Observing Optimization

Fol­lowup to: Op­ti­miza­tion and the Singularity

In “Op­ti­miza­tion and the Sin­gu­lar­ity” I pointed out that his­tory since the first repli­ca­tor, in­clud­ing hu­man his­tory to date, has mostly been a case of non­re­cur­sive op­ti­miza­tion—where you’ve got one thingy do­ing the op­ti­miz­ing, and an­other thingy get­ting op­ti­mized. When evolu­tion builds a bet­ter amoeba, that doesn’t change the struc­ture of evolu­tion - the mu­tate-re­pro­duce-se­lect cy­cle.

But there are ex­cep­tions to this rule, such as the in­ven­tion of sex, which af­fected the struc­ture of nat­u­ral se­lec­tion it­self—trans­form­ing it to mu­tate-re­com­bine-mate-re­pro­duce-se­lect.

I was sur­prised when Robin, in “Eliezer’s Meta-Level Deter­minism” took that idea and ran with it and said:

...his view does seem to make testable pre­dic­tions about his­tory. It sug­gests the in­tro­duc­tion of nat­u­ral se­lec­tion and of hu­man cul­ture co­in­cided with the very largest ca­pa­bil­ity growth rate in­creases. It sug­gests that the next largest in­creases were much smaller and co­in­cided in biol­ogy with the in­tro­duc­tion of cells and sex, and in hu­mans with the in­tro­duc­tion of writ­ing and sci­ence. And it sug­gests other rate in­creases were sub­stan­tially smaller.

It hadn’t oc­curred to me to try to de­rive that kind of testable pre­dic­tion. Why? Well, par­tially be­cause I’m not an economist. (Don’t get me wrong, it was a vir­tu­ous step to try.) But also be­cause the whole is­sue looked to me like it was a lot more com­pli­cated than that, so it hadn’t oc­curred to me to try to di­rectly ex­tract pre­dic­tions.

What is this “ca­pa­bil­ity growth rate” of which you speak, Robin? There are old, old con­tro­ver­sies in evolu­tion­ary biol­ogy in­volved here.

Just to start by point­ing out the ob­vi­ous—if there are fixed re­sources available, only so much grass to be eaten or so many rab­bits to con­sume, then any evolu­tion­ary “progress” that we would rec­og­nize as pro­duc­ing a bet­ter-de­signed or­ganism, may just re­sult in the dis­place­ment of the old allele by the new allele—not any in­crease in the pop­u­la­tion as a whole. It’s quite pos­si­ble to have a new wolf that ex­pends 10% more en­ergy per day to be 20% bet­ter at hunt­ing, and in this case the sus­tain­able wolf pop­u­la­tion will de­crease as new wolves re­place old.

If I was go­ing to talk about the effect that a meta-level change might have on the “op­ti­miza­tion ve­loc­ity” of nat­u­ral se­lec­tion, I would talk about the time for a new adap­ta­tion to re­place an old adap­ta­tion af­ter a shift in se­lec­tion pres­sures—not the to­tal pop­u­la­tion or to­tal bio­mass or to­tal mor­pholog­i­cal com­plex­ity (see be­low).

Like­wise in hu­man his­tory—farm­ing was an im­por­tant in­no­va­tion for pur­poses of op­ti­miza­tion, not be­cause it changed the hu­man brain all that much, but be­cause it meant that there were a hun­dred times as many brains around; and even more im­por­tantly, that there were sur­pluses that could sup­port spe­cial­ized pro­fes­sions. But many in­no­va­tions in hu­man his­tory may have con­sisted of new, im­proved, more harm­ful weapons—which would, if any­thing, have de­creased the sus­tain­able pop­u­la­tion size (though “no effect” is more likely—fewer peo­ple means more food means more peo­ple).

Or similarly: there’s a talk some­where where ei­ther War­ren Buffett or Charles Munger men­tions how they hate to hear about tech­nolog­i­cal im­prove­ments in cer­tain in­dus­tries—be­cause even if in­vest­ing a few mil­lion can cut the cost of pro­duc­tion by 30% or what­ever, the bar­ri­ers to com­pe­ti­tion are so low that the con­sumer cap­tures all the gain. So they have to in­vest to keep up with com­peti­tors, and the in­vestor doesn’t get much re­turn.

