No Evolutions for Corporations or Nanodevices

“The laws of physics and the rules of math don’t cease to ap­ply. That leads me to be­lieve that evolu­tion doesn’t stop. That fur­ther leads me to be­lieve that na­ture —bloody in tooth and claw, as some have termed it —will sim­ply be taken to the next level...
“[Get­ting rid of Dar­wi­nian evolu­tion is] like try­ing to get rid of grav­i­ta­tion. So long as there are limited re­sources and mul­ti­ple com­pet­ing ac­tors ca­pa­ble of pass­ing on char­ac­ter­is­tics, you have se­lec­tion pres­sure.”
—Perry Met­zger, pre­dict­ing that the reign of nat­u­ral se­lec­tion would con­tinue into the in­definite fu­ture.

In evolu­tion­ary biol­ogy, as in many other fields, it is im­por­tant to think quan­ti­ta­tively rather than qual­i­ta­tively. Does a benefi­cial mu­ta­tion “some­times spread, but not always”? Well, a psy­chic power would be a benefi­cial mu­ta­tion, so you’d ex­pect it to spread, right? Yet this is qual­i­ta­tive rea­son­ing, not quan­ti­ta­tive—if X is true, then Y is true; if psy­chic pow­ers are benefi­cial, they may spread. In Evolu­tions Are Stupid, I de­scribed the equa­tions for a benefi­cial mu­ta­tion’s prob­a­bil­ity of fix­a­tion, roughly twice the fit­ness ad­van­tage (6% for a 3% ad­van­tage). Only this kind of nu­mer­i­cal think­ing is likely to make us re­al­ize that mu­ta­tions which are only rarely use­ful are ex­tremely un­likely to spread, and that it is prac­ti­cally im­pos­si­ble for com­plex adap­ta­tions to arise with­out con­stant use. If psy­chic pow­ers re­ally ex­isted, we should ex­pect to see ev­ery­one us­ing them all the time—not just be­cause they would be so amaz­ingly use­ful, but be­cause oth­er­wise they couldn’t have evolved in the first place.

“So long as there are limited re­sources and mul­ti­ple com­pet­ing ac­tors ca­pa­ble of pass­ing on char­ac­ter­is­tics, you have se­lec­tion pres­sure.” This is qual­i­ta­tive rea­son­ing. How much se­lec­tion pres­sure?

While there are sev­eral can­di­dates for the most im­por­tant equa­tion in evolu­tion­ary biol­ogy, I would pick Price’s Equa­tion, which in its sim­plest for­mu­la­tion reads:

change in av­er­age char­ac­ter­is­tic = co­var­i­ance(rel­a­tive fit­ness, char­ac­ter­is­tic)

This is a very pow­er­ful and gen­eral for­mula. For ex­am­ple, a par­tic­u­lar gene for height can be the Z, the char­ac­ter­is­tic that changes, in which case Price’s Equa­tion says that the change in the prob­a­bil­ity of pos­sess­ing this gene equals the co­var­i­ance of the gene with re­pro­duc­tive fit­ness. Or you can con­sider height in gen­eral as the char­ac­ter­is­tic Z, apart from any par­tic­u­lar genes, and Price’s Equa­tion says that the change in height in the next gen­er­a­tion will equal the co­var­i­ance of height with rel­a­tive re­pro­duc­tive fit­ness.

(At least, this is true so long as height is straight­for­wardly her­i­ta­ble. If nu­tri­tion im­proves, so that a fixed geno­type be­comes taller, you have to add a cor­rec­tion term to Price’s Equa­tion. If there are com­plex non­lin­ear in­ter­ac­tions be­tween many genes, you have to ei­ther add a cor­rec­tion term, or calcu­late the equa­tion in such a com­pli­cated way that it ceases to en­lighten.)

Many en­light­en­ments may be at­tained by study­ing the differ­ent forms and deriva­tions of Price’s Equa­tion. For ex­am­ple, the fi­nal equa­tion says that the av­er­age char­ac­ter­is­tic changes ac­cord­ing to its co­var­i­ance with rel­a­tive fit­ness, rather than its ab­solute fit­ness. This means that if a Frodo gene saves its whole species from ex­tinc­tion, the av­er­age Frodo char­ac­ter­is­tic does not in­crease, since Frodo’s act benefited all geno­types equally and did not co­vary with rel­a­tive fit­ness.

It is said that Price be­came so dis­turbed with the im­pli­ca­tions of his equa­tion for al­tru­ism that he com­mit­ted suicide, though he may have had other is­sues. (Over­com­ing Bias does not ad­vo­cate com­mit­ting suicide af­ter study­ing Price’s Equa­tion.)

