[ Note 2: Warning: the writing style of this post is kind of jank. It blends formal and informal register, uses punctuation loosely and italics liberally, and contains long sections of extended quotes. If that kind of thing tends to put you off, please don't read this post. Unless you're in a mood for risk. Or unless you're so into evolutionary biology, or mesa-optimizer theory, that you'd read any new hypothesis no matter how poorly written. ]
Since biology hypotheses - though when fully understood they vary wildly in their explanatory power! - are unusually hard to crisply distinguish from one another [given the high dimensionality of their study-objects], correct attribution of priority is even harder in biology than it is in pure game theory. Please inform me of any unambiguous inaccuracies; otherwise, I issue a repeat, intensified apology for the blurriness here, along with my duplicated conviction that LW having a version of the history part is worth the fog. ]
[ Note 4: I can imagine, if I try, the publication of this post being bad for capabilities. But given that all the [vague outlines of] techniques discussed are, as I understand it, already well-known to ML, if it is bad for capabilities it will probably be bad by way of emotionally inspiring somebody. So: if you find yourself emotionally inspired by this post to work on AI capabilities, try this one. If that's not to your taste, and you're still feeling inspired to go work on capabilities - because, for example, you see my argument about the capabilities part of the mesa-optimizer but you're not sure the inner and outer optimizer I describe in this post are that misaligned with each other, not really, not if you just - . . . then I wish you happiness, but I can only wish you failure. You have missed the entire point, and I publish this only in spite of you. Your posts will not age well. ]
"This subject of sexual selection has been treated at full length in the present work, simply because an opportunity was here first afforded me. I have been struck with the likeness of many of the half-favourable criticisms on sexual selection, with those which appeared at first on natural selection; such as, that it would explain some few details, but certainly was not applicable to the extent to which I have employed it. My conviction of the power of sexual selection remains unshaken; but it is probable, or almost certain, that several of my conclusions will hereafter be found erroneous; this can hardly fail to be the case in the first treatment of a subject." [Darwin, The Descent of Man, and Selection in Relation to Sex, 1871]
.
"So, the question is, if greenflies and elm trees don't do it, why do the rest of us go to such lengths to mix our genes up with somebody else's before we make a baby? It does seem an odd way to proceed. Why did sex, that bizarre perversion of straightforward replication, ever arise in the first place? What is the good of sex?
This is an extremely difficult question for the evolutionist to answer. Most serious attempts to answer it involve sophisticated mathematical reasoning. I am frankly going to evade it except to say one thing. This is that at least some of the difficulty that theorists have with explaining the evolution of sex results from the fact that they habitually think of the individual as trying to maximize the number of his genes that survive. In these terms, sex appears paradoxical because it is an 'inefficient' way for an individual to propagate her genes: each child has only 50 per cent of the individual's genes, the other 50 per cent being provided by the sexual partner. [ . . . ]
A gene 'for' sexuality manipulates all the other genes for its own selfish ends. So does a gene for [chromosomal] crossing-over. There are even genes - called mutators - that manipulate the rates of copying-errors in other genes. By definition, a copying error is to the disadvantage of the gene which is miscopied. But if it is to the advantage of the selfish mutator gene that induces it, the mutator can spread through the gene pool. Similarly, if crossing-over benefits a gene for crossing-over, that is a sufficient explanation for the existence of crossing-over. And if sexual, as opposed to non-sexual, reproduction benefits a gene for sexual reproduction, that is a sufficient explanation for the existence of sexual reproduction. Whether or not it benefits all the rest of an individual's genes is completely irrelevant. Seen from the selfish gene's point of view, sex is not so bizarre after all.
This comes perilously close to being a circular argument, since the existence of sexuality is a precondition for the whole chain of reasoning that leads to the gene being regarded as the unit of selection. I believe there are ways of escaping from the circularity, but this book is not the place to pursue the question. Sex exists. That much is true." [Dawkins, The Selfish Gene, 1976]
pt I. State and History of the Field
I. Darwin
Darwin's On the Origin of Species by Means of Natural Selection, or the Preservation of Favoured Races in the Struggle for Life, published in 1859, was the first time anyone had ever heard of a plausible mechanism other than God by which the Nature they saw in front of them, could have been arranged.
Darwin wrote of the "survival of the fittest": those individuals not fit to survive, would die out and not pass on their characteristics to the next generation, thus "naturally" changing the character of the species [within its existing variation] to be one more suited to survival. If the environment changed to be one where different characteristics were required for survival - say, your breeding population of finches moves from a large landmass with lots of different kinds of food, to an island with one major kind of food you can exploit - then the species could change too, say by the iterative dying-off, generation by generation, of those would-have-been-parents with beaks least suited to foraging the new food source, resulting in the whole breeding population having newly specialized beaks.
Darwin noted that this mechanism - «natural selection [by the deaths of the less fit for survival]» - could not explain certain superfluous or even survival-hindering characteristics, such as the bright, heavy plumage and loud [predator-attracting] calls of male birds. To account for such traits, Darwin introduced «sexual selection» as a separate, distinct mechanism:
"[ . . . ] [W]hat I have called sexual selection [ . . . ] depends, not on a struggle for existence in relation to other organic beings or to external conditions, but on a struggle between the individuals of one sex, generally the males, for the possession of the other sex. The result is not death to the unsuccessful competitor, but few or no offspring." [Darwin 1859]
Darwin identified two key pathways of sexual selection: first, by competition among males to essentially take mates by force as parcels of territory; or, alternatively, by the female's free election of suitors.
"Generally, the most vigorous males, those which are best fitted for their places in nature, will leave most progeny. But in many cases, victory will depend not on general vigour, but on having special weapons, confined to the male sex. A hornless stag or spurless cock would have a poor chance of leaving offspring. [ . . . ] [T]he war is, perhaps, severest between the males of polygamous animals, and these seem oftenest provided with special weapons. [ . . . ]
.
Amongst birds, the contest is often of a more peaceful character. All those who have attended to the subject, believe that there is the severest rivalry between the males of many species to attract by singing the females. [ . . . ] I can see no good reason to doubt that female birds, by selecting, during thousands of generations, the most melodious or beautiful males, according to their standard of beauty, might produce a marked effect." [Darwin 1859]
I think "The Simple Math of Evolution" [and certainly 99+% of contemporary academic evo-bio work] is wanting to point toward a single, dense monomechanism that is supposed to be responsible for all genetic changes that can be described as "evolutionary". Call the monomechanism «evolution by natural selection for inclusive relative reproductive fitness», or something.
Notice how Darwin himself didn't have that impulse at all. He was perfectly comfortable with a model where two distinct mechanisms - «natural selection» and «sexual selection» - are acting simultaneously, each producing a different character of effect.
[ It's tempting to think of CICOism. "Natural selection for inclusive relative reproductive fitness causes evolution." "Calories cause fat." One feels the semantic echoes of what was once a complex web of heuristics, having been compressed into a dogma of monocause, to fit more easily into people's heads. ]
II. Hardy
Hardy [1908] had already demonstrated mathematically by Fisher's time that - assuming random mating in a sexually reproducing population with particulate inheritance such that all genes or units of inheritance have matching loci between every pair of mating partners - if the frequencies of two mutually exclusive [because they are at the same particulate locus] and exhaustive, alleles in the population are as p : q, then the frequencies of the genotypes in the population must go as p2:2pq:q2 [ the proof of this being, essentially, the Punnett square ].
III. Fisher
Fisher [1930] noted that, though Darwin's theory had implicitly been one of infinitesimal gradations in heritable character, a Mendelian or particulate modification was necessary, to explain two things:
[1] the high rate of observed conserved variation within species together with the fact that,
[2] under assumptions of random mating* and a reasonably low mutation rate, heritable variation [according to Fisher] should rapidly trend to zero.
[ *The argument that Fisher actually goes with, to make the case that heritable variation should rapidly go to zero under "blending" inheritence, is based on a proof he gives later in the book about how he expects the "chance survival" of individuals generation-to-generation to influence genetic variation. Fisher assumes no affiliative [or anti-affiliative] mating, for the sake of this argument.
An alternative, just as easily workable framing, is the assumption that there will be some affiliative mating, i.e. that individuals will tend more to mate with individuals of their own genotype, which will induce an acceleration of any existing asymmetry of alleles [assuming stable population size], as the preferred breeding pool for the less common homozygotes progressively shrinks. ]
Fisher's writing illuminates how the existence of an underlying substrate of finite genetic particles, is implied by the existence of quasi-stable species. Without genetic particles, how can one explain the fact that, after a certain, seemingly hard cutoff of genetic divergence, reproduction between individuals is no longer feasible? We know species must branch because we see an archeological record that clearly implies some existing species must have branched off from a shared common ancestor. Branching implies gradualness - during speciation, there must be some period during which each proto-species has limited ability to reproduce with the other.
But in practice - as evolutionary biologists have since remarked - speciation seems to be a rare, step-change event, like supernovae, that we rarely observe in process. It was known, at least, in Fisher's time, that partial mutual fertility between breeding populations was rare, and that the norm was binary mutual-fertility-vs-mutual-infertility - defining hard cutoffs for species. Infinitesimal theories of inheritance require more "epicycles" to explain this dynamic, than discrete ones.
But even introducing Mendelian inheritance doesn't seem to me to fully explain a hard species boundary.
[ Sidenote: Fisher, in this paper, points out how close he thinks Darwin could have gotten, to independently deriving a Mendelian theory of inheritance:
"[H]e would certainly have inferred that each organism must receive a definite proportion of its genes from each parent [ . . . ] The simplification that, apart from sex [ . . . ] the contributions of the two parents were equal, would not have been confidently assumed without the evidence of reciprocal crosses; [ . . . ] our imaginary theorist would scarcely have failed to imagine a conceptual framework in which each gene had its proper place or locus, which could be occupied alternatively, had the percentage been different, by a gene of a different kind." [Fisher 1930]
Now, Fisher doubts specifically that Darwin [or a Darwin-era theorist] could have deduced without "reciprocal cross[ing]" experiments that the contributions of the two parents must be equal - i.e., that each organism must have an exemplar of each locus, more or less, to have explained Darwin's observations.