I’m try­ing to mea­sure the op­ti­miza­tion ve­loc­ity of in­for­ma­tion, not pro­duc­tion or growth rates. At the tail end of a very long pro­cess, knowl­edge fi­nally does trans­late into power—guns or nan­otech­nol­ogy or what­ever. But along that long way, if you’re mea­sur­ing the num­ber of ma­te­rial copies of the same stuff (how many wolves, how many peo­ple, how much grain), you may not be get­ting much of a glimpse at op­ti­miza­tion ve­loc­ity. Too many com­pli­ca­tions along the causal chain.

And this is not just my prob­lem.

Back in the bad old days of pre-1960s evolu­tion­ary biol­ogy, it was widely taken for granted that there was such a thing as progress, that it pro­ceeded for­ward over time, and that mod­ern hu­man be­ings were at the apex.

Ge­orge Willi­ams’s Adap­ta­tion and Nat­u­ral Selec­tion, mark­ing the so-called “Willi­ams Revolu­tion” in ev-bio that flushed out a lot of the ro­man­ti­cism and an­thro­po­mor­phism, spent most of one chap­ter ques­tion­ing the seem­ingly com­mon-sen­si­cal met­rics of “progress”.

Biol­o­gists some­times spoke of “mor­pholog­i­cal com­plex­ity” in­creas­ing over time. But how do you mea­sure that, ex­actly? And at what point in life do you mea­sure it if the or­ganism goes through mul­ti­ple stages? Is an am­phibian more ad­vanced than a mam­mal, since its genome has to store the in­for­ma­tion for mul­ti­ple stages of life?

“There are life cy­cles enor­mously more com­plex than that of a frog,” Willi­ams wrote. “The lowly and ‘sim­ple’ liver fluke...” goes through stages that in­clude a wa­ter­borne stage that swims us­ing cilia; finds and bur­rows into a snail and then trans­forms into a sporo­cyst; that re­pro­duces by bud­ding to pro­duce re­dia; that mi­grate in the snail and re­pro­duce asex­u­ally; then trans­form into cer­caria, that, by wig­gling a tail, bur­rows out of the snail and swims to a blade of grass; where they trans­form into dor­mant metac­er­caria; that are eaten by sheep and then hatch into a young fluke in­side the sheep; then trans­form into adult flukes; which spawn fluke zy­gotes… So how “ad­vanced” is that?

Willi­ams also pointed out that there would be a limit to how much in­for­ma­tion evolu­tion could main­tain in the genome against de­gen­er­a­tive pres­sures—which seems like a good prin­ci­ple in prac­tice, though I made some mis­takes on OB in try­ing to de­scribe the the­ory. Tax­onomists of­ten take a cur­rent form and call the his­tor­i­cal trend to­ward it “progress”, but is that up­ward mo­tion, or just sub­sti­tu­tion of some adap­ta­tions for other adap­ta­tions in re­sponse to chang­ing se­lec­tion pres­sures?

“To­day the fish­ery biol­o­gists greatly fear such ar­chaic fishes as the bowfin, garpikes , and lam­prey, be­cause they are such out­stand­ingly effec­tive com­peti­tors,” Willi­ams noted.

So if I were talk­ing about the effect of e.g. sex as a meta-level in­no­va­tion, then I would ex­pect e.g. an in­crease in the to­tal bio­chem­i­cal and mor­pholog­i­cal com­plex­ity that could be main­tained—the lift­ing of a pre­vi­ous up­per bound, fol­lowed by an ac­cre­tion of in­for­ma­tion. And I might ex­pect a change in the ve­loc­ity of new adap­ta­tions re­plac­ing old adap­ta­tions.

But to get from there, to some­thing that shows up in the fos­sil record—that’s not a triv­ial step.

