Automatic for the people

Sum­mary: Things we could do about tech­nolog­i­cal un­em­ploy­ment, if there was tech­nolog­i­cal un­em­ploy­ment. (Novel bit is a top-down macroe­co­nomic calcu­la­tion.)

Con­fi­dence: 90% that the solu­tions I call ‘vi­cious’ would be. 60% that the ones I call ‘non­vi­cious’ would be. Worth em­pha­sis­ing up here that there is lit­tle ev­i­dence for tech’ un­em­ploy­ment right now.

Cross­posted from gleech.org.


Au­tonomous trucks are now in use and are already safer and more fuel-effi­cient than hu­man driven ones. (Truck drivers are ~2% of the en­tire Amer­i­can work­force.)

Crap jour­nal­ism (that is, 80% of (UK) jour­nal­ism) is now fully au­tomat­able. Au­to­matic art is quite good and im­prov­ing fast. Con­sider also the cock­tail bar­tender. And so on: maybe half of all jobs are at risk of be­ing au­to­mated, as­sum­ing the rate of AI progress just stays con­stant (“over an un­speci­fied pe­riod, per­haps a decade or two”).

Au­toma­tion is maybe the main way that tech­nol­ogy im­proves most peo­ple’s lives: aside from sta­tus ex­cep­tions like Ap­ple prod­ucts, big re­duc­tions in man­u­fac­tur­ing cost usu­ally mean big re­duc­tion in the end cost of goods. Ob­vi­ously, re­plac­ing labour costs with lower-marginal-cost ma­chines benefits rich ma­chine-own­ers most, but au­toma­tion also al­lows gi­ant price cuts in all kinds of things; over the last two cen­turies these cuts have trans­formed so­ciety, in­creas­ing equal­ity enor­mously by mak­ing things af­ford­able for the first time.

Be­sides the ob­vi­ous ex­am­ple—that we now pro­duce a vol­ume of food far be­yond the needs of the en­tire world pop­u­la­tion (2940kcal per per­son per day, though with ter­rible dis­tribu­tive failures) - con­sider that a sin­gle or­di­nary shirt takes 508 hours of labour to pro­duce on a spin­ning wheel and hand-loom—so you would ex­pect to pay some­thing above $3600 at cur­rent min­i­mum wage (still $900 at 1400CE wage lev­els).

Get­ting costs as near to zero as pos­si­ble is the way we will solve the easy prob­lem of hu­man ex­is­tence, scarcity of ba­sic goods and leisure. (The hard prob­lem of hu­man life is bore­dom and dig­nity and mean­ing and all that.)

Types of automation

I couldn’t find a rigor­ous list of the types of au­toma­tion, so here’s my at­tempt:

  • First-wave: Me­chan­i­cal con­trol. ma­chines which perform one task au­to­mat­i­cally be­cause of the ex­act shape of their com­po­nents. e.g. a) Cams, linked wheels ma­chined to cre­ate fixed se­quences of lin­ear mo­tions. an in­stance in the year 200BCE. b) Gover­nors. c) Cataracts.

  • Se­cond-wave: Numer­i­cal con­trol. Pro­grammable ma­chines, where chang­ing the pro­gram (punch-cards) changes the product. Grossly phys­i­cal, ana­log al­gorithms. The first fully-auto pro­duc­tive ma­chine was de­signed and built by Vau­can­son in 1745 (not Fal­con or Bou­chon or Jac­quard).

  • Third-wave: Com­puter nu­mer­i­cal con­trol. One ma­chine, whose be­havi­our is dic­tated by al­gorithms en­coded as digi­tal data. The birth of soft­ware. Me­mory al­lows for ~zero hu­man over­sight once task in­struc­tions are cre­ated and loaded. Ar­bi­trar­ily fine man­ual work, performed bet­ter than hu­mans and much faster. 1957 or 1958.

  • Fourth-wave: Ma­chine-learn­ing model con­trol: No need to pro­gram, af­ter suit­able do­main al­gorithms have been writ­ten; mod­els are au­to­mat­i­cally cal­ibrated given enough data. Trained mod­els can go be­yond or­di­nary con­trol flow, dis­cern­ing non-nec­es­sary and non-suffi­cient con­di­tions in­volved in so­phis­ti­cated tasks like driv­ing. Early suc­cesses came in speech recog­ni­tion (e.g. Tan­gora and DRAGON in the early 80s, Rabiner et al in 1985), be­gin­ning the long march to re­place stenog­ra­phers and sec­re­taries (and ul­ti­mately ex­per­i­men­tal lin­guists). The defin­ing in­stance of the fourth wave may be driver­less cars: 5% of de­vel­oped-world jobs at risk. Some­times called the ‘sec­ond ma­chine age’.

  • Fifth-wave: Ar­tifi­cial gen­eral in­tel­li­gence: No need for pro­gram­ming (pro­gram­ming ei­ther new tasks, or any fu­ture de­sign of im­proved ma­chines).