In fact, in 1871, in On The Descent of Man, and Selection in Relation to Sex [ although Fisher apparently was not aware of this ], Darwin came very close to postulating a theory of inheritance that was correct not only in being Mendelian-particulate, but also in that it correctly predicted the majority of the genome being sexually "equipotential" [ somatic, or sexually agnostic ]. [ A [largely] sexually equipotential genome, together with an assumption of loci being mirrored in each parent, logically implies both sexes must contribute equal hereditary information to each sex of offspring ]. Darwin made this conjecture on the basis of observation and theoretical reasoning, without any Mendel-type experiments at all:
"[I]n certain breeds of the fowl, spurs regularly appear in the young and healthy females. But in truth they are simply developed in the female; for in every breed each detail in the structure of the spur is transmitted through the female to her male offspring. Many cases will hereafter be given, where the female exhibits, more or less perfectly, characters proper to the male, in whom they must have been first developed, and then transferred to the female. [ . . . ] [I]n all cases of reversion, characters are transmitted through two, three, or many generations, and then are developed under certain unknown favorable conditions. [ . . . ] According to this hypothesis, every unit or cell of the body throws off gemmules or undeveloped atoms, which are transmitted to the offspring of both sexes, and are multiplied by self-division. They may remain undeveloped during the early years of life or during successive generations, and their development into units or cells, like those from which they were derived, depends on their affinity for, and union with other units or cells previously developed in the due order of growth." [Darwin 1871]
s/cells/proteins, and Darwin here is just correctly predicting how Mendelian inheritance does in fact work in sexually-reproducing species. Hence, fully Mendelian inheritance is a more straightforward implication of the Darwinian theory of sexual selection, than Darwin's heirs have ever recognized. ]
IV. Bateman
Bateman [1948] is considered to have confirmed Darwin's theory, by experimentally demonstrating that female Drosophila are indeed choosier than males - in the sense that Bateman's male Drosophila population had higher inter-individual variance in actual fertility [going by the visually 'genetically-tagged' offspring], than the females did.
"J.B.S. Haldane in 1955 briefly alluded to the principle in limited circumstances [Haldane famously joked that he would willingly die for two brothers or eight cousins], and R.A. Fisher mentioned a similar principle even more briefly in 1930. However, it was not until 1964 that W.D. Hamilton generalised the concept and developed it mathematically [ . . . ]"
Haldane was referring to how my brother, sharing as he does half my genes, is no less genetically related to me than my child [whose survival is obviously in some sense my concern according to traditional evolutionary thinking], and that this principle extends to all my genetic relatives, to an exponentially-discounted degree by distance of relation.
Armed with such intuition, Hamilton proposed an etiological explanation for the apparently selfless behavior of worker individuals in the eusocial insects. Say, by some chance circumstance, some genetically-marked caste of animals within a species - say beta wolves - becomes situated to best aid the reproduction of their own genes, not by seeking mating opportunities themselves, but by helping their reproductive [say alpha] relatives mate. Then that caste of individuals would risk reaching a tipping point, such that over successive generations, the kin-altruistic behavior would be reinforced, while breeding behavior [and capacity] would atrophy. This, in retrospect, was obviously what had happened with the eusocial worker insects.
[ A weak version - the "gay uncle" hypothesis - has since been proposed to explain human homosexuality; it doesn't feel to me like it fits empirically, but it's a valid hypothesis within evolutionary theory as far as anybody knows, and it illustrates that such hypotheses are generally viable. ]
VI. Trivers
[ This section contains a run of extended quotes from and relating to Robert Trivers. ]
In the early 1970s, Trivers basically revolutionized the whole of behavioral evolutionary biology to be about games [in the technical sense], with behavioral strategies locally optimized over successive generations. He argued that cases of heritable reciprocal altruism could be explained by cases where it was to individuals' selfish genetic advantage that they cooperate with conspecifics under certain conditions - e.g. pack hunting.
Trivers's idea of reciprocal altruism as having the capacity to evolve via phenotypes 'predicting' game-theoretic equilibria with conspecifics was somewhat different from the explanation that had been proposed in the 1960s by W.D. Hamilton for the the apparently selfless behavior of worker individuals in the eusocial insects. Hamilton's explanation was etiological; Trivers's was [although he didn't explicitly put it like this] more acausal, or logical-decision-theoretic. His focus was on the games that behavioral phenotypes could logically "see" - which, in his experience, even within tight-knit social groups, were layered and often adversarial. "The argument [in favor of reciprocal altruism] will therefore apply, unlike Hamilton's [1964], to altruistic acts exchanged between members of different species." [Trivers, "The Evolution of Reciprocal Altruism", 1971]
Trivers also sought to explain the aspects of these layered, adversarial games that culminated in less Pareto-optimal equlibria.
He pointed out that the fitness-maximizing behavioral incentives of parents did not entirely align with the fitness-maximizing behavioral incentives of their children. E.g., children - if, as Trivers did, you are modeling children as local instantiations of fitness-maximization in their capacity as children - want to monopolize their parents' capacity to care for them for as long as possible, to acquire the physical strength and possibly tutelage for more mating success, while parents want them to shuffle their resources on to more children.
This implies parent-child conflict - a war that pits the vast Parent genetic cellular automaton, against the vast Child genetic cellular automaton [the like with males vs females, etc.], over many generations.
Trivers claimed the effects of anisogamy on sex dimorphism were mediated by differential parental investment of males vs females in offspring.
Extended quotes from [and relating to] Trivers, attempting to show via cutaway view the depth of his theory:
Convergently deriving utility theory for a biological theory of the family
"In early 1971 I decided to work on what I thought of as a biological theory of the family, a theory that would have variables such as sex and age and other relevant parameters and would derive how natural selection was acting on members of the family. There was only one problem: individuals in a family were related to each other, and somehow you had to take this relatedness into account when describing natural selection acting on the participants. I puzzled about the matter for some time, in my usual style, consulted with advanced graduate students and relevant faculty, but without getting any help. What surprises me so much about this, in retrospect, was that I already knew Hamiliton's kinship theory - he had, in fact, solved the very problem that was bedeviling me, and I had by then lectured on his work. But for some reason I still narrowly conceived his work as explaining altruistic traits per se, not as covering all interactions between kin. [ . . . ]
.
The key parameter turned out to be r, or degree of relatedness, the chance that one individual shares an identical copy of any given gene with another individual by direct descent (typically, 1/2 in both directions for parent-offpsring). I narrowed the paper to mother-offspring conflict [ . . . ] because I was thinking about mammals, but then later [ . . . ] made the argument more general [ . . . ]
.
Later, when I would lecture on this work to economists, they would come up afterward and say, 'We like the way you think - just like an economist!' Some wondered whether I had learned graphing techniques by studying economics. [ . . . ] I had never had a course in economics, and the similarity [ . . . ] occurred just because of similarity in logic. Economists thought in terms of something they called 'utility', not reproductive success, but still something that could be conceptualized in terms of benefits and costs. [ . . . ]" [Trivers, Natural Selection and Social Theory, Selected Papers of Robert Trivers, 2002]
Parent-offspring conflict
At Harvard, I often watched fledgling conflict in pigeons. Both parents acted very solicitously to newly hatched chicks, stroking the neck of a chick, for example, to induce gaping and then feeding it, but toward the end of parental care, parents were harassed almost continuously by their fully grown offspring and often flew onto very narrow ledges to escape the incessant begging [ . . . ] manipulation - parents trying to make the offspring act more altruistic and less selfish than it would otherwise act on its own." [Trivers, Natural Selection and Social Theory, Selected Papers of Robert Trivers, 2002]
Coparenting as adversarial, under a natural extension of Hamilton's theory
"«Parental investment» - defined by Trivers [1972] [is] 'any investment by the parent in an individual offspring that increases the offspring's chance of surviving [and hence reproductive success] at the cost of [the] parent's ability to invest in other offspring'. An often overlooked component of Trivers's definition is that investment is not measured at the time when the parent provides care for the young, but rather in the long term, by how much it takes away from the parent's future success." [McGraw, Szekeley, & Young, "Social Behavior: Genes, Ecology and Evolution", 2010] [Note: Trivers originally took and modified the idea of measuring "parental expenditure", as Fisher had called it, from a 1930 Fisher paper speculating on the ability of parents to vary the sex ratio of their offspring.]
.
"The conditions under which selection favors male parental investment have not been specified for any group of animals. Except for the case of polygyny in birds, the role of female choice has not been explored; instead, it is commonly assumed that, whenever two individuals can raise more individuals together than one alone could, natural selection will favor male parental investment (Lack 1968, p. 149), an assumption that overlooks the effects of both male-male competition and female choice. [ . . . ] An important consequence of the early evolutionary differentiation of the sex cells and subsequent sperm competition is that male sex cells remain tiny compared to female sex cells [ . . . ] Parental investment in the young can be viewed as a sequence of discrete investments by each sex. [ . . . ] In the human species, for example, a copulation costing the male virtually nothing may trigger a nine-month investment by the female that is not trivial, followed, if she wishes, by a fifteen-year investment in the offspring that is considerable. [ . . . ]
.
[N]atural selection may favor either partner deserting even if one has invested more in the young than the other. This is because the desertion may put the deserted partner in a cruel bind: he has invested so much that he loses considerably if he also deserts the young, even though, which should make no difference to him, the partner would lose even more. [ . . . ] Two neighboring pairs of wrens happened to fledge their young simultaneously and could not tell their young apart, so both pairs fed all six young indiscriminately, until one pair 'deserted' to raise another brood, leaving their neighbors to feed all six young, which they did, even though this meant they were, in effect, being taken advantage of." [Trivers, "Parental Investment and Sexual Selection", 1972]
The evolutionary benefit of self-deceptive or repressive instincts
"I was interested in self-deception well before I became interested in evolutionary biology [ . . . ] Regarding deception, one of my most vivid childhood memories is my first realization of how pervasive and stupid patterns of human deception could be. I was about six or seven years old, as I remember it, when I had spotted [ . . . ] may have been a knife, if not some extra special toy [ . . . ] cost $6.00. [ . . . ] I The man brought me the knife, I paid, and he told me that I was $1 short. [ . . . ] If the price was really $7.00, I wanted to know, why did the sign in the window say [ . . . ] $6.00? [ . . . ] We went to the display [ . . . ] the small .98 written next to the big 6 [ . . . ] What sense did it make, I asked the man, to misrepresent the true price of an item [ . . . ] He assured me that the practice was widespread. [ . . . ] Although there are oblique references to human self-deception in my paper on reciprocal altruism, I chose the foreward to Richard Dawkins's book to first state the connection to self-deception." [Trivers, Natural Selection and Social Theory, Selected Papers of Robert Trivers, 2002]
.
"[I]f (as Dawkins argues) deceit is fundamental in animal communication, then there must be strong selection to spot deception and this ought, in turn, to select for a degree of self-deception, rendering some facts and motives unconscious so as not to betray - by the subtle signs of self-knowledge - the deception being practiced. Thus, the conventional view that natural selection favors nervous systems which produce ever more accurate images of the world must be a very naïve view of mental evolution." [Trivers, foreword to "The Selfish Gene", 1976]
"The Evolution of Psychodynamic Mechanisms", a chapter in Tooby's and Cosmides's The Adapted Mind [the major work on Lorenzian!behaviorist evolutionary psychology] begins with the above quote from Trivers, and extrapolates his logic to many areas of human and animal life.
