Sorry; one last. In parallel to my earlier objection to the word, let me argue here for an endorsement of it that I think does live alongside (and that is already very widely and conventionally recognized).
I can think of a sense in which “design” has a primary meaning that isn’t about intention, and that is central to Nick’s and EricC’s points. In my preferred language of function, we can often, in some big system, identify some functions that are more primitive or “atomic” or “elementary” (by some criterion, relatively speaking), and others that are more composite or compound. When the primitive functions have lots of possible combinations, and we see a particular one realized, its particularity can become much higher than the particularity of the elements. In very large spaces, the quantitative difference can be so large that it seems to be (perhaps is?) qualitative. Or at least it wants other concepts to analyze and try to understand it. Where “design” has a role in the discourse that the more generic term “function” doesn’t have, my sense of its ordinary usage is to signal this axis from the more elementary to the more composite, and to remark on the increasing particularity as the prior possibility-space of compositions gets large. Navigating this combinatorial expanse is what human designers do, and so we associate it with them in common usage. From this standpoint, intention is a means to the end of choosing among many counterfactuals, if other human abilities have enabled us to first generate and then scan them. So there becomes a pertinent question: how large a particularity in an apparently combinatorial space can selection (of whatever variety) plausibly be suggested to have produced? This is probably more to the center of Nick’s concern in the source email for this sub-thread. Now we’re in the domain of several things that people have done for a long time, and had good thoughts about: 1. “credit propagation” to refine components of composite functions, as John Holland tried to study it in his Classifier Systems; 2. Herb Simon’s efforts to see where combinatorial spaces are not as unstructured or neutral as they first appear to us, so that selection is indeed less implausible than we naively suppose; 3. Any of my ongoing interests in paths of least resistance, which are just outgrowths of the previous two. All of these are not meant to show that “hard” (in the sense of NP at scale) problems of selection are solved; they are meant to dissipate the perception of those problems as hard. Valiant’s arguments are that they must be dissipated, or we have mis-characterized them. That, of course, is a Bayesian prior. If Valiant is wrong, that will be interesting to learn. When I say that the difference of large combinatorial spaces can be qualitative, at the same time as I argue that their crypticity is probably not as high as we think it is, I mean that we will still need new pattern-concepts to deal with the kinds of organization that look cryptic but that afford navigation paths for suitably-configured (matched) searchers. As everybody here also already recognizes, the successes of ML are going to show that recursive reinforcement protocols can identify a lot more patterns than we come to expect before now. I also want to add to the earlier remark something that I forgot but that I think is important: There is a big asymmetry between the distillation power of iterating those reinforcement programs in the forward direction, and carrying out statistical inference about what they have found and how they have generated internal “representations” of it, in the backward direction. This asymmetry is part, though not all, of why selection can extract quite complicated functions, but population genetics can still be a very partial tool for pinning down mechanistic models. As ML systems learn to discover affordable abstractions (as in the examples of programming that Marcus sends), I wonder how much our view will change, of adding syntactic layers to quantity-based pattern representation. Okay; I think with this I will not have written something so biased that I don’t actually believe it, Eric > On Mar 31, 2026, at 7:31, Santafe <[email protected]> wrote: > > So the second reply to NST: > >> On Mar 30, 2026, at 19:34, Nicholas Thompson <[email protected]> wrote: >> >> DES et al., >> >> I don't understand how we can have a conversation about evolution without >> engaging the problem of design. Darwin's theory was an attempt to explain >> design without reference to a designer; natural selection was that >> explanation. Unless I'm out of my mind—which of course is a >> possibility—population biology does not address that problem at all. > > This is a change of subject from the earlier protest email that I just > replied to, about what is or isn’t a tautology, but as long as we acknowledge > that, I am happy to shift with it. > > I don’t find myself ever using the term “design”, and — although I did > criticize it off-hand yesterday — I had to reflect a moment about why. I > suspect that, if I spent a lot of time thinking about it, I would decide I > really