So au­toma­tion has been hap­pen­ing for (many) hun­dreds of years, but in the past it prob­a­bly didn’t pro­duce long-term, “tech­nolog­i­cal” un­em­ploy­ment. This is prob­a­bly be­cause peo­ple were eas­ily able to think up new pro­fes­sions given new tech and cul­ture, and since the in­creased pro­duc­tivity trans­lated into lower costs for the au­to­mated good, which stim­u­lated other parts of the econ­omy, and peo­ple could re­train for the sub­se­quent new jobs. The pre­sent wave might be differ­ent: it might ac­tu­ally re­duce the num­ber of available jobs per­ma­nently, be­cause the ma­chines now en­ter­ing the work­force can be ap­plied to many of the jobs that peo­ple could re­train to do.

If so, our econ­omy—re­source al­lo­ca­tion based on em­ploy­ment (which we use as a poor proxy vari­able for pro­duc­tivity) - is a lo­cal max­i­mum and we can­not ex­pect to ar­rive at a good out­come with­out ac­tivism, since:

  1. The new ma­chine-learn­ing au­toma­tion could fully re­place around half of jobs.

  2. Since these jobs in­volve even our high­est cog­ni­tive fac­ul­ties, it is pos­si­ble that we won’t think up new pro­duc­tive jobs for the re­placed work­ers, like we did in the past.

  3. So au­toma­tion could pro­duce an un­prece­dent­edly high un­em­ploy­ment rate, ~60%.

  4. Most ex­ist­ing un­em­ploy­ment welfare sys­tems are in­ad­e­quate and de­grad­ing.

  5. So with­out in­ter­ven­tion, a crash in hu­man welfare is eas­ily pos­si­ble.

But, un­less we au­to­mate a lot more, we the species will never have enough wealth to offer a de­cent ba­sic in­come, and ev­ery­one will con­tinue to waste half their lives at work. Like C20th peas­ants.


How much is there for ev­ery­one?

You can fol­low my calcu­la­tion here.

Gross world product di­vided by pop­u­la­tion is $10,600 per per­son! - but this naive dis­tri­bu­tion would be im­pos­si­ble, even as­sum­ing all the poli­ti­cal will in the world. De­pre­ci­a­tion costs bring this down to ~$9,300 ; main­tain­ing our pre­sent lev­els of R&D in­vest­ment brings it down fur­ther to $9,100.

We also have to con­sider the “dead­weight loss” of tax­a­tion (how much you have to spend to col­lect the tax + how much un­pro­duc­tive tax avoidance be­havi­our you cause + how much it dis­cour­ages eco­nomic ac­tivity + etc). The re­search on this is shock­ingly vague (this gives es­ti­mates be­tween 2.5% and 30%!). Lower bound takes us to $8,800; the 30% up­per bound takes us all the way down to $5,900. The mass carve-up we’re talk­ing about goes well be­yond any ex­ist­ing tax rate, and avoidance does scale in pro­por­tion to rates; so we prob­a­bly have to as­sume the dead­weight would be worse than any yet ex­pe­rienced. Call it 30%.

Most peo­ple will want to main­tain gov­ern­ment ser­vices at around their cur­rent level (be­sides the gi­ant ba­sic in­come ex­pen­di­ture); re­mem­ber that this could knock an­other 30% off our available in­come flow (or a mere 28% off if we lose the mil­i­tary). Half of that is welfare and pen­sions, which are be­ing re­placed here (in our heads). So we’re down to $4,300.

So the cur­rent econ­omy, carved up sus­tain­ably, would yield some frac­tion of $4,000 per per­son per year. Even given that most house­holds would get about 4 of these in­comes, this is sim­ply not enough for free­dom, given the needs or tastes of an av­er­age hu­man.

(The above does not con­sider a host of other sad re­al­ities: e.g. rich peo­ple like their money; e.g. this much equal­ity would de­stroy en­tire in­dus­tries (lux­ury goods!), and so fur­ther sap the available pie; e.g. there would be a re­ces­sion effect from all the fully-alienated work­ers down­ing their tools—yes, there could also be a stim­u­lus effect from in­creas­ing poor peo­ple’s spend­ing, but it’s ex­tremely difficult to say which sign pre­vails.

Worse: only the di­rect cost of tax­a­tion is fac­tored in above, with­out the amount that the rich man­age to spirit away. We would need some­thing like a world gov­ern­ment for it to work even this well (badly), to stamp out tax havens and trans­fer pric­ing and all that jolly fi­nan­cial danc­ing.)

My point is not that this is the ex­act figure we’ll have to work with - in­stead it points up our paltry pre­sent ca­pac­ity. Growth would be nec­es­sary, even in an ideal world with­out na­tion­al­ism, greed, in­effi­ciency (...)