"Confusion often results because the term 'unconscious' sometimes refers generally to anything that is outside of conscious awareness and sometimes refers to the more specific "dynamic unconscious", a special repository for mental contents that would be accessible to consciousness, except that they are actively repressed. Freud was not the first to recognize the existence of the dynamic unconscious, but he was one of the first to systematically explore and describe it (Ellenberger, 1970). Similar confusion results because "repression" describes two things: (a) the general capacity for keeping things unconscious (the meaning we will use), and (b) the more specific defense mechanism of simply "forgetting" things that are unacceptable [ . . . ]
.
offspring may be manipulated to behave in ways that are not in their best interests (for instance, by parental sanctions against sibling conflict). Slavin observes that deception (and, therefore, self-deception) is the best strategy for the otherwise powerless child. The child's wishes that are unacceptable to the parent remain conscious, while those that would be punished are pursed unconsciously. [ . . . ]
.
self-deception could increase fitness by increasing the ability to pursue selfish motives without detection. [ . . . ]
Repression makes it easier to overlook a friend's transgression. A personal slight might have been a misunderstanding instead of a defection. Even if it was a defection, it might best be ignored in order to maintain the relationship." [Barkow and Lloyd, edited by Tooby and Cosmides, The Adapted Mind, 1992]
VII. Dawkins
Dawkins reconciled these two explanations of altruism in "The Selfish Gene" [1976]: both kin-selected altruism and reciprocal altruism can be seen as special cases of the principle that what is selfishly-in-a-zero-sum-game advantageous from the perspective of a gene trying to maximize the proportion of the next generation that are its copies [the presumed fixed generation size being the game's zero-sum element], may not necessarily cash out as selfish-looking behavior under every circumstance when conspecifics meet. In the case of reciprocal altruism, the genes predict that they will encounter [any] positive-sum games with the copies of themselves in their conspecifics, and code for the corresponding self-benefiting strategy. In the case of the eusocial insects, being within non-reproducing individuals themselves, the genes of worker ants and worker bees see their advantage in maximizing the fertility of the hive - so, cooperation with conspecifics becomes the default for worker insects, since most conspecifics worker insects encounter are not, from their genes's perspective, mating competition.
II. Unsolved Problems
VIII. Why Has Sexual Reproduction Remained The Default Mode In Plants And Animals?
But biologists remain confused about why sexual reproduction remains so ubiquitous in the animal and plant kingdoms. Plants frequently, and animals occasionally, re-evolve asexual modes of reproduction, and replicate to appreciable local success [ eg Pando, whiptail lizards ]. Asexual reproduction is seemingly less costly, requiring no mate search, mating, or coordination between parents around co-parenting. So why have these local successes been dying in their cradles, leaving sexual reproduction as the global default, for hundreds of millions of years?
[ As I understand it, there's an implicit background consensus as to why, if there should be sexes at all, there should be two sexes, and it's the same reason that memetic equilibria driven by FPTP elections tend toward two political parties [which is also the same reason that, eg, everyone in a city ends up having to go to city center to visit the gas station or grocery store nearest them] - the Median Voter Theorem. Which is valid enough, as I see it. An interesting question is whether some larger trap tends us toward always having these same two sexes - exploratory-and-materially-thrifty vs immobile-and-materially-generous - or whether these particular two "political parties" just coincidentally got fixed for terran biology at the start. [I have no idea.] ]
IX. Why Are There Discrete Plant And Animal Species?
As previously noted, it's the author's vague understanding that recent evolutionary biology has veered toward viewing speciation as usually a step change event. Modern biologists mainly study speciation via the fossil record, but if you consider that we don't seem to observe many wild species that are partially mutually fertile [in terms of mating behavior rather than theoretic genetic compatibility, which is much more common - think ligers, or the theoretically possible chimp-human hybrid, or the unholy sturddlefish] conceiving of speciation as an unusual, brief aberration is just theory catching up to where our face-value observations point.
We can easily infer that speciation must occur at some point [even if we don't know how to replicate it in domestic animals, exactly]. If we observe 1,000 wild breeding populations, and 1 of them is apparently in the process of speciating, then if we assume it is valid to generalize from our sample, and if we assume our samples were taken over an identical and "average" period of time, then we can take the average over species and induce over time, and say that probably a breeding population observed continuously over 1 million years will only be in the process of speciating for on the order of 1,000 of those years.
In any case, the reality we see around us is that in nature, in terms of who actually reproduces with whom, discrete species are the overwhelming supermajority of all macroscopic animals. As for plants and microorganisms, I have no particular idea - but I guess that the situation will be similar, though meliorated, in plants [which don't need complex mating behaviors to mediate intermarriage], and much more ambiguous in microorganisms [which frequently reproduce asexually].
A knowledgeable acquaintance has informed me that bacteria, specifically, usually don't do sex, and instead do horizontal gene transfer, and that this makes bacterial species classification unwieldy.
X. Why Is Direct Conflict The Norm in Animal Life?
Trivers's implicitly game-theoretic view of evolutionary biology is obviously correct, but it contains an assumptive kernel that feels unnatural to me: ubiquitous zero-sum games. Now, "unnatural" seems like a strange word to apply to a feature of a true theory of biology, which we usually refer to as "the natural world".
But in fact Darwin would have recognized the majority of Trivers's objects of study - the conflicts between parents and children, between rival and rival, and between mate and mate - as external to the domain that the theory of «natural selection» could describe. Unless, that is, food scarcity were so extreme that members of a breeding population were frequently starving to death due to being outcompeted for food by their own conspecifics. But such circumstances are, as far as I know, rare in animals, the exceptions being infant broods of fish and arthropods.
If it doesn't have anything to do with Darwin's "survival of the fittest", then how can Trivers's "biological theory of the family" illuminate so much of animal behavior?
Trivially, Darwin's «sexual selection» could account for the direct-conflict-dense mating scene shown by Trivers's analysis. But biologists in general seem to assign inferior power to sexual selection, if they distinguish it from natural selection at all - consigning it most centrally to the modification of essentially cosmetic characteristics [which is in line with how the somatic genome is, in fact, unisex, with a very small proportion of the genome bearing sole responsibility for eliciting sexual dimorphism in a low-information, stereotyped way - such that large changes cannot occur to the genome of one sex only, because members of the opposite sex resulting from recombination of these genomes would be nonviable and leave no offspring]. This tracks Darwin's consistent reference to sexual selection as "less rigid" or "less rigorous" than natural selection, seeming to imply the standard impression that it has less power:
"The result is not death to the unsuccessful competitor, but few or no offspring. Sexual selection is, therefore, less rigorous than natural selection." [Darwin 1859]
Of course, today biologists see this "lower rigorousness" in a different light. Today we would say that the only significant fact about individual death, from the long viewpoint of evolution, is that it reduces number of offspring - rendering natural selection but a special case of sexual selection, in species that already happen to be sexually reproducing.
But the issue that sexual differences can only be skin-deep [as each gene must be at least reasonably fit to function from within both sexes], which Darwin did not position his theory to address, still remains.
Then again, Darwin notes elsewhere in Origin of Species:
"Whatever the cause may be of the variability of secondary sexual characters, as they are highly variable, sexual selection will have had a wide scope for action, and may thus have succeeded in giving to the species of the same group a greater amount of difference in these than in other aspects". [Darwin 1859]
So the sum of our priors on the efficacy of sexual selection - when we are assessing whether it is powerful enough to have produced Trivers's effects, without any group selection or anything extra and consensus-aberring like that, is:
- We know sex differences in must be superficial
- While low in information, sex differences can be large in magnitude
which we might, if we squint, allow to add up to the complex scene of mating games, instincts, and behaviors that Trivers records.
But what about parent-child conflict? Such is firmly outside the scope Darwin envisioned for the possible effects of his «sexual selection».
So our situation is:
Direct within-species conflict over survival-essential material resources does not crop up every generation to fuel an equilibrium of mutual hostility. So it can't be happening via «natural selection», or fitness-selective death.
The obvious scope of «sexual selection» is insufficient to locally explain most of Trivers's cases of omnidirectional domestic hostility.
Yet, when we're looking at animal behavior, we must admit that Trivers is right and within the family, omnidirectional hostility seems to be the default.
Why is this the case? That is, why is it the case that everyone, in Trivers's world, is [almost] always fighting?
There is, in our genes' remembered experience, always basically food for everyone to survive. We [that is, our genes] are only child, or parent, for half a lifetime. We are only male or female for one lifetime. There is seemingly not time to accrete the propensity to be deeply anything, dispositionally.
So why are animals, in family life, so deeply Machiavellian? Why is such a high fraction of animal behavior geared toward winning zero-sum games against conspecifics?
III. The Case That Sexual Selection is a Mesa-Optimizer
Dawkins: Survival machines and the conspecific
"We are survival machines - robot vehicles blindly programmed to preserve the selfish molecules known as genes."
.
"Survival machines that can simulate the future are one jump ahead of survival machines who can only learn on the basis of overt trial and error."
.
"To a survival machine, another survival machine (which is not its own child or another close relative) is part of its environment, like a rock or a river or a lump of food. It is something that gets in the way, or something that can be exploited. It differs from a rock or a river in one important respect: it is inclined to hit back."
.
"Survival machines of the same species tend to impinge on each others' lives more directly. [ . . . ] [H]alf the population of one's own species may be potential mates, and potentially hard-working and exploitable parents to one's children. [ . . . ] [M]embers of the same species, being machines for preserving genes in the same kind of place, with the same kind of way of life, are particularly direct competitors for all the resources [ . . . ] Moles and blackbirds compete with each other for worms and for everything else. If they are members of the same sex, they may also compete for mating partners. [ . . . ] The logical policy for a survival machine might therefore seem to be to murder its rivals, and then, preferably, to eat them." [Dawkins 1976]
Trivers: Behavioral equilibria analyzed from first principles
"Hamilton did something else in his famous 1964 paper that was deceptively simple - he defined the four major categories of social interaction in terms of their effects on the reproductive success of the two individuals involved. Thus, "altruistic" behavior was behavior that caused a loss to the actor and a benefit to the recipient where these were defined in terms of effects on their reproductive success. Selfish behavior was the reverse, while in cooperative behavior both parties benefited and in spiteful behavior neither party benefited: each suffered a cost. This fourfold classification of behavior, or social traits, more broadly, had the benefit of immediately stating how natural selection was acting on the interaction from the standpoint of each of the two individuals. [ . . . ] Hamilton could have called the behaviors x, y, w, and z, so as to avoid any but alphabetical connotations. [ . . . ]
.
When I came into biology at age twenty-two, never having had a course in biology and knowing next to nothing about animal behavior, my knowledge was almost entirely restricted to our own species. In adult humans it was obvious that, though kinship was a very important factor - blood being thicker than water - it could not explain all phenomena. We had strong positive feelings toward friends, and we were willing to act altruistically toward them and others. Kinship could not explain this. What could?
