you actu­ally dropped your faith in FAH by ~10%? Or at all? What sort of evi­dence would con­vince you FAH is harm­ful?

Doing it less wrong: Rosetta @ home

So what’s the right way then? Look at a very sim­i­lar grid com­put­ing pro­ject, (RAH). RAH has only 1⁄175th the com­put­ing power of FAH and pre­sum­ably con­sumes pro­por­tion­ately less elec­tric­i­ty; hence it directly kills peo­ple a year and indi­rectly kills . This is three orders of mag­ni­tude fewer deaths. And with its applied focus, the are a lit­tle bit bet­ter than FAH.

And RAH has made an addi­tional con­tri­bu­tion in demon­strat­ing how new hybrid approaches to the pro­tein-­fold­ing prob­lem might work; I pre­vi­ously men­tioned Foldit crack­ing a long-­s­tand­ing pro­tein-­fold­ing prob­lem, but Foldit did this by refin­ing a RAH-generated approx­i­ma­tion—and not a FAH approx­i­ma­tion. Foldit failed on the other tar­gets where it had no RAH start­ing point.9 (As far as I can tell, FAH has never par­tic­i­pated in that par­tic­u­lar com­pe­ti­tion, so one can­not com­pare its results with either RAH or RAH+Foldit.)

When you run a few num­bers, this seems like a pretty uncon­tro­ver­sial con­clu­sion. Lots of things are worse char­i­ties than the best char­i­ties (al­most by def­i­n­i­tion); why be so wed­ded to the idea that FAH is not one of those char­i­ties? (Why do geeks in par­tic­u­lar seem offended by crit­i­cism of FAH?) I think it has to do with our real rea­sons for a lot of things—­so­cial sta­tus10. Phil­an­thropy is often for such worth­less activ­i­ties (does the MoMA really need dona­tions from its board of direc­tors so it can buy the lat­est art­work to have been priced into the stratos­phere?) and peo­ple so unin­ter­ested in whether the char­ity actu­ally helps11 that the truth of the mat­ter—a straight­for­ward cash-­for-s­ta­tus bar­gain—is obvi­ous12, but it’s not so obvi­ous that char­i­ties them­selves seek sta­tus-rais­ing activ­i­ties and so are biased towards fund­ing bizarre & novel new activ­i­ties13—and what is more bizarre & novel than build­ing a world­wide super­com­puter to cal­cu­late the fold­ing of pro­teins?

It is sad and pitiable that we spend so many bil­lions on things like dog food and cos­met­ics rather than sav­ing lives; but isn’t it even sad­der that we can avoid that error, and try to do good, and still fail? The only thing sad­der, I think, would be if we could know of our fail­ure and go on sup­port­ing FAH. If char­ity truly was not about help­ing.

Appendix

“The Exposed Nest”, Frost

, “The Exposed Nest” (first pub­lished in , 1916):

YOU were forever finding some new play.
So when I saw you down on hands and knees
In the meadow, busy with the new-cut hay,
Trying, I thought, to set it up on end,
I went to show you how to make it stay,
If that was your idea, against the breeze,
And, if you asked me, even help pretend
To make it root again and grow afresh.
But 'twas no make-believe with you to-day,
Nor was the grass itself your real concern,
Though I found your hand full of wilted fern,
Steel-bright June-grass, and blackening heads of clover.
'Twas a nest full of young birds on the ground
The cutter-bar had just gone champing over
(Miraculously without tasting flesh)
And left defenseless to the heat and light.
You wanted to restore them to their right
Of something interposed between their sight
And too much world at once-could means be found.
The way the nest-full every time we stirred
Stood up to us as to a mother-bird
Whose coming home has been too long deferred,
And care for them in such a change of scene
And might our meddling make her more afraid.
That was a thing we could not wait to learn.
We saw the risk we took in doing good,
But dared not spare to do the best we could
Though harm should come of it; so built the screen
All this to prove we cared. Why is there then
No more to tell? We turned to other things.
I haven't any memory---have you? -
Of ever coming to the place again
To see if the birds lived the first night through,
And so at last to learn to use their wings.