I re­call read­ing, some­where or other, about an ev-bio con­tro­versy that en­sued when one party spoke of the “sud­den burst of cre­ativity” rep­re­sented by the Cam­brian ex­plo­sion, and won­dered why evolu­tion was pro­ceed­ing so much more slowly nowa­days. And an­other party re­sponded that the Cam­brian differ­en­ti­a­tion was mainly visi­ble post hoc—that the groups of an­i­mals we have now, first differ­en­ti­ated from one an­other then, but that at the time the differ­ences were not as large as they loom nowa­days. That is, the ac­tual ve­loc­ity of adap­ta­tional change wasn’t re­mark­able by com­par­i­son to mod­ern times, and only hind­sight causes us to see those changes as “stak­ing out” the an­ces­try of the ma­jor an­i­mal groups.

I’d be sur­prised to learn that sex had no effect on the ve­loc­ity of evolu­tion. It looks like it should in­crease the speed and num­ber of sub­sti­tuted adap­ta­tions, and also in­crease the com­plex­ity bound on the to­tal ge­netic in­for­ma­tion that can be main­tained against mu­ta­tion. But to go from there, to just look­ing at the fos­sil record and see­ing faster progress—it’s not just me who thinks that this jump to phe­nomenol­ogy is ten­ta­tive, difficult, and con­tro­ver­sial.

Should you ex­pect more spe­ci­a­tion af­ter the in­ven­tion of sex, or less? The first im­pulse is to say “more”, be­cause sex seems like it should in­crease the op­ti­miza­tion ve­loc­ity and speed up time. But sex also cre­ates mu­tu­ally re­pro­duc­ing pop­u­la­tions, that share genes among them­selves, as op­posed to asex­ual lineages—so might that act as a cen­tripetal force?

I don’t even pro­pose to an­swer this ques­tion, just point out that it is ac­tu­ally quite stan­dard for the phe­nomenol­ogy of evolu­tion­ary the­o­ries—the ques­tion of which ob­serv­ables are pre­dicted—to be a ma­jor difficulty. Un­less you’re deal­ing with re­ally easy qual­i­ta­tive ques­tions like “Should I find rab­bit fos­sils in the pre-Cam­brian?” (I try to only make pre­dic­tions about AI, us­ing my the­ory of op­ti­miza­tion, when it looks like an easy ques­tion.)

Yes, it’s more con­ve­nient for sci­en­tists when the­o­ries make eas­ily testable, read­ily ob­serv­able pre­dic­tions. But when I look back at the his­tory of life, and the his­tory of hu­man­ity, my first pri­or­ity is to ask “What’s go­ing on here?”, and only af­ter­ward see if I can man­age to make non-ob­vi­ous retro­d­ic­tions. I can’t just start with the goal of hav­ing a con­ve­nient phe­nomenol­ogy. Or similarly: the the­o­ries I use to or­ga­nize my un­der­stand­ing of the his­tory of op­ti­miza­tion to date, have lots of pa­ram­e­ters, e.g. the op­ti­miza­tion-effi­ciency curve that de­scribes op­ti­miza­tion out­put as a func­tion of re­source in­put, or the ques­tion of how many low-hang­ing fruit ex­ist in the neigh­bor­hood of a given search point. Does a larger pop­u­la­tion of wolves in­crease the ve­loc­ity of nat­u­ral se­lec­tion, by cov­er­ing more of the search neigh­bor­hood for pos­si­ble mu­ta­tions? If so, is that a log­a­r­ith­mic in­crease with pop­u­la­tion size, or what? - But I can’t just wish my the­o­ries into be­ing sim­pler.

If Robin has a sim­pler causal model, with fewer pa­ram­e­ters, that stands di­rectly be­hind ob­serv­ables and eas­ily coughs up testable pre­dic­tions, which fits the data well, and ob­vi­ates the need for my own ab­strac­tions like “op­ti­miza­tion effi­ciency” -

- then I may have to dis­card my own at­tempts at the­o­riz­ing. But ob­serv­ing a se­ries of ma­te­rial growth modes doesn’t con­tra­dict a causal model of op­ti­miza­tion be­hind the scenes, be­cause it’s a pure phe­nomenol­ogy, not it­self a causal model—it doesn’t say whether a given in­no­va­tion had any effect on the op­ti­miza­tion ve­loc­ity of the pro­cess that pro­duced fu­ture ob­ject-level in­no­va­tions that ac­tu­ally changed growth modes, etcetera.