Well, reciprocity, in some form, could obviously do the trick - that is, you scratch my back and I'll scratch yours - but reciprocity required some thinking to get the argument right. When we are scratching each other's backs we are simultaneously trading benefits and suffering costs. That does not create much of an evolutionary problem. But what about when we act nicely toward an individual and the return benefit, if any, must come later? This raised some interesting evolutionary problems. So, I saw that what in the human species was obviously a major area of life involving deep and complex problems was not explained by Hamilton's theory, and required some new explanation. Note that the use of the term 'altruism' helped immediately in thinking about reciprocity or reciprocal altruism. Reciprocity, after all, can be negative - reciprocal spite - as Frans deWaal is fond of emphasizing." [Trivers, Natural Selection and Social Theory, Selected Papers of Robert Trivers, 2002]
Darwin: The illegibility of evolution
"and finally of sexual selection, by which characters of use to one sex are often gained and then transmitted more or less perfectly to the other sex, though of no use to the sex. But structures thus indirectly gained, although at first of no advantage to a species, may subsequently have been taken advantage of by its modified descendants, under new conditions of life and newly acquired habits.
If green woodpeckers alone had existed, and we did not know that there were many black and pied kinds, I dare say that we should have thought that the green colour was a beautiful adaptation to conceal this tree-frequenting bird from its enemies; and consequently that it was a character of importance, and had been acquired through natural selection; as it is, the colour is probably in chief part due to sexual selection." [Darwin 1871]
Dawkins: Co-adaptation and evolutionary stability
"In the case of genes [ . . . ] co-adapted gene complexes may arise in the gene pool. A large set of genes concerned with mimicry in butterflies became tightly linked together on the same chromosome, so tightly that they can be treated as one gene. [ . . . ] the more sophisticated idea of an evolutionarily stable set of genes. Mutually suitable teeth, claws, guts, and sense organs evolved in carnivore gene pools, while a different stable set of characteristics emerged from herbivore gene pools. [ . . . ]
.
I conjecture that co-adapted meme-complexes evolve in the same kind of way as co-adapted gene-complexes. Selection favours memes that exploit their cultural environment to their own advantage. This cultural environment consists of other memes which are also being selected. The meme pool therefore comes to have the attributes of an evolutionarily stable set, which new memes find it hard to invade." [Dawkins 1976]
Dawkins [repeated from top of post]: Genes for sex as potentially self-justifying
"So, the question is, if greenflies and elm trees don't do it, why do the rest of us go to such lengths to mix our genes up with somebody else's before we make a baby? It does seem an odd way to proceed. Why did sex, that bizarre perversion of straightforward replication, ever arise in the first place? What is the good of sex?
This is an extremely difficult question for the evolutionist to answer. Most serious attempts to answer it involve sophisticated mathematical reasoning. I am frankly going to evade it except to say one thing. This is that at least some of the difficulty that theorists have with explaining the evolution of sex results from the fact that they habitually think of the individual as trying to maximize the number of his genes that survive. In these terms, sex appears paradoxical because it is an 'inefficient' way for an individual to propagate her genes: each child has only 50 per cent of the individual's genes, the other 50 per cent being provided by the sexual partner. [ . . . ]
A gene 'for' sexuality manipulates all the other genes for its own selfish ends. So does a gene for [chromosomal] crossing-over. There are even genes - called mutators - that manipulate the rates of copying-errors in other genes. By definition, a copying error is to the disadvantage of the gene which is miscopied. But if it is to the advantage of the selfish mutator gene that induces it, the mutator can spread through the gene pool. Similarly, if crossing-over benefits a gene for crossing-over, that is a sufficient explanation for the existence of crossing-over. And if sexual, as opposed to non-sexual, reproduction benefits a gene for sexual reproduction, that is a sufficient explanation for the existence of sexual reproduction. Whether or not it benefits all the rest of an individual's genes is completely irrelevant. Seen from the selfish gene's point of view, sex is not so bizarre after all.
This comes perilously close to being a circular argument, since the existence of sexuality is a precondition for the whole chain of reasoning that leads to the gene being regarded as the unit of selection. I believe there are ways of escaping from the circularity, but this book is not the place to pursue the question. Sex exists. That much is true." [Dawkins 1976]
XI. Hamilton's Theory of the Ubiquity of Sex: Parasite Load
Since Dawkins's time, the most popular theory of the ubiquity of sexual reproduction has been W.D. Hamilton's: Sexual reproduction prevails because, compared to asexual reproduction, it offers a rate of adaptation that can more robustly defend against antagonistically co-evolving parasites.
But to make this hypothesis, Hamilton and Zuk implicitly assume that, while the host population is a priori variation-poor [or mutation- or adaptation-slow], the parasite population has an a priori uncapped adaptation rate, which the host population can use to bootstrap itself by iteratively selecting on resistance to parasites through female choice. I think this theory fails to hang together.
Why should female choice be better at selecting for resistance to parasites, than simple natural selection, or selective death?
More to the point, why should the a priori parasite adaptation rate be uncapped, while the a priori host adaptation rate is assumed impoverished? Yes, hosts have longer generation times and more complex organ systems. But, panspermia or no panspermia, we can assume that the most causally antecedent ancestor of all life on Earth reproduced asexually, just because asexual reproduction is simpler and it would be easier for it to happen "by chance" "in a vacuum". If parasites in general were really so fast-mutating as to impede asexual life as hard as Hamilton imagines, how did our asexual ancestors have the leisure to evolve sexuality while not getting eaten alive by parasites? Yes, the anthropic principle could explain it, but that increases the Kolmogorov complexity of Hamilton's hypothesis.
The role that "parasite load" is filling, in Hamilton's theory, is the role of "environmental pressure which is better handled by sexual reproduction, than asexual reproduction".
But as Dawkins points out when discussing a gene[-complex] for sexual reproduction that may promote sexual reproduction in the process of promoting itself, no such special environmental pressure need exist, in order for sexual reproduction to simply be more robust.
The hard question is, what's special about it?
XII. Sexual Selection Can Generate Its Own Synthetic Training Data
Dawkins may or may not have been the first to note [elsewhere in The Selfish Gene, if I'm not mistaken] that sexual selection 'eats its own tail', in the sense that Generation 2's mating pool or sexual selection environment is a direct product of mating events in Generation 1, and so on. I am still not entirely sure Darwin never mentioned this fact, but after reading and re-skimming the majority of Darwin's writing on the topic of sexual selection, I did not find any remarks upon it. This fact - the fact that the environment for sexual selection, in Generation 2, is generated directly by the outcome of sexual selection in Generation 1 - can also be seen as "sexual selection can generate its own synthetic training data".
XIII. Thought Experiment: TCG Bot Evolution
Let's construct a clunky metaphor for natural selection that also looks like an AI system, so we can more concretely see how the concept of "generating one's own synthetic training data" applies and can become important here.
[ Note: Because of the simplicity of this metaphor, it would be possible for someone to actually code [a version of] the simulation up, and run it. I don't necessarily expect the results of this, if it actually happened, to be a resounding win for my hypothesis here; my hypothesis says something about what happens under massive, evolutionary-timescale amounts of optimization pressure, under the particular constraints experienced by macroscopic organisms in our biosphere which we don't actually understand.
I'm throwing this simulation out there as a thought experiment to illustrate what the sort of thing I think happened in our universe even is, not as literal exact conditions that, on my theory, are necessarily sufficient to produce a properly isomorphic outcome.
This doesn't mean my theory can explain anything, or that I'm assuming its truth; I'm asking you, dear reader, to look at the biological animal and plant kingdoms and assess whether my theory explains what you see there better than previous theories. ]
Imagine we have two Magic decks. These will be our genomes.
[ I'm picking Magic decks because:
- the rules of Magic allow for infinitely complex homebrewed decks [most of which complexity is necessarily unexplored]
- deck content is highly variable and does a lot to constrain players' strategy, and hence determine the outcome of games
- Magic decks are legible only as deceptively simple data, and totally illegible as code for a player strategy [which they may nonetheless be]. ]
Our phenotypes will be bots that play Magic. Every day, each bot initiates [at least] 12 game-rounds of Magic. [Consider all bots as equally competent, differentiated only by the quality of their decks]. After each of these 12 game-rounds, each bot moves 1 distance unit on a spatial grid, in a direction that initializes as random, but can be coded for by its deck, based on the contents of the adjacent squares up to 2 units away, up to and including the contents of a nearby bot's deck [this is so sexually-reproducing bots can identify compatible mates].
Our grid will be populated with filler bots that have terrible decks, so that every time each player bot moves a square, it has something to play against [it meets some resistance]. Player bots play NPC bots one time. But if a player bot meets another player bot, no matter what else is happening, they must play each other at least ten times, and until either each player has lost at least once, or one of the players is dead.
We'll strengthen the resemblance to agar.io by giving each player bot a running score that is
10
+ 10x its total # victories over NPC bots
- 5% the integral of its historical score with respect to time [so if I'm currently 50 points, for my current time tick I subtract 5% * 20 * (1 tick) from my score, making my new score 19 points]
- the number of times another PC has beaten it
+ the number of times it's beaten another PC
If your score goes to 0, you die. This death mechanism constitutes «natural selection», our outer optimizer.
When it's 60 days old, each bot becomes reproductive.
Bot A is going to be asexually reproductive. Every 10 days, it's going to split off a fork of itself [mitosis], which will go on to bop around the grid and play Magic games. When the cloning happens, each card in the cloned fork is going to have x% chance of mutating y distance in a random direction - a feature added or removed, a stat or move modified, or a classifier changed.
Bot B is going to be sexually reproductive. Every 10 days, it's going to spend at most 5 days attempting to locate a compatible mate; if it manages to meet up with one, each will clone itself, shuffle and cut the clone [meiosis] with per-card mutation rate x% equal to the asexual mutation rate, merge and reshuffle half of each clone deck to produce 1 new bot, and discard the excess cards. Then its timer resets and it waits 10 days to mate again.
We can see that, at the start of the game [ as is the case for macroscopic animals ], in terms of which bot looks more promising for ultimate population, the A-type bot has 2 major advantages:
It reproduces faster, with each pair of mitoses producing twice as many offspring as each pair-of-meioses-that-results-in-a-successful-recombination.
It doesn't have to spend any energy on mating, or any energy on bearing children that is not directly invested into producing [what is approximately] an exact genetic copy of itself.
Of course, to make the B-type bots viable at the start, we'll obviously have to introduce at least 2 mutually-compatible mates, to be parents of the next generation. And we might have to introduce thousands of B-type bots, to prevent immediate population decline from inbreeding. Having made itself apparent, this genetic finnickiness might constitute a third relative advantage for Bot A:
It can expand just as easily from very small population sizes as from moderate ones.
