Conjuring An Evolution To Serve You

GreyThumb.blog offers an in­ter­est­ing analogue be­tween re­search on an­i­mal breed­ing and the fall of En­ron. Be­fore 1995, the way an­i­mal breed­ing worked was that you would take the top in­di­vi­d­ual perform­ers in each gen­er­a­tion and breed from them, or their par­ents. A cock­erel doesn’t lay eggs, so you have to ob­serve daugh­ter hens to de­ter­mine which cock­erels to breed. Sounds log­i­cal, right? If you take the hens who lay the most eggs in each gen­er­a­tion, and breed from them, you should get hens who lay more and more eggs.

Be­hold the awe­some power of mak­ing evolu­tion work for you! The power that made but­terflies—now con­strained to your own pur­poses! And it worked, too. Per-cow milk out­put in the US dou­bled be­tween 1905 and 1965, and has dou­bled again since then.

Yet con­jur­ing Aza­thoth oft has un­in­tended con­se­quences, as some re­searchers re­al­ized in the 1990s. In the real world, some­times you have more than an­i­mal per farm. You see the prob­lem, right? If you don’t, you should prob­a­bly think twice be­fore try­ing to con­jure an evolu­tion to serve you—magic is not for the un­para­noid.

Select­ing the hen who lays the most eggs doesn’t nec­es­sar­ily get you the most effi­cient egg-lay­ing metabolism. It may get you the most dom­i­nant hen, that pecked its way to the top of the peck­ing or­der at the ex­pense of other hens. In­di­vi­d­ual se­lec­tion doesn’t nec­es­sar­ily work to the benefit of the group, but a farm’s pro­duc­tivity is de­ter­mined by group out­puts.

In­deed, for some strange rea­son, the in­di­vi­d­ual breed­ing pro­grams which had been so suc­cess­ful at in­creas­ing egg pro­duc­tion now re­quired hens to have their beaks clipped, or be housed in in­di­vi­d­ual cages, or they would peck each other to death.

While the con­di­tions for group se­lec­tion are only rarely right in Na­ture, one can read­ily im­pose gen­uine group se­lec­tion in the lab­o­ra­tory. After only 6 gen­er­a­tions of ar­tifi­cially im­posed group se­lec­tion—breed­ing from the hens in the best groups, rather than the best in­di­vi­d­ual hens—av­er­age days of sur­vival in­creased from 160 to 348, and egg mass per bird in­creased from 5.3 to 13.3 kg. At 58 weeks of age, the se­lected line had 20% mor­tal­ity com­pared to the con­trol group at 54%. A com­mer­cial line of hens, al­lowed to grow up with un­clipped beaks, had 89% mor­tal­ity at 58 weeks.

And the fall of En­ron? Jeff Skil­ling fan­cied him­self an evolu­tion-con­jurer, it seems. (Not that he, like, knew any evolu­tion­ary math or any­thing.) Every year, ev­ery En­ron em­ployee’s perfor­mance would be eval­u­ated, and the bot­tom 10% would get fired, and the top perform­ers would get huge raises and bonuses. Un­for­tu­nately, as GreyThumb points out:

“Every­one knows that there are many things you can do in any cor­po­rate en­vi­ron­ment to give the ap­pear­ance and im­pres­sion of be­ing pro­duc­tive. En­ron’s cor­po­rate en­vi­ron­ment was par­tic­u­larly con­ducive to this: its prin­ci­pal busi­ness was en­ergy trad­ing, and it had large densely pop­u­lated trad­ing floors peo­pled by high-pow­ered traders that would sit and play the mar­kets all day. There were, I’m sure, many things that a trader could do to up his perfor­mance num­bers, ei­ther by cheat­ing or by gam­ing the sys­tem. This gam­ing of the sys­tem prob­a­bly in­cluded gam­ing his fel­low traders, many of whom were close enough to rub elbows with.

