Observing Optimization

Eliezer Yudkowsky21 Nov 2008 5:39 UTC
12 points
28 comments6 min readLW linkArchive

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.

Eliezer Yudkowsky21 Nov 2008 5:39 UTC
12 points
28 comments6 min readLW linkArchive