One of the en­light­en­ments which may be gained by med­i­tat­ing upon Price’s Equa­tion is that “limited re­sources” and “mul­ti­ple com­pet­ing ac­tors ca­pa­ble of pass­ing on char­ac­ter­is­tics” are not suffi­cient to give rise to an evolu­tion. “Things that repli­cate them­selves” is not a suffi­cient con­di­tion. Even “com­pe­ti­tion be­tween repli­cat­ing things” is not suffi­cient.

Do cor­po­ra­tions evolve? They cer­tainly com­pete. They oc­ca­sion­ally spin off chil­dren. Their re­sources are limited. They some­times die.

But how much does the child of a cor­po­ra­tion re­sem­ble its par­ents? Much of the per­son­al­ity of a cor­po­ra­tion de­rives from key officers, and CEOs can­not di­vide them­selves by fis­sion. Price’s Equa­tion only op­er­ates to the ex­tent that char­ac­ter­is­tics are her­i­ta­ble across gen­er­a­tions. If great-great-grand­chil­dren don’t much re­sem­ble their great-great-grand­par­ents, you won’t get more than four gen­er­a­tions’ worth of cu­mu­la­tive se­lec­tion pres­sure—any­thing that hap­pened more than four gen­er­a­tions ago will blur it­self out. Yes, the per­son­al­ity of a cor­po­ra­tion can in­fluence its spinoff—but that’s noth­ing like the her­i­ta­bil­ity of DNA, which is digi­tal rather than ana­log, and can trans­mit it­self with 10^-8 er­rors per base per gen­er­a­tion.

With DNA you have her­i­ta­bil­ity last­ing for mil­lions of gen­er­a­tions. That’s how com­plex adap­ta­tions can arise by pure evolu­tion—the digi­tal DNA lasts long enough for a gene con­vey­ing 3% ad­van­tage to spread it­self over 768 gen­er­a­tions, and then an­other gene de­pen­dent on it can arise. Even if cor­po­ra­tions repli­cated with digi­tal fidelity, they would cur­rently be at most ten gen­er­a­tions into the RNA World.

Now, cor­po­ra­tions are cer­tainly se­lected, in the sense that in­com­pe­tent cor­po­ra­tions go bust. This should log­i­cally make you more likely to ob­serve cor­po­ra­tions with fea­tures con­tribut­ing to com­pe­tence. And in the same sense, any star that goes nova shortly af­ter it forms, is less likely to be visi­ble when you look up at the night sky. But if an ac­ci­dent of stel­lar dy­nam­ics makes one star burn longer than an­other star, that doesn’t make it more likely that fu­ture stars will also burn longer—the fea­ture will not be copied onto other stars. We should not ex­pect fu­ture as­tro­physi­cists to dis­cover com­plex in­ter­nal fea­tures of stars which seem de­signed to help them burn longer. That kind of me­chan­i­cal adap­ta­tion re­quires much larger cu­mu­la­tive se­lec­tion pres­sures than a once-off win­now­ing.

Think of the prin­ci­ple in­tro­duced in Ein­stein’s Ar­ro­gance—that the vast ma­jor­ity of the ev­i­dence re­quired to think of Gen­eral Rel­a­tivity had to go into rais­ing that one par­tic­u­lar equa­tion to the level of Ein­stein’s per­sonal at­ten­tion; the amount of ev­i­dence re­quired to raise it from a de­liber­ately con­sid­ered pos­si­bil­ity to 99.9% cer­tainty was triv­ial by com­par­i­son. In the same sense, com­plex fea­tures of cor­po­ra­tions which re­quire hun­dreds of bits to spec­ify, are pro­duced pri­mar­ily by hu­man in­tel­li­gence, not a hand­ful of gen­er­a­tions of low-fidelity evolu­tion. In biol­ogy, the mu­ta­tions are purely ran­dom and evolu­tion sup­plies thou­sands of bits of cu­mu­la­tive se­lec­tion pres­sure. In cor­po­ra­tions, hu­mans offer up thou­sand-bit in­tel­li­gently de­signed com­plex “mu­ta­tions”, and then the fur­ther se­lec­tion pres­sure of “Did it go bankrupt or not?” ac­counts for a hand­ful of ad­di­tional bits in ex­plain­ing what you see.