Yet [if the reader buys the analogy so far] the analog of B-type bots, in our biosphere, seem to have won. And they seem to be winning - sexual reproduction seems convergent. Asexual reproduction in macroscopic animals keeps evolving, and could in principle take over at any time. Yet it doesn't.
What could A-type bots be worse at?
Imagine we start this simulation off, "under versimilar conditions, with sufficient number and variety of both A- and B-type bots" - whatever that would mean. We get a few false starts where either the A-type replicators or the B-type replicators or both sputter to death in fewer than a thousand generations, tweak parameters so both types are actually workable, and then hit fast-forward x 10^n.
From our god's-eye view, at first we see an explosion of A-type replication. Small enclaves of B-type bots are able to survive the explosion only because there's enough space in our grid [representing the large but not actually infinite quantity of available sunlight and geothermal energy on the Earth's surface] that the A-type replicators can almost totally dominate it, and still leave incidental pockets of space for sluggish little B-types.
Then, slowly, the B-types speciate.
While initially low-information enough that they all identified one another as compatible mates, happenstance geographical segregation means that some pairings no longer generate decks that can survive to breeding age without dying to nearby adult replicators [which here is standing in for the difficulty of embryonic development itself].
In order to promote itself at this stage, one would at first naïvely imagine a gene should code for promiscuous mating - that is, if I'm a card in a B-type deck, I should code for my bot to attempt to mate with any bot whose deck reads as more than [really this would be a complicated cost-benefit calculation taking into account my offspring's expected lifespan vs my expected lifespan], say, 50% likely capable of creating viable offspring with me. I'll waste some mating rounds on offspring that turn out to be nonviable, but the alternative - being highly discriminating - means I would waste lots of mating rounds, and spend more days per mating round, due to not being able to find any mate that meets my high standards for prospective compatibility at all.
But this assumption about the correct mate selection strategy misses one subtle feature of the meta-game.
As a mere gene made out of Magic cards, I'm obviously too dumb to understand LDT. Natural selection, however, isn't, in the sense that, on a long enough timescale, on a large enough playing field, where the relevant play-strategies are simple enough for selection to act on them, strategies that violate LDT will be weeded out until only LDT ones are left.
So even though a gene can't really know a priori that it should choose strategies that help-rather-than-hurt the replication of sets of copies of itself, it can end up acting as though it knows this, via all its competitor genes which didn't code for LDT strategies, getting weeded out of the population.
[ Note that this is not Wynne-Edwards - Lorenzian «groupselection». Wynne-Edwards - Lorenzian groupselection predicts that individual, realized animals should act to altruistically sacrifice their own replication for the survival of the group, which doesnot happen. Groupselectionism fails to be valid because it assumes that selection cares about the survival of realized [groups of] animals. Really selection only cares about the reproductive fitness of [groups of] genes.
But it can care about the reproductive fitness of groups of genes, distributed or not, so long as - see Dawkins - it can act on them as a unit. ]
So we have some distributed sets of copies of cards that code for promiscuous mating, and some distributed sets of copies of cards that code for selective mating. Of the two kinds of distributed sets-of-copies, which tend - under our current dynamic - to be better at replicating?
It's not a trivial question to answer. The trick, for sensibly predicting what in fact happens in the very long run here, is to look at the very long run from the start. That's the thing about optimizers, even very locally dumb optimizers like natural selection: they act like they have more information than you've actually seen them collecting, and more predictive power than whatever printouts of their internal calculations you've seen can justify.
Say I'm a [set-of-copies of] card[s] deciding whether to code for promiscuous or selective mating, in this simulation-game. From the viewpoint of natural selection - our outer optimizer - I do get to "decide" based on a view of the situation 1,000,000 generations out. At least, if I'm at the far right tail of the fitness distribution, I get to look in retrospect like I was making the decision on that much foreknowledge, because of the amount of optimization pressure natural selection will have been capable of grinding me through in the meantime.
Both gene-copy-sets [because they have equal mutation rate and equal generation time] are optimized at an equal rate by natural selection - that is, the pressure to not die against NPC bots. Also, both gene-copy-sets, a priori, before any selection is done, will be equally vulnerable to other players. The only major difference seems to be that the A-type bots reproduce [more than] twice as quickly.
But distributed sets-of-gene-copies which code for mating sufficiently promiscuous that it does not result in discrete species, are never acted upon by sexual selection. Their Magic decks can't be honed by the thousands of rounds of the intricate, implicit self-play that is constituted by the mating endeavor [for both sexes] when they can be sure their offspring will be viable, and can instead expend effort scrutinizing, wooing, and holding their [genetically very familiar] mate to ensure that their offspring will be optimal.
And the 1,000,000-[or 1,000,000,000-]generation view, says that distributed sets-of-gene-copies, which get sexual selection and can thus generate their own synthetic training data at all, win.
The selectively-mating gene copy-sets, because they get sexual selection, are smarter at adapting to grind the NPC bots than the promiscuously-mating gene copy-sets. They're also smarter at adapting to predate on the promiscuously-mating gene copy-sets, than the promiscuously-mating gene copy-sets are at adapting to predate on them. These two things alone set up the selective-mating genes to vastly outnumber the promiscuous-mating genes.
Are there other in-principle training-speedup mesa-strategies which natural selection missed, in our timeline? I don't know, but I think it's evident that it found this one.
IV. Sexual Selection Analyzed as a Mesa-Optimizer: Frames and Implications
"Angels on the sideline
Puzzled and amused
Why did Father give these humans free will?
Now they're all confused
Don't these talking monkeys know that
Eden has enough to go around?
Plenty in this holy garden, silly monkeys
Where there's one, you're bound to divide it
Right in two" [TOOL, "Right in Two"]
Relation To Within-Species Hostility
Our three questions from Section II were:
- Why has sexual selection remained the default mode in plants and animals?
- Why are there discrete plant and animal species?
- Why is direct conflict the norm in animal life?
The one I haven't even attempted to answer yet is the third.
"Relative Fitness"
For individuals belonging to no discretely bounded mutually-fertile population, "relative reproductive fitness" is not available as a selection criterion. This is, of course, just another way of saying that if you don't have a hard boundary as to which members of your local animal community belong to your breeding population, sexual selection is not available.
But the connection to the oft-pointed-to sexual-optimization target of "relative reproductive fitness" is worth examining.
Biologists often make a [valid] argument that animal behaviors which decrease the coherence of the group - finding ways to "cheat" by philandering, cuckolding, slacking/mooching, betrayal, etc. - are explicable by the priority that the animal's instincts [/genes] are assigning to maximizing its relative share of the next generation - as opposed to the group's absolute next-generation size.
To some extent, we can account for the fact that so many of the most complex animals' instincts are geared toward deep, complex zero-sum games, because
- when our selection environment [i.e., complex conspecifics] gets very complex, natural selection must look very hard into the future [i.e. optimize very hard], so sexual selection is employed
- sexual selection works by self-play
- self-play is zero-sum
.
So when sexual selection is in play, and you ask for more complexity, you will [modulo everything] get more "free-riding" zero-sum-game-oriented behaviors. They're not truly free-riding, in the sense that they're a logical extension of the inner-optimizing heuristic speed-adaptation mechanism. But we still feel they are free-riding in some sense.
Why This Is An Example Of Mesa-Optimization
The above might read as self-defeating, in a way. If the direction of natural selection is to create animals that are fit for their environment, in what sense can animals that waste most of their time on aggressing and undermining each other be "best", according to natural selection's view?
But it's only a paradox if you see «natural selection» as still being in control.
Hubinger et al.'s original definition of "mesa-optimizer" was
"the type of learned optimization that occurs when a learned model (such as a neural network) is itself an optimizer"
.
I've added an additional criterion that mesa- optimizers [as opposed to meta- optimizers] should be removed from their base or outer optimizers in a direction that looks to us like alienation, rather than actualization.
"'Meta' is for when you go outside the system. 'Meta-reasoning'. 'Meta' refers to things that are strictly more powerful than the object-level versions of themselves. 'Meta-cult'. 'Meta-dispute'. 'Meta-currency'.
'Meta' seemed to modify object-level referents to make them more all-inclusive - allowing us to modify an object-level word such as 'cat', to predicate over a sort of ideal or Platonic fusion referent that counterfactually-exists in our preferred, as opposed to the real, social consensus. I'd rather join a meta-cult than a cult, I'd rather win a meta-dispute than a dispute, and I'd rather have 10 meta-dollars than 10 dollars. And I think I'd rather pet a meta-cat, than a cat. [ . . . ]
.
"'Meta-'X is more desirable to me.
'Mesa-'X is more desirable to you - and you are not me, so to me it looks like you're dragging my group's precious social consensus, selfishly closer to yourself. [Which is fine if you're aligned enough with me! In fact, these concepts aren't exactly diametrically opposed - if our world is composed of identical agents, meta-X and mesa-X will be exactly the same for all X.]"
The category boundary of "optimizer", among what Hubinger et al. refer to as "learned models", is ill-defined at present.
In any case, mesa-optimizer theory is broadly oriented toward finding ways to protect against what we might call usurping mesa-optimizers. We can tell these mesa-optimizers are at least optimizers relative to us, because they're quite capable of optimizing over us - that is, seizing control of their creators.
It's widely accepted that human general intelligence constitutes a usurping mesa-optimizer relative to evolution. Even though evolution still has effects on the world, for thousands of years, humans have been optimizing the world around themselves, far harder than evolution is capable of optimizing the world around itself. We use birth control and spend most of our time doing things that don't advance our reproduction at all; instead, we do things we care about.
«Natural selection», in selecting for organisms that meet its fitness criteria of "survive and reproduce the most", accidentally ran into a paperclip-maximizer, and lost out accordingly. «Sexual selection» doesn't care at all who survives and reproduces the most; it only cares who has the highest proportion of relatedness in the breeding pool, over the next N generations. We might not think that's a very sensible or aesthetic goal, relative to natural selection's more humble and peaceable desires. Natural selection definitely wouldn't, if it was smart enough to answer. Nonetheless, for animals on Earth, sexual selection has been the hardest-optimizing force for hundreds of millions of years, as the work of Trivers, Dawkins, deWaal, [ironically] Lorenz, and others can attest.
Well, it was. Again, until humans.
Corrections to the theory will be received with gratitude and joy.
[ Note 1: This post is a follow-up to "Thoughts on Evo-Bio Math and Mesa-Optimization: Maybe We Need To Think Harder About 'Relative' Fitness?" ]
[ Note 2: Warning: the writing style of this post is kind of jank. It blends formal and informal register, uses punctuation loosely and italics liberally, and contains long sections of extended quotes. If that kind of thing tends to put you off, please don't read this post. Unless you're in a mood for risk. Or unless you're so into evolutionary biology, or mesa-optimizer theory, that you'd read any new hypothesis no matter how poorly written. ]
[ Note 3: The earlier parts of this post are somewhat an entry into my series of woefully incomplete histories.