1. The most effi­cient super­com­put­ers do bet­ter in terms of ; see the Green500 list. For exam­ple, man­ages to turn 1.2 megawatts into 1.19 petaflops.↩︎

2. In 2011, Ger­many decided to begin shut­ting down its due to the ; ignor­ing the points of dis­analo­gies (Ger­many is not sub­ject to tsunamis, is not run­ning out­dated designs from the 1960s, has more effec­tive cor­po­rate gov­er­nance etc), this was a pre­dictable recipe for not replac­ing it all with renew­able power (as envi­ron­men­tal­ists might hope) but increas­ing use of filthy fuel sources. This has indeed hap­pened; from “Europe con­sum­ing more coal”, Wash­ing­ton Post (2013-02-07):

In Ger­many, which by some mea­sures is pur­su­ing the most wide-rang­ing green goals of any major indus­tri­al­ized coun­try, a 2011 deci­sion to shut­ter nuclear power plants means that domes­ti­cally pro­duced , also known as brown coal, is fill­ing the gap . Power plants that burn the sticky, sul­furous, high­-e­mis­sions fuel are run­ning at full throt­tle, with many tal­ly­ing 2012 as their high­est-de­mand year since the early 1990s. Sev­eral new coal power plants have been unveiled in recent month­s—even though solar panel instal­la­tions more than dou­bled last year…Lig­nite sup­plied 25.6% of Ger­many’s elec­tric­ity in 2012, up from 22.7% in 2010. Hard black coal sup­plied an addi­tional 19.1% last year, and it was also on the rise.

Above I quoted esti­mates that coal kills 15 peo­ple per ter­awat­t-hour. says Ger­many used 590 ter­awat­t-hours in 2010; lig­nite burn­ing increased by 25.6 − 22.7 = 2.9%. So how many direct lig­nite-­coal-based deaths have been caused by just 2 years of just Ger­mans pan­ick­ing over Fukushi­ma? .

Wikipedia lists no direct deaths from Fukushima and even­tual plant worker deaths; Hoeve & Jacob­son 2012 (dis­cus­sion) esti­mated that total global radi­a­tion-re­lated deaths from Fukushima will most likely be 180 and will not exceed 2500. Hence, Ger­many may already have killed more peo­ple due to its fear of Fukushi­ma-­like events than Fukushima killed.↩︎