Ad­vanced molec­u­lar nan­otech­nol­ogy—the ar­tifi­cial sort, not biol­ogy—should be able to copy it­self with digi­tal fidelity through thou­sands of gen­er­a­tions. Would Price’s Equa­tion thereby gain a foothold?

Cor­re­la­tion is co­var­i­ance di­vided by var­i­ance, so if A is highly pre­dic­tive of B, there can be a strong “cor­re­la­tion” be­tween them even if A is rang­ing from 0 to 9 and B is only rang­ing from 50.0001 and 50.0009. Price’s Equa­tion runs on co­var­i­ance of char­ac­ter­is­tics with re­pro­duc­tion—not cor­re­la­tion! If you can com­press var­i­ance in char­ac­ter­is­tics into a tiny band, the co­var­i­ance goes way down, and so does the cu­mu­la­tive change in the char­ac­ter­is­tic.

The Fore­sight In­sti­tute sug­gests, among other sen­si­ble pro­pos­als, that the repli­ca­tion in­struc­tions for any nan­ode­vice should be en­crypted. More­over, en­crypted such that flip­ping a sin­gle bit of the en­coded in­struc­tions will en­tirely scram­ble the de­crypted out­put. If all nan­ode­vices pro­duced are pre­cise molec­u­lar copies, and more­over, any mis­takes on the as­sem­bly line are not her­i­ta­ble be­cause the offspring got a digi­tal copy of the origi­nal en­crypted in­struc­tions for use in mak­ing grand­chil­dren, then your nan­ode­vices ain’t gonna be doin’ much evolv­ing.

You’d still have to worry about pri­ons—self-repli­cat­ing as­sem­bly er­rors apart from the en­crypted in­struc­tions, where a robot arm fails to grab a car­bon atom that is used in as­sem­bling a ho­mologue of it­self, and this causes the offspring’s robot arm to like­wise fail to grab a car­bon atom, etc., even with all the en­crypted in­struc­tions re­main­ing con­stant. But how much cor­re­la­tion is there likely to be, be­tween this sort of trans­mis­si­ble er­ror, and a higher re­pro­duc­tive rate? Let’s say that one nan­ode­vice pro­duces a copy of it­self ev­ery 1000 sec­onds, and the new nan­ode­vice is mag­i­cally more effi­cient (it not only has a prion, it has a benefi­cial prion) and copies it­self ev­ery 999.99999 sec­onds. It needs one less car­bon atom at­tached, you see. That’s not a whole lot of var­i­ance in re­pro­duc­tion, so it’s not a whole lot of co­var­i­ance ei­ther.

And how of­ten will these nan­ode­vices need to repli­cate? Un­less they’ve got more atoms available than ex­ist in the so­lar sys­tem, or for that mat­ter, the visi­ble Uni­verse, only a small num­ber of gen­er­a­tions will pass be­fore they hit the re­source wall. “Limited re­sources” are not a suffi­cient con­di­tion for evolu­tion; you need the fre­quently iter­ated death of a sub­stan­tial frac­tion of the pop­u­la­tion to free up re­sources. In­deed, “gen­er­a­tions” is not so much an in­te­ger as an in­te­gral over the frac­tion of the pop­u­la­tion that con­sists of newly cre­ated in­di­vi­d­u­als.

This is, to me, the most fright­en­ing thing about grey goo or nan­otech­nolog­i­cal weapons—that they could eat the whole Earth and then that would be it, noth­ing in­ter­est­ing would hap­pen af­ter­ward. Di­a­mond is sta­bler than pro­teins held to­gether by van der Waals forces, so the goo would only need to re­assem­ble some pieces of it­self when an as­ter­oid hit. Even if pri­ons were a pow­er­ful enough idiom to sup­port evolu­tion at all—evolu­tion is slow enough with digi­tal DNA!—less than 1.0 gen­er­a­tions might pass be­tween when the goo ate the Earth and when the Sun died.

To sum up, if you have all of the fol­low­ing prop­er­ties:

  • En­tities that replicate

  • Sub­stan­tial vari­a­tion in their characteristics

  • Sub­stan­tial vari­a­tion in their reproduction

  • Per­sis­tent cor­re­la­tion be­tween the char­ac­ter­is­tics and reproduction

  • High-fidelity long-range her­i­ta­bil­ity in characteristics

  • Fre­quent birth of a sig­nifi­cant frac­tion of the breed­ing population

  • And all this re­mains true through many iterations

Then you will have sig­nifi­cant cu­mu­la­tive se­lec­tion pres­sures, enough to pro­duce com­plex adap­ta­tions by the force of evolu­tion.