can’t stand the term (for this context), and would more-actively > refuse to use it, protected by an automatic pain-reflex. Specifically, the > pain of cognitive Bad Faith, as Glen uses the term and as I would use it too. > The reason is that it is the worst of all worlds: its overwhelming semantic > content is prejudice: the metaphorical imposition of human intention, while > the aspect of it that could ever be unpacked into something systematic is so > weak and arbitrary that it is a good model for horoscope language. It > _invites_ conversations like Pangloss’s, in which whatever happens is glossed > ex post to be the Best of All Possible Worlds. Again, remember the Vampires: > don’t complain to me that they do things you dislike, if you are the one who > invited them in. > > (Now, here, I do want to acknowledge the earlier sub-thread that JonZ has > immortalized: if we were teleported back into the God-suffused world that > Darwin inhabited, maybe we would be granted no other terms for discourse, and > would then have to suffer whatever fate befell us. But let me not suppose > that world here, because for today we haven’t yet been pushed back down into > it.) > > There is very much a word I would like to use probably-everywhere you want > the vampire-word of “design”, and if we were in physics or engineering, I > could use it that way. It is the word “function”. It is unfortunate that, > in evolution-leaning biology (and thus bled into the rest of biology), the > word “function” was captured early-on by the theologians, and imprinted with > almost all the same prejudicial metaphor as “design”. And it turns out that > no matter how long you tell biologists that you want to use the term in its > physics/engineering meaning, they will simply refuse to hear you, continue to > hear you as meaning their usage, and then complain that whatever you said > doesn’t parse. I don’t know what to do about that. So I am going to use it > in the physics/engineering sense here, and I guess damn the torpedos. (Is > that what they say on ships? "I guess damn the torpedos"?) > > If I had to speak carefully about what “function” carries in physics or > engineering, I would say it entails that we have agreed upon a > system/environment decomposition, with the substitutability that multiple > different systems could be embedded in the same environment, glued at > (more-or-less, closely-enough) the same interfaces. A “function” performed > by the system is then some kind of change delivered through the > system/environment interface, which can be defined from what is done at the > interface or in the environment, independently of the specific identification > of the system doing it. (There is a math of conditional independence, yada > yada yada; you know it and the list always discusses it; Markov Blankets and > other stuff I don’t even know the definitions of; let me not go off down that > track here.) > > (For example, obviously not meant to be inclusive: “The chemical conversion X > -> Y is a function that can be performed by many different reaction > networks.") > > My aim (at whatever level of semantic tedium is needed to carry it through) > is that sentences such as “different systems can perform the same function” > are meaningful and sensible. I don’t regard this as any real problem, beyond > ad hoc work setting up terms for description and discourse. > > Anyway — god it’s all tiresome — when we are talking about “functions”, there > are all sorts of questions of criterion that are _easy_ to make operational > and concrete. Performing a chemical conversion. Absorbing light energy to > produce a charge separation. Making convex poly-glycine bodies that focus > light as converging lenses. Making electrochemically-active membrane systems > that induce waves from receptors for charge separations. Realizing a kind of > system-catalytic control architecture that inputs polynucleotides as > templates, and generates polypeptides as products, thus admitting lots of > control-theoretic and information-theoretic markings that we formalize in the > Central Dogma. And so forth. > > Each of these functions is a describable capability in the physical world. > We can then ask other sensible (if often rather difficult) questions about: > How rare are systems with that capability? How “hard” are they to discover > (making use of computer-science formalisms to try to capture “hardness”, and > requiring us to declare null models to define “discovery” as an absorption > phenomenon of stochastic motion in combinatorial spaces). And on and on. > > Then, a whole world of questions opens up to us in an entirely non-confusing > form, as far as I can tell: Given two premises: 1) things fall apart (which > again we can make operational at the cost of tedium); and 2) in > temporally-extended processes, amplification and attenuation (the latter both > from merely falling apart, and from more active processes) become available > in certain kinds of systems (which must have memory, yada yada yada…) will > this-or-that particular model of a heredity-and-selection dynamic give a > defensible