3. $12.65m may be low for the cost of pow­er. A March 2011 pre­sen­ta­tion by Kathy Yelick, “Exas­cale Com­put­ing: Tech­ni­cal Chal­lenges”, esti­mates each megawatt used at$1m, which obvi­ously adds on another $2.5m to our esti­mate, which we esti­mated con­ser­v­a­tively in any case.↩︎ 4. An inter­est­ing exam­ple being the art project , which is exactly what it sounds like, and earned the KLF global lash­ings of scorn and crit­i­cism about waste—de­spite the fact that the sole con­se­quence of burn­ing some money is some altru­is­tic redis­tri­b­u­tion of wealth (ev­ery­one else’s quids are worth slightly more) and that there would not have been a whis­per of cen­sure if they had instead self­ishly bought a man­sion or an Aston Mar­tin car or over­priced wine. The asym­me­try of the reac­tion to the burn­ing and reac­tion to nor­mal prof­li­gate con­spic­u­ous con­sump­tion through­out soci­ety is fas­ci­nat­ing and says some­thing about human cog­ni­tive biases with regard to acts of commission/omission, I think, as well as wide­spread illu­sions about money & wealth.↩︎ 5. A non-nu­tri­en­t-based approach would be mid­wife train­ing; “Even a small pilot project cost­ing only$20,244 saved the lives of 97 infants, the authors esti­mat­ed, mean­ing that it cost just $208 per life saved.”↩︎ 6. The mar­ginal effec­tive­ness of the best char­i­ties is huge; the best char­i­ties do orders of mag­ni­tude more good than mediocre or bad char­i­ties. A thought-ex­am­ple from “Effec­tive Char­ity” A hypo­thet­i­cal char­ity run­ning pro­grams like Vil­lageReach’s which embez­zled 95% of its bud­get and had cor­re­spond­ingly greatly reduced cost-­ef­fec­tive­ness would still be doing far more good per dol­lar than the Make-A-Wish Foun­da­tion or the least effec­tive devel­op­ing world char­i­ties do. This exam­ple makes it clear how pro­foundly use­less the over­head ratio is for assess­ing the rel­a­tive qual­ity of a char­i­ty. This holds true for less mor­tal char­i­ties; Nicholas Kristof cites an exam­ple with 2 orders of mag­ni­tude dif­fer­ence: In much of the devel­op­ing world, most kids have intesti­nal worms, leav­ing them sick, ane­mic and more likely to miss school. Deworm­ing is very cheap (a pill cost­ing a few pen­nies), and, in the exper­i­ment he did with Edward Miguel, it resulted in 25% less absen­teeism. Even years lat­er, the kids who had been ran­domly cho­sen to be dewormed were earn­ing more money than other kids. Kre­mer esti­mates that the cost of keep­ing a kid in school for an addi­tional year by build­ing schools or by sub­si­diz­ing school uni­forms is more than$100, while by deworm­ing kids, the cost drops to $3.50. (In a pinch, kids can usu­ally go to ‘school’ in a church or mosque with­out a uni­for­m.) ↩︎ 7. Orig­i­nal 2011 paper: “Crys­tal struc­ture of a monomeric retro­vi­ral pro­tease solved by pro­tein fold­ing game play­ers”. Gen­eral back­ground on Foldit: , Cooper et al 2010.↩︎ 8. For exam­ple, bcoburn remarks: “I per­son­ally expected that P(AI) would drop-off roughly lin­early as n increased, so this cer­tainly seems coun­ter-in­tu­itive to me.”