Since biology hypotheses - though when fully understood they vary wildly in their explanatory power! - are unusually hard to crisply distinguish from one another [given the high dimensionality of their study-objects], correct attribution of priority is even harder in biology than it is in pure game theory. Please inform me of any unambiguous inaccuracies; otherwise, I issue a repeat, intensified apology for the blurriness here, along with my duplicated conviction that LW having a version of the history part is worth the fog. ]
[ Note 4: I can imagine, if I try, the publication of this post being bad for capabilities. But given that all the [vague outlines of] techniques discussed are, as I understand it, already well-known to ML, if it is bad for capabilities it will probably be bad by way of emotionally inspiring somebody. So: if you find yourself emotionally inspired by this post to work on AI capabilities, try this one. If that's not to your taste, and you're still feeling inspired to go work on capabilities - because, for example, you see my argument about the capabilities part of the mesa-optimizer but you're not sure the inner and outer optimizer I describe in this post are that misaligned with each other, not really, not if you just - . . . then I wish you happiness, but I can only wish you failure. You have missed the entire point, and I publish this only in spite of you. Your posts will not age well. ]
.
pt I. State and History of the Field
I. Darwin
Darwin's On the Origin of Species by Means of Natural Selection, or the Preservation of Favoured Races in the Struggle for Life, published in 1859, was the first time anyone had ever heard of a plausible mechanism other than God by which the Nature they saw in front of them, could have been arranged.
Darwin wrote of the "survival of the fittest": those individuals not fit to survive, would die out and not pass on their characteristics to the next generation, thus "naturally" changing the character of the species [within its existing variation] to be one more suited to survival. If the environment changed to be one where different characteristics were required for survival - say, your breeding population of finches moves from a large landmass with lots of different kinds of food, to an island with one major kind of food you can exploit - then the species could change too, say by the iterative dying-off, generation by generation, of those would-have-been-parents with beaks least suited to foraging the new food source, resulting in the whole breeding population having newly specialized beaks.
Darwin noted that this mechanism - «natural selection [by the deaths of the less fit for survival]» - could not explain certain superfluous or even survival-hindering characteristics, such as the bright, heavy plumage and loud [predator-attracting] calls of male birds. To account for such traits, Darwin introduced «sexual selection» as a separate, distinct mechanism:
Darwin identified two key pathways of sexual selection: first, by competition among males to essentially take mates by force as parcels of territory; or, alternatively, by the female's free election of suitors.
.
I think "The Simple Math of Evolution" [and certainly 99+% of contemporary academic evo-bio work] is wanting to point toward a single, dense monomechanism that is supposed to be responsible for all genetic changes that can be described as "evolutionary". Call the monomechanism «evolution by natural selection for inclusive relative reproductive fitness», or something.
Notice how Darwin himself didn't have that impulse at all. He was perfectly comfortable with a model where two distinct mechanisms - «natural selection» and «sexual selection» - are acting simultaneously, each producing a different character of effect.
[ It's tempting to think of CICOism. "Natural selection for inclusive relative reproductive fitness causes evolution." "Calories cause fat." One feels the semantic echoes of what was once a complex web of heuristics, having been compressed into a dogma of monocause, to fit more easily into people's heads. ]
II. Hardy
Hardy [1908] had already demonstrated mathematically by Fisher's time that - assuming random mating in a sexually reproducing population with particulate inheritance such that all genes or units of inheritance have matching loci between every pair of mating partners - if the frequencies of two mutually exclusive [because they are at the same particulate locus] and exhaustive, alleles in the population are as p : q, then the frequencies of the genotypes in the population must go as p2:2pq:q2 [ the proof of this being, essentially, the Punnett square ].
III. Fisher
Fisher [1930] noted that, though Darwin's theory had implicitly been one of infinitesimal gradations in heritable character, a Mendelian or particulate modification was necessary, to explain two things:
[1] the high rate of observed conserved variation within species together with the fact that,
[2] under assumptions of random mating* and a reasonably low mutation rate, heritable variation [according to Fisher] should rapidly trend to zero.
[ *The argument that Fisher actually goes with, to make the case that heritable variation should rapidly go to zero under "blending" inheritence, is based on a proof he gives later in the book about how he expects the "chance survival" of individuals generation-to-generation to influence genetic variation. Fisher assumes no affiliative [or anti-affiliative] mating, for the sake of this argument.
An alternative, just as easily workable framing, is the assumption that there will be some affiliative mating, i.e. that individuals will tend more to mate with individuals of their own genotype, which will induce an acceleration of any existing asymmetry of alleles [assuming stable population size], as the preferred breeding pool for the less common homozygotes progressively shrinks. ]
Fisher's writing illuminates how the existence of an underlying substrate of finite genetic particles, is implied by the existence of quasi-stable species. Without genetic particles, how can one explain the fact that, after a certain, seemingly hard cutoff of genetic divergence, reproduction between individuals is no longer feasible? We know species must branch because we see an archeological record that clearly implies some existing species must have branched off from a shared common ancestor. Branching implies gradualness - during speciation, there must be some period during which each proto-species has limited ability to reproduce with the other.
But in practice - as evolutionary biologists have since remarked - speciation seems to be a rare, step-change event, like supernovae, that we rarely observe in process. It was known, at least, in Fisher's time, that partial mutual fertility between breeding populations was rare, and that the norm was binary mutual-fertility-vs-mutual-infertility - defining hard cutoffs for species. Infinitesimal theories of inheritance require more "epicycles" to explain this dynamic, than discrete ones.
But even introducing Mendelian inheritance doesn't seem to me to fully explain a hard species boundary.
[ Sidenote: Fisher, in this paper, points out how close he thinks Darwin could have gotten, to independently deriving a Mendelian theory of inheritance:
Now, Fisher doubts specifically that Darwin [or a Darwin-era theorist] could have deduced without "reciprocal cross[ing]" experiments that the contributions of the two parents must be equal - i.e., that each organism must have an exemplar of each locus, more or less, to have explained Darwin's observations.
In fact, in 1871, in On The Descent of Man, and Selection in Relation to Sex [ although Fisher apparently was not aware of this ], Darwin came very close to postulating a theory of inheritance that was correct not only in being Mendelian-particulate, but also in that it correctly predicted the majority of the genome being sexually "equipotential" [ somatic, or sexually agnostic ]. [ A [largely] sexually equipotential genome, together with an assumption of loci being mirrored in each parent, logically implies both sexes must contribute equal hereditary information to each sex of offspring ]. Darwin made this conjecture on the basis of observation and theoretical reasoning, without any Mendel-type experiments at all:
s/cells/proteins, and Darwin here is just correctly predicting how Mendelian inheritance does in fact work in sexually-reproducing species. Hence, fully Mendelian inheritance is a more straightforward implication of the Darwinian theory of sexual selection, than Darwin's heirs have ever recognized. ]IV. Bateman
Bateman [1948] is considered to have confirmed Darwin's theory, by experimentally demonstrating that female Drosophila are indeed choosier than males - in the sense that Bateman's male Drosophila population had higher inter-individual variance in actual fertility [going by the visually 'genetically-tagged' offspring], than the females did.
V. Hamilton
Quoting Wikipedia:
Haldane was referring to how my brother, sharing as he does half my genes, is no less genetically related to me than my child [whose survival is obviously in some sense my concern according to traditional evolutionary thinking], and that this principle extends to all my genetic relatives, to an exponentially-discounted degree by distance of relation.
Armed with such intuition, Hamilton proposed an etiological explanation for the apparently selfless behavior of worker individuals in the eusocial insects. Say, by some chance circumstance, some genetically-marked caste of animals within a species - say beta wolves - becomes situated to best aid the reproduction of their own genes, not by seeking mating opportunities themselves, but by helping their reproductive [say alpha] relatives mate. Then that caste of individuals would risk reaching a tipping point, such that over successive generations, the kin-altruistic behavior would be reinforced, while breeding behavior [and capacity] would atrophy. This, in retrospect, was obviously what had happened with the eusocial worker insects.
[ A weak version - the "gay uncle" hypothesis - has since been proposed to explain human homosexuality; it doesn't feel to me like it fits empirically, but it's a valid hypothesis within evolutionary theory as far as anybody knows, and it illustrates that such hypotheses are generally viable. ]
VI. Trivers
[ This section contains a run of extended quotes from and relating to Robert Trivers. ]
In the early 1970s, Trivers basically revolutionized the whole of behavioral evolutionary biology to be about games [in the technical sense], with behavioral strategies locally optimized over successive generations. He argued that cases of heritable reciprocal altruism could be explained by cases where it was to individuals' selfish genetic advantage that they cooperate with conspecifics under certain conditions - e.g. pack hunting.
Trivers's idea of reciprocal altruism as having the capacity to evolve via phenotypes 'predicting' game-theoretic equilibria with conspecifics was somewhat different from the explanation that had been proposed in the 1960s by W.D. Hamilton for the the apparently selfless behavior of worker individuals in the eusocial insects. Hamilton's explanation was etiological; Trivers's was [although he didn't explicitly put it like this] more acausal, or logical-decision-theoretic. His focus was on the games that behavioral phenotypes could logically "see" - which, in his experience, even within tight-knit social groups, were layered and often adversarial. "The argument [in favor of reciprocal altruism] will therefore apply, unlike Hamilton's [1964], to altruistic acts exchanged between members of different species." [Trivers, "The Evolution of Reciprocal Altruism", 1971]
Trivers also sought to explain the aspects of these layered, adversarial games that culminated in less Pareto-optimal equlibria.
He pointed out that the fitness-maximizing behavioral incentives of parents did not entirely align with the fitness-maximizing behavioral incentives of their children. E.g., children - if, as Trivers did, you are modeling children as local instantiations of fitness-maximization in their capacity as children - want to monopolize their parents' capacity to care for them for as long as possible, to acquire the physical strength and possibly tutelage for more mating success, while parents want them to shuffle their resources on to more children.
This implies parent-child conflict - a war that pits the vast Parent genetic cellular automaton, against the vast Child genetic cellular automaton [the like with males vs females, etc.], over many generations.
Trivers claimed the effects of anisogamy on sex dimorphism were mediated by differential parental investment of males vs females in offspring.
Extended quotes from [and relating to] Trivers, attempting to show via cutaway view the depth of his theory:
Convergently deriving utility theory for a biological theory of the family
.
.
Parent-offspring conflict
Coparenting as adversarial, under a natural extension of Hamilton's theory
.
.
The evolutionary benefit of self-deceptive or repressive instincts
.
"The Evolution of Psychodynamic Mechanisms", a chapter in Tooby's and Cosmides's The Adapted Mind [the major work on Lorenzian!behaviorist evolutionary psychology] begins with the above quote from Trivers, and extrapolates his logic to many areas of human and animal life.
.
.