So­cially con­scious peo­ple are these days am­biva­lent about eco­nomic growth, of­ten for en­vi­ron­men­tal rea­sons. But con­sider the en­light­ened defi­ni­tion of “pro­duc­tivity”: it is not “amount of out­put”, but the amount of out­put per unit of in­put. This is pure gain, and is ac­tu­ally en­vi­ron­men­tally pos­i­tive, since it could re­duce re­source use and waste. But we do need out­put growth too: e.g. un­til ev­ery para­plegic on earth who wants one of these has one.

The above just uses world in­come; what about us­ing world wealth? Even if we liqui­date the whole of the world’s wealth (our stock of money, as op­posed to the GDP, a flow), it would only provide a uni­ver­sal ba­sic in­come for three and a half years.


Prob­a­bly vi­cious solu­tions to the worst case:

  1. Halt progress on au­toma­tion, pre­serv­ing cur­rent em­ploy­ment. (Via gov­ern­ment ban, suc­cess­ful hos­tility by or­ganised labour, mass mon­key wrench­ing). I count this as vi­cious, even though it is much bet­ter than the worst-case, since it leaves al­most all of us very un­free, for­ever.

  2. or I guess you could try huge un­em­ploy­ment plus an au­thor­i­tar­ian crack­down on des­per­ate masses, see how that goes.

  3. Na­tion­al­ised robot fac­to­ries, or full cy­ber-com­mu­nism. Food and clothes and houses guaran­teed to all, at least. Leave aside the his­tor­i­cal failure of com­mand economies; imag­ine here that a new big-data Kan­torovich man­ages to make it fairly effi­cient.

    Even grant­ing this gi­ant as­sump­tion, this is a dan­ger­ous move. To­tal state con­trol of the means of pro­duc­tion is too eas­ily twisted. Now, this could be just my emo­tional over­re­ac­tion to read­ing about e.g. Maoist China, with its food ter­ror­ism and peas­ant-rob­bing. But it rings ma­lign: to­tal con­trol of pro­duc­tion by any en­tity is a ter­rible un­nec­es­sary risk.

  4. “Back to the land” prim­i­tivism. Hu­mans re­turn to sub­sis­tence farm­ing as a means of sur­vival.

  5. Just rais­ing the min­i­mum wage with­out do­ing any­thing else. Misses the point en­tirely. (‘Should the min­i­mum wage be called the “Robot Em­ploy­ment Act?”’ – Cowen and Tabar­rok.)

  6. Mass hu­man aug­men­ta­tion, to keep up with the ma­chines. (At min­i­mum, just tra­di­tional ex­ter­nally-hosted soft­ware: Hu­mans us­ing chess as­sis­tant pro­grams were beat­ing solo su­per­com­put­ers un­til rel­a­tively re­cently, c. 2006.) Pos­si­bly vi­cious, but not be­cause there’s any­thing wrong with tran­shu­man­ism: be­cause do­ing it for purely eco­nomic rea­sons is 1) a nev­erend­ing pro­cess, since the ma­chines will im­prove as fast, and 2) it would prob­a­bly be de­struc­tive of some dis­tinc­tive hu­man virtues (e.g. seren­ity, play, re­flec­tion, aes­thetic in­ter­est). Only good if peo­ple re­ally can’t feel dig­nified with­out hav­ing a lead­ing pro­duc­tive role in things.

Po­ten­tially non­vi­cious solu­tions:

  1. Prop up the liberal mixed econ­omy:

    • with a pro­gramme of mass em­ployee stock own­er­ship.

    • or by carv­ing each full-time job into sev­eral part-time ones, plus heavy wage sub­sidies.

    • or with a uni­ver­sal ba­sic in­come funded through higher tax­a­tion.

    • Get the gov­ern­ment to buy ev­ery 18 year old a se­ri­ous stock port­fo­lio (??)

2. More or less vague sug­ges­tions for a very differ­ent so­cial struc­ture, like em­bar­rass­ingly de­cen­tral­ised group­ings, with their own mini­fac­to­ries...


I am not very sure of any of the above; the ac­tual stats on pro­duc­tivity growth are wor­ry­ing for the op­po­site rea­son: it has been too slow to sup­port wages for a long time. Any­way other pow­er­ful forces (e.g. global out­sourc­ing, the de­cay of unions) be­sides robots have led to the 40-year de­cline in labour’s share of global in­come. But those will pro­duce similar dystopian prob­lems if the trend con­tinues, and there’s enough of a risk of the above sce­nario for us to put a lot of thought and effort into pro­tect­ing peo­ple, ei­ther way.


Fac­to­ries that run ‘lights out’ are fully au­to­mated and re­quire no hu­man pres­ence on-site… these fac­to­ries can be run with the lights off.

– Wiki

Not only is it lights-out—we turn off the air con­di­tion­ing and heat too.

– an ex­ec­u­tive at Fuji Au­to­matic Numer­i­cal Con­trol

No nominations.
No reviews.