↩︎ 9. From “Crys­tal struc­ture of a monomeric retro­vi­ral pro­tease solved by pro­tein fold­ing game play­ers” 2011: De novo struc­ture pre­dic­tion remains an excep­tion­ally chal­leng­ing prob­lem, and very few pre­dic­tions with atomic accu­racy have been made in the his­tory of [the com­pe­ti­tion] . For CASP9 tar­get T0581, start­ing from an extended chain, the Rosetta Server, which car­ried out a large-s­cale search for the low­est­-en­ergy struc­ture using com­put­ing power from RAH vol­un­teers (http://boinc.bakerlab.org/rosetta/), pro­duced a remark­ably accu­rate model (Fig. 1a; com­pare red and blue). How­ev­er, the server ranked this model fourth out of the five sub­mis­sions. The Foldit Void Crush­ers team cor­rectly selected this near-­na­tive model and fur­ther improved it by accu­rately mov­ing the ter­mi­nal helix, pro­duc­ing the best model for this tar­get of any group and one of the best over­all pre­dic­tions at CASP9 (ref. 4) (Fig. 1a; com­pare yel­low and blue). Thus, in a sit­u­a­tion where one model out of sev­eral is in a near-­na­tive con­for­ma­tion, Foldit play­ers can rec­og­nize it and improve it to become the best mod­el. Unfor­tu­nately for the other Free Mod­el­ing tar­gets, there were no sim­i­larly out­stand­ing Rosetta Server start­ing mod­els, so Foldit play­ers sim­ply tun­neled to the near­est incor­rect local min­i­ma.” ↩︎ 10. Sta­tus as I use it here is a bit com­plex, more than a lit­tle idio­syn­crat­ic, and as much a par­a­digm as any sim­ple prop­er­ty. To get an idea of what I mean, see , Less­Wrong/Over­com­ingBias, and descrip­tion of in Dis­cover your Inner Econ­o­mist.↩︎ 11. It’s inter­est­ing how lit­tle effort goes into eval­u­at­ing or rank­ing char­i­ties; the num­ber of orga­ni­za­tions can be counted on one hand, and one of the most promi­nent attempts, , is amaz­ingly recent (2006). In this vein, I was struck by a com­ment in the New York Time’s “For Fed­eral Pro­grams, a Taste of Mar­ket Dis­ci­pline”: “A recent review found that 10 major social pro­grams had been rig­or­ously eval­u­ated over the past two decades, using the sci­en­tific gold stan­dard of ran­dom assign­ment. Only one of the 10—, for infants, tod­dlers preg­nant wom­en—was a clear suc­cess. Yet all 10 still exist, and largely in their orig­i­nal form.” ↩︎ 12. The New Yorker, with its focus on New York City’s upper-crust, recently made this clear to me yet again with its cov­er­age of the , who would more usu­ally be greeted by said upper-crust with deri­sion than acclaim; from 2010’s “Covert Oper­a­tions: The bil­lion­aire broth­ers who are wag­ing a war against Obama”: "On May 17th, a black­-tie audi­ence at the Met­ro­pol­i­tan Opera House applauded as a tall, jovial-look­ing bil­lion­aire took the stage. It was the sev­en­ti­eth annual spring gala of Amer­i­can Bal­let The­atre, and David H. Koch was being cel­e­brated for his gen­eros­ity as a mem­ber of the board of trustees; he had recently donated$2.5 mil­lion toward the com­pa­ny’s upcom­ing sea­son, and had given many mil­lions before that. Koch received an award while flanked by two of the gala’s co-chairs, Blaine Trump, in a peach-­col­ored gown, and Car­o­line Kennedy Schloss­berg, in emer­ald green. Kennedy’s moth­er, Jacque­line Kennedy Onas­sis, had been a patron of the bal­let and, coin­ci­den­tal­ly, the pre­vi­ous owner of a Fifth Avenue apart­ment that Koch had bought, in 1995, and then sold, eleven years lat­er, for thir­ty-two mil­lion dol­lars, hav­ing found it too small.