VII. Dawkins
Dawkins reconciled these two explanations of altruism in "The Selfish Gene" [1976]: both kin-selected altruism and reciprocal altruism can be seen as special cases of the principle that what is selfishly-in-a-zero-sum-game advantageous from the perspective of a gene trying to maximize the proportion of the next generation that are its copies [the presumed fixed generation size being the game's zero-sum element], may not necessarily cash out as selfish-looking behavior under every circumstance when conspecifics meet. In the case of reciprocal altruism, the genes predict that they will encounter [any] positive-sum games with the copies of themselves in their conspecifics, and code for the corresponding self-benefiting strategy. In the case of the eusocial insects, being within non-reproducing individuals themselves, the genes of worker ants and worker bees see their advantage in maximizing the fertility of the hive - so, cooperation with conspecifics becomes the default for worker insects, since most conspecifics worker insects encounter are not, from their genes's perspective, mating competition.
II. Unsolved Problems
VIII. Why Has Sexual Reproduction Remained The Default Mode In Plants And Animals?
But biologists remain confused about why sexual reproduction remains so ubiquitous in the animal and plant kingdoms. Plants frequently, and animals occasionally, re-evolve asexual modes of reproduction, and replicate to appreciable local success [ eg Pando, whiptail lizards ]. Asexual reproduction is seemingly less costly, requiring no mate search, mating, or coordination between parents around co-parenting. So why have these local successes been dying in their cradles, leaving sexual reproduction as the global default, for hundreds of millions of years?
[ As I understand it, there's an implicit background consensus as to why, if there should be sexes at all, there should be two sexes, and it's the same reason that memetic equilibria driven by FPTP elections tend toward two political parties [which is also the same reason that, eg, everyone in a city ends up having to go to city center to visit the gas station or grocery store nearest them] - the Median Voter Theorem. Which is valid enough, as I see it. An interesting question is whether some larger trap tends us toward always having these same two sexes - exploratory-and-materially-thrifty vs immobile-and-materially-generous - or whether these particular two "political parties" just coincidentally got fixed for terran biology at the start. [I have no idea.] ]
IX. Why Are There Discrete Plant And Animal Species?
As previously noted, it's the author's vague understanding that recent evolutionary biology has veered toward viewing speciation as usually a step change event. Modern biologists mainly study speciation via the fossil record, but if you consider that we don't seem to observe many wild species that are partially mutually fertile [in terms of mating behavior rather than theoretic genetic compatibility, which is much more common - think ligers, or the theoretically possible chimp-human hybrid, or the unholy sturddlefish] conceiving of speciation as an unusual, brief aberration is just theory catching up to where our face-value observations point.
We can easily infer that speciation must occur at some point [even if we don't know how to replicate it in domestic animals, exactly]. If we observe 1,000 wild breeding populations, and 1 of them is apparently in the process of speciating, then if we assume it is valid to generalize from our sample, and if we assume our samples were taken over an identical and "average" period of time, then we can take the average over species and induce over time, and say that probably a breeding population observed continuously over 1 million years will only be in the process of speciating for on the order of 1,000 of those years.
In any case, the reality we see around us is that in nature, in terms of who actually reproduces with whom, discrete species are the overwhelming supermajority of all macroscopic animals. As for plants and microorganisms, I have no particular idea - but I guess that the situation will be similar, though meliorated, in plants [which don't need complex mating behaviors to mediate intermarriage], and much more ambiguous in microorganisms [which frequently reproduce asexually].
A knowledgeable acquaintance has informed me that bacteria, specifically, usually don't do sex, and instead do horizontal gene transfer, and that this makes bacterial species classification unwieldy.
X. Why Is Direct Conflict The Norm in Animal Life?
Trivers's implicitly game-theoretic view of evolutionary biology is obviously correct, but it contains an assumptive kernel that feels unnatural to me: ubiquitous zero-sum games. Now, "unnatural" seems like a strange word to apply to a feature of a true theory of biology, which we usually refer to as "the natural world".
But in fact Darwin would have recognized the majority of Trivers's objects of study - the conflicts between parents and children, between rival and rival, and between mate and mate - as external to the domain that the theory of «natural selection» could describe. Unless, that is, food scarcity were so extreme that members of a breeding population were frequently starving to death due to being outcompeted for food by their own conspecifics. But such circumstances are, as far as I know, rare in animals, the exceptions being infant broods of fish and arthropods.
If it doesn't have anything to do with Darwin's "survival of the fittest", then how can Trivers's "biological theory of the family" illuminate so much of animal behavior?
Trivially, Darwin's «sexual selection» could account for the direct-conflict-dense mating scene shown by Trivers's analysis. But biologists in general seem to assign inferior power to sexual selection, if they distinguish it from natural selection at all - consigning it most centrally to the modification of essentially cosmetic characteristics [which is in line with how the somatic genome is, in fact, unisex, with a very small proportion of the genome bearing sole responsibility for eliciting sexual dimorphism in a low-information, stereotyped way - such that large changes cannot occur to the genome of one sex only, because members of the opposite sex resulting from recombination of these genomes would be nonviable and leave no offspring]. This tracks Darwin's consistent reference to sexual selection as "less rigid" or "less rigorous" than natural selection, seeming to imply the standard impression that it has less power:
Of course, today biologists see this "lower rigorousness" in a different light. Today we would say that the only significant fact about individual death, from the long viewpoint of evolution, is that it reduces number of offspring - rendering natural selection but a special case of sexual selection, in species that already happen to be sexually reproducing.
But the issue that sexual differences can only be skin-deep [as each gene must be at least reasonably fit to function from within both sexes], which Darwin did not position his theory to address, still remains.
Then again, Darwin notes elsewhere in Origin of Species:
So the sum of our priors on the efficacy of sexual selection - when we are assessing whether it is powerful enough to have produced Trivers's effects, without any group selection or anything extra and consensus-aberring like that, is:
- We know sex differences in must be superficial
- While low in information, sex differences can be large in magnitude
which we might, if we squint, allow to add up to the complex scene of mating games, instincts, and behaviors that Trivers records.
But what about parent-child conflict? Such is firmly outside the scope Darwin envisioned for the possible effects of his «sexual selection».
So our situation is:
Direct within-species conflict over survival-essential material resources does not crop up every generation to fuel an equilibrium of mutual hostility. So it can't be happening via «natural selection», or fitness-selective death.
The obvious scope of «sexual selection» is insufficient to locally explain most of Trivers's cases of omnidirectional domestic hostility.
Yet, when we're looking at animal behavior, we must admit that Trivers is right and within the family, omnidirectional hostility seems to be the default.
Why is this the case? That is, why is it the case that everyone, in Trivers's world, is [almost] always fighting?
There is, in our genes' remembered experience, always basically food for everyone to survive. We [that is, our genes] are only child, or parent, for half a lifetime. We are only male or female for one lifetime. There is seemingly not time to accrete the propensity to be deeply anything, dispositionally.
So why are animals, in family life, so deeply Machiavellian? Why is such a high fraction of animal behavior geared toward winning zero-sum games against conspecifics?
III. The Case That Sexual Selection is a Mesa-Optimizer
Dawkins: Survival machines and the conspecific
.
.
.
Trivers: Behavioral equilibria analyzed from first principles
.
Darwin: The illegibility of evolution
Dawkins: Co-adaptation and evolutionary stability
.
Dawkins [repeated from top of post]: Genes for sex as potentially self-justifying
XI. Hamilton's Theory of the Ubiquity of Sex: Parasite Load
Since Dawkins's time, the most popular theory of the ubiquity of sexual reproduction has been W.D. Hamilton's: Sexual reproduction prevails because, compared to asexual reproduction, it offers a rate of adaptation that can more robustly defend against antagonistically co-evolving parasites.
But to make this hypothesis, Hamilton and Zuk implicitly assume that, while the host population is a priori variation-poor [or mutation- or adaptation-slow], the parasite population has an a priori uncapped adaptation rate, which the host population can use to bootstrap itself by iteratively selecting on resistance to parasites through female choice. I think this theory fails to hang together.
The role that "parasite load" is filling, in Hamilton's theory, is the role of "environmental pressure which is better handled by sexual reproduction, than asexual reproduction".
But as Dawkins points out when discussing a gene[-complex] for sexual reproduction that may promote sexual reproduction in the process of promoting itself, no such special environmental pressure need exist, in order for sexual reproduction to simply be more robust.
The hard question is, what's special about it?
XII. Sexual Selection Can Generate Its Own Synthetic Training Data
Dawkins may or may not have been the first to note [elsewhere in The Selfish Gene, if I'm not mistaken] that sexual selection 'eats its own tail', in the sense that Generation 2's mating pool or sexual selection environment is a direct product of mating events in Generation 1, and so on. I am still not entirely sure Darwin never mentioned this fact, but after reading and re-skimming the majority of Darwin's writing on the topic of sexual selection, I did not find any remarks upon it. This fact - the fact that the environment for sexual selection, in Generation 2, is generated directly by the outcome of sexual selection in Generation 1 - can also be seen as "sexual selection can generate its own synthetic training data".
XIII. Thought Experiment: TCG Bot Evolution
Let's construct a clunky metaphor for natural selection that also looks like an AI system, so we can more concretely see how the concept of "generating one's own synthetic training data" applies and can become important here.
[ Note: Because of the simplicity of this metaphor, it would be possible for someone to actually code [a version of] the simulation up, and run it. I don't necessarily expect the results of this, if it actually happened, to be a resounding win for my hypothesis here; my hypothesis says something about what happens under massive, evolutionary-timescale amounts of optimization pressure, under the particular constraints experienced by macroscopic organisms in our biosphere which we don't actually understand.
I'm throwing this simulation out there as a thought experiment to illustrate what the sort of thing I think happened in our universe even is, not as literal exact conditions that, on my theory, are necessarily sufficient to produce a properly isomorphic outcome.
This doesn't mean my theory can explain anything, or that I'm assuming its truth; I'm asking you, dear reader, to look at the biological animal and plant kingdoms and assess whether my theory explains what you see there better than previous theories. ]
Imagine we have two Magic decks. These will be our genomes.
[ I'm picking Magic decks because:
- the rules of Magic allow for infinitely complex homebrewed decks [most of which complexity is necessarily unexplored]
- deck content is highly variable and does a lot to constrain players' strategy, and hence determine the outcome of games
- Magic decks are legible only as deceptively simple data, and totally illegible as code for a player strategy [which they may nonetheless be]. ]
Our phenotypes will be bots that play Magic. Every day, each bot initiates [at least] 12 game-rounds of Magic. [Consider all bots as equally competent, differentiated only by the quality of their decks]. After each of these 12 game-rounds, each bot moves 1 distance unit on a spatial grid, in a direction that initializes as random, but can be coded for by its deck, based on the contents of the adjacent squares up to 2 units away, up to and including the contents of a nearby bot's deck [this is so sexually-reproducing bots can identify compatible mates].