“So En­ron was ap­ply­ing se­lec­tion at the in­di­vi­d­ual level ac­cord­ing to met­rics like in­di­vi­d­ual trad­ing perfor­mance to a group sys­tem whose perfor­mance was, like the hen­houses, an emer­gent prop­erty of group dy­nam­ics as well as a re­sult of in­di­vi­d­ual fit­ness. The re­sult was more or less the same. In­stead of in­creas­ing over­all pro­duc­tivity, they got mean chick­ens and ac­tual pro­duc­tivity de­clined. They were se­lect­ing for traits like ag­gres­sive­ness, so­cio­pathic ten­den­cies, and dishon­esty.”

And the moral of the story is: Be care­ful when you set forth to con­jure the blind idiot god. Peo­ple look at a pretty but­terfly (note se­lec­tivity) and think: “Evolu­tion de­signed them—how pretty—I should get evolu­tion to do things for me, too!” But this is qual­i­ta­tive rea­son­ing, as if evolu­tion were ei­ther pre­sent or ab­sent. Ap­ply­ing 10% se­lec­tion for 10 gen­er­a­tions is not go­ing to get you the same amount of cu­mu­la­tive se­lec­tion pres­sure as 3.85 billion years of nat­u­ral se­lec­tion.

I have pre­vi­ously em­pha­sized that the evolu­tion-of-foxes works at cross-pur­poses to the evolu­tion-of-rab­bits; there is no uni­tary Evolu­tion God to praise for ev­ery beauty of Na­ture. Aza­thoth has ten mil­lion hands. When you con­jure, you don’t get the evolu­tion, the Maker of But­terflies. You get an evolu­tion, with char­ac­ter­is­tics and strength that de­pend on your ex­act con­ju­ra­tion. If you just take ev­ery­thing you see in Na­ture and at­tribute it to “evolu­tion”, you’ll start think­ing that some cute lit­tle con­ju­ra­tion which runs for 20 gen­er­a­tions will get you ar­ti­facts on the or­der of but­terflies. Try 3.85 billion years.

Same caveat with the won­ders of simu­lated evolu­tion on com­put­ers, pro­duc­ing a ra­dio an­tenna bet­ter than a hu­man de­sign, etcetera. Th­ese are some­times hu­man-com­pet­i­tive (more of­ten not) when it comes to op­ti­miz­ing a con­tin­u­ous de­sign over 57 perfor­mance crite­ria, or breed­ing a de­sign with 57 el­e­ments. Any­thing be­yond that, and mod­ern evolu­tion­ary al­gorithms are defeated by the same ex­po­nen­tial ex­plo­sion that con­sumes the rest of AI. Yes, evolu­tion­ary al­gorithms have a le­gi­t­i­mate place in AI. Con­sult a ma­chine-learn­ing ex­pert, who knows when to use them and when not to. Even biolog­i­cally in­spired ge­netic al­gorithms with sex­ual mix­ing, rarely perform bet­ter than beam searches and other non-biolog­i­cally-in­spired tech­niques on the same prob­lem.

And for this weak­ness, let us all be thank­ful. If the blind idiot god did not take a mil­lion years in which to do any­thing com­pli­cated, It would be bloody scary. 3.85 billion years of nat­u­ral se­lec­tion pro­duced molec­u­lar nan­otech­nol­ogy (cells) and Ar­tifi­cial Gen­eral In­tel­li­gence (brains), which even we hu­mans aren’t go­ing to get for a few more decades. If there were an alien demide­ity, moral­ity-and-aes­thet­ics-free, of­ten blindly suici­dal, ca­pa­ble of wield­ing nan­otech and AGI in real time, I’d put aside all other con­cerns and figure out how to kill it. As­sum­ing that I hadn’t already been en­slaved be­yond all de­sire of es­cape. Look at the trou­ble we’re hav­ing with bac­te­ria, which go through gen­er­a­tions fast enough that their evolu­tions are learn­ing to evade our an­tibiotics af­ter only a few decades’ respite.

You re­ally don’t want to con­jure Aza­thoth at full power. You re­ally, re­ally don’t. You’ll get more than pretty but­terflies.