The gala marked the social ascent of Koch, who, at the age of sev­en­ty, has become one of the city’s most promi­nent phil­an­thropists. In 2008, he donated a hun­dred mil­lion dol­lars to mod­ern­ize Lin­coln Cen­ter’s New York State The­atre build­ing, which now bears his name. He has given twenty mil­lion to the Amer­i­can Museum of Nat­ural His­to­ry, whose dinosaur wing is named for him. This spring, after notic­ing the decrepit state of the foun­tains out­side the Met­ro­pol­i­tan Museum of Art, Koch pledged at least ten mil­lion dol­lars for their ren­o­va­tion. He is a trustee of the muse­um, per­haps the most cov­eted social prize in the city, and serves on the board of Memo­r­ial Sloan-Ket­ter­ing Can­cer Cen­ter, where, after he donated more than forty mil­lion dol­lars, an endowed chair and a research cen­ter were named for him.

One dig­ni­tary was con­spic­u­ously absent from the gala: the even­t’s third hon­orary co-chair, Michelle Oba­ma. Her office said that a sched­ul­ing con­flict had pre­vented her from attend­ing. Yet had the First Lady shared the stage with Koch it might have cre­ated an awk­ward tableau…"

A quick cal­cu­la­tion: , so a vastly incom­plete tally of the Koch dona­tions is $200 mil­lion or roughly 200,000 dead Africans. Does any­one want to argue that the Kochs’ phil­an­thropy is even remotely close to being effi­cient, and that these dona­tions were any­thing but pur­chas­ing sta­tus? In some cas­es, one won­ders why they even pre­tend; “More Cash to Go to a Hall Than to Haiti”, The New York Times: “Even if the even­t’s nearly$200,000 worth of tick­ets sell out, less than $8,000 from the sales will go to the cause. The con­cert, though, is expected to raise some mon­ey, thanks mainly to a$50,000 sub­sidy by the Mont­blanc com­pany and $10,000 by CAMI Music, the con­cert’s pre­sen­ter and Mr. Lang’s man­age­ment agen­cy…No hard and fast guide­lines exist on how much money raised in a ben­e­fit should go for expens­es, and it is not unusual for galas to raise lit­tle money or even lose it…In an account­ing pro­vided by CAMI Music, the costs will total$181,590. If the hall sells out, box office pro­ceeds will total $189,793, exclud­ing com­pli­men­tary tick­ets.” ↩︎ 13. From the GiveWell blog, “After”Extra­or­di­nary and Unortho­dox" comes the Val­ley of Death" (re­plac­ing the rel­e­vant adjec­tives with ‘sta­tus’ is left as an exer­cise for the read­er): “…it’s hard for me to see a big dif­fer­ence between it and the$100 mil­lion Gates Grand Chal­lenges Explo­rations,”a unique ini­tia­tive that sup­ports inno­v­a­tive research of unortho­dox ideas" in global health (though the pro­posal above does not explic­itly spec­ify a sec­tor, all three of its exam­ples are in global health as well).

Speak­ing more infor­mal­ly, I’ve heard sim­i­lar con­cepts empha­sized by most major fun­ders I’ve spo­ken with. Any­one who has dealt with major foun­da­tions should rec­og­nize the desire to find a com­pletely new, rev­o­lu­tion­ary, neglected oppor­tu­nity that just needs some seed fund­ing to explode.

I do believe that the best oppor­tu­ni­ties are the under­-­funded ones. Yet I’m not sure that tiny, neglected inno­va­tions are the best places to look for these oppor­tu­ni­ties—­pre­cisely because that’s where all the major fun­ders seem to be look­ing. I sub­mit that the bet­ter place to look for neglected oppor­tu­ni­ties is the “val­ley of death” between proof of con­cept and large-s­cale roll­out.

…There’s no glory in fund­ing the roll­out. Vil­lageReach already has shown that what it’s doing has worked; nobody can claim to be bril­liant for spot­ting it. And Vil­lageReach does­n’t need help design­ing its pro­gram (this has been cited to me explic­itly as a draw­back from the per­spec­tive of some major fun­der­s)."

“Atti­tudes toward evi­dence seem less key than we would have guessed. When we started GiveWell, we and most of our sup­port­ers imag­ined that new cus­tomers could be found in cer­tain indus­tries where peo­ple are accus­tomed to using mea­sure­ment to eval­u­ate and learn from their deci­sions. We hoped these peo­ple would res­onate with our desire to bring feed­back loops into areas where feed­back loops don’t nat­u­rally exist. But we’ve found that a lot of them don’t, largely because impact isn’t the main thing they’re aim­ing for when they give. Peo­ple give for many rea­son­s—to main­tain friend­ships, to over­come guilt and cog­ni­tive dis­so­nance, to achieve recog­ni­tion—and a given donor is unlikely to be inter­ested in GiveWell unless achiev­ing impact is at the top of his/her list.”

On the other hand,

“GiveWell cus­tomers never seem inter­ested in pub­lic recog­ni­tion. In our first year, we con­sid­ered post­ing acknowl­edg­ments to major sup­port­ers on our web­site, but there was no inter­est. Since then, we have had many cus­tomers who require anonymity (even when we ask them to take pub­lic credit for our sake) and no cus­tomers who’ve requested that we pub­licly thank them or oth­er­wise help them get recog­ni­tion.”

It is a lit­tle dif­fi­cult for Han­son­ian the­o­ries of char­ity to explain wholly anony­mous char­i­ty. Pos­si­bly char­ity in such cases is due to sig­nal­ing within a small group, rather than pub­lic sig­nal­ing to thou­sands or mil­lions of peo­ple (the gen­eral pop­u­lace). And who knows; maybe there is gen­uine altru­ism in this world of dust and delu­sions.↩︎