Our grid will be populated with filler bots that have terrible decks, so that every time each player bot moves a square, it has something to play against [it meets some resistance]. Player bots play NPC bots one time. But if a player bot meets another player bot, no matter what else is happening, they must play each other at least ten times, and until either each player has lost at least once, or one of the players is dead.
We'll strengthen the resemblance to agar.io by giving each player bot a running score that is
10
+ 10x its total # victories over NPC bots
- 5% the integral of its historical score with respect to time [so if I'm currently 50 points, for my current time tick I subtract 5% * 20 * (1 tick) from my score, making my new score 19 points]
- the number of times another PC has beaten it
+ the number of times it's beaten another PC
If your score goes to 0, you die. This death mechanism constitutes «natural selection», our outer optimizer.
When it's 60 days old, each bot becomes reproductive.
Bot A is going to be asexually reproductive. Every 10 days, it's going to split off a fork of itself [mitosis], which will go on to bop around the grid and play Magic games. When the cloning happens, each card in the cloned fork is going to have x% chance of mutating y distance in a random direction - a feature added or removed, a stat or move modified, or a classifier changed.
Bot B is going to be sexually reproductive. Every 10 days, it's going to spend at most 5 days attempting to locate a compatible mate; if it manages to meet up with one, each will clone itself, shuffle and cut the clone [meiosis] with per-card mutation rate x% equal to the asexual mutation rate, merge and reshuffle half of each clone deck to produce 1 new bot, and discard the excess cards. Then its timer resets and it waits 10 days to mate again.
We can see that, at the start of the game [ as is the case for macroscopic animals ], in terms of which bot looks more promising for ultimate population, the A-type bot has 2 major advantages:
Of course, to make the B-type bots viable at the start, we'll obviously have to introduce at least 2 mutually-compatible mates, to be parents of the next generation. And we might have to introduce thousands of B-type bots, to prevent immediate population decline from inbreeding. Having made itself apparent, this genetic finnickiness might constitute a third relative advantage for Bot A:
Yet [if the reader buys the analogy so far] the analog of B-type bots, in our biosphere, seem to have won. And they seem to be winning - sexual reproduction seems convergent. Asexual reproduction in macroscopic animals keeps evolving, and could in principle take over at any time. Yet it doesn't.
What could A-type bots be worse at?
Imagine we start this simulation off, "under versimilar conditions, with sufficient number and variety of both A- and B-type bots" - whatever that would mean. We get a few false starts where either the A-type replicators or the B-type replicators or both sputter to death in fewer than a thousand generations, tweak parameters so both types are actually workable, and then hit fast-forward x 10^n.
From our god's-eye view, at first we see an explosion of A-type replication. Small enclaves of B-type bots are able to survive the explosion only because there's enough space in our grid [representing the large but not actually infinite quantity of available sunlight and geothermal energy on the Earth's surface] that the A-type replicators can almost totally dominate it, and still leave incidental pockets of space for sluggish little B-types.
Then, slowly, the B-types speciate.
While initially low-information enough that they all identified one another as compatible mates, happenstance geographical segregation means that some pairings no longer generate decks that can survive to breeding age without dying to nearby adult replicators [which here is standing in for the difficulty of embryonic development itself].
In order to promote itself at this stage, one would at first naïvely imagine a gene should code for promiscuous mating - that is, if I'm a card in a B-type deck, I should code for my bot to attempt to mate with any bot whose deck reads as more than [really this would be a complicated cost-benefit calculation taking into account my offspring's expected lifespan vs my expected lifespan], say, 50% likely capable of creating viable offspring with me. I'll waste some mating rounds on offspring that turn out to be nonviable, but the alternative - being highly discriminating - means I would waste lots of mating rounds, and spend more days per mating round, due to not being able to find any mate that meets my high standards for prospective compatibility at all.
But this assumption about the correct mate selection strategy misses one subtle feature of the meta-game.
As a mere gene made out of Magic cards, I'm obviously too dumb to understand LDT. Natural selection, however, isn't, in the sense that, on a long enough timescale, on a large enough playing field, where the relevant play-strategies are simple enough for selection to act on them, strategies that violate LDT will be weeded out until only LDT ones are left.
So even though a gene can't really know a priori that it should choose strategies that help-rather-than-hurt the replication of sets of copies of itself, it can end up acting as though it knows this, via all its competitor genes which didn't code for LDT strategies, getting weeded out of the population.
[ Note that this is not Wynne-Edwards - Lorenzian «groupselection». Wynne-Edwards - Lorenzian groupselection predicts that individual, realized animals should act to altruistically sacrifice their own replication for the survival of the group, which does not happen. Groupselectionism fails to be valid because it assumes that selection cares about the survival of realized [groups of] animals. Really selection only cares about the reproductive fitness of [groups of] genes.
But it can care about the reproductive fitness of groups of genes, distributed or not, so long as - see Dawkins - it can act on them as a unit. ]
So we have some distributed sets of copies of cards that code for promiscuous mating, and some distributed sets of copies of cards that code for selective mating. Of the two kinds of distributed sets-of-copies, which tend - under our current dynamic - to be better at replicating?
It's not a trivial question to answer. The trick, for sensibly predicting what in fact happens in the very long run here, is to look at the very long run from the start. That's the thing about optimizers, even very locally dumb optimizers like natural selection: they act like they have more information than you've actually seen them collecting, and more predictive power than whatever printouts of their internal calculations you've seen can justify.
Say I'm a [set-of-copies of] card[s] deciding whether to code for promiscuous or selective mating, in this simulation-game. From the viewpoint of natural selection - our outer optimizer - I do get to "decide" based on a view of the situation 1,000,000 generations out. At least, if I'm at the far right tail of the fitness distribution, I get to look in retrospect like I was making the decision on that much foreknowledge, because of the amount of optimization pressure natural selection will have been capable of grinding me through in the meantime.
Both gene-copy-sets [because they have equal mutation rate and equal generation time] are optimized at an equal rate by natural selection - that is, the pressure to not die against NPC bots. Also, both gene-copy-sets, a priori, before any selection is done, will be equally vulnerable to other players. The only major difference seems to be that the A-type bots reproduce [more than] twice as quickly.
But distributed sets-of-gene-copies which code for mating sufficiently promiscuous that it does not result in discrete species, are never acted upon by sexual selection. Their Magic decks can't be honed by the thousands of rounds of the intricate, implicit self-play that is constituted by the mating endeavor [for both sexes] when they can be sure their offspring will be viable, and can instead expend effort scrutinizing, wooing, and holding their [genetically very familiar] mate to ensure that their offspring will be optimal.
And the 1,000,000-[or 1,000,000,000-]generation view, says that distributed sets-of-gene-copies, which get sexual selection and can thus generate their own synthetic training data at all, win.
The selectively-mating gene copy-sets, because they get sexual selection, are smarter at adapting to grind the NPC bots than the promiscuously-mating gene copy-sets. They're also smarter at adapting to predate on the promiscuously-mating gene copy-sets, than the promiscuously-mating gene copy-sets are at adapting to predate on them. These two things alone set up the selective-mating genes to vastly outnumber the promiscuous-mating genes.
Are there other in-principle training-speedup mesa-strategies which natural selection missed, in our timeline? I don't know, but I think it's evident that it found this one.
IV. Sexual Selection Analyzed as a Mesa-Optimizer: Frames and Implications
Relation To Within-Species Hostility
Our three questions from Section II were:
- Why has sexual selection remained the default mode in plants and animals?
- Why are there discrete plant and animal species?
- Why is direct conflict the norm in animal life?
The one I haven't even attempted to answer yet is the third.
"Relative Fitness"
For individuals belonging to no discretely bounded mutually-fertile population, "relative reproductive fitness" is not available as a selection criterion. This is, of course, just another way of saying that if you don't have a hard boundary as to which members of your local animal community belong to your breeding population, sexual selection is not available.
But the connection to the oft-pointed-to sexual-optimization target of "relative reproductive fitness" is worth examining.
Biologists often make a [valid] argument that animal behaviors which decrease the coherence of the group - finding ways to "cheat" by philandering, cuckolding, slacking/mooching, betrayal, etc. - are explicable by the priority that the animal's instincts [/genes] are assigning to maximizing its relative share of the next generation - as opposed to the group's absolute next-generation size.
To some extent, we can account for the fact that so many of the most complex animals' instincts are geared toward deep, complex zero-sum games, because
- when our selection environment [i.e., complex conspecifics] gets very complex, natural selection must look very hard into the future [i.e. optimize very hard], so sexual selection is employed
- sexual selection works by self-play
- self-play is zero-sum
.
So when sexual selection is in play, and you ask for more complexity, you will [modulo everything] get more "free-riding" zero-sum-game-oriented behaviors. They're not truly free-riding, in the sense that they're a logical extension of the inner-optimizing heuristic speed-adaptation mechanism. But we still feel they are free-riding in some sense.
Why This Is An Example Of Mesa-Optimization
The above might read as self-defeating, in a way. If the direction of natural selection is to create animals that are fit for their environment, in what sense can animals that waste most of their time on aggressing and undermining each other be "best", according to natural selection's view?
But it's only a paradox if you see «natural selection» as still being in control.
Hubinger et al.'s original definition of "mesa-optimizer" was
.
I've added an additional criterion that mesa- optimizers [as opposed to meta- optimizers] should be removed from their base or outer optimizers in a direction that looks to us like alienation, rather than actualization.
.
The category boundary of "optimizer", among what Hubinger et al. refer to as "learned models", is ill-defined at present.
In any case, mesa-optimizer theory is broadly oriented toward finding ways to protect against what we might call usurping mesa-optimizers. We can tell these mesa-optimizers are at least optimizers relative to us, because they're quite capable of optimizing over us - that is, seizing control of their creators.
It's widely accepted that human general intelligence constitutes a usurping mesa-optimizer relative to evolution. Even though evolution still has effects on the world, for thousands of years, humans have been optimizing the world around themselves, far harder than evolution is capable of optimizing the world around itself. We use birth control and spend most of our time doing things that don't advance our reproduction at all; instead, we do things we care about.
«Natural selection», in selecting for organisms that meet its fitness criteria of "survive and reproduce the most", accidentally ran into a paperclip-maximizer, and lost out accordingly. «Sexual selection» doesn't care at all who survives and reproduces the most; it only cares who has the highest proportion of relatedness in the breeding pool, over the next N generations. We might not think that's a very sensible or aesthetic goal, relative to natural selection's more humble and peaceable desires. Natural selection definitely wouldn't, if it was smart enough to answer. Nonetheless, for animals on Earth, sexual selection has been the hardest-optimizing force for hundreds of millions of years, as the work of Trivers, Dawkins, deWaal, [ironically] Lorenz, and others can attest.
Well, it was. Again, until humans.
Corrections to the theory will be received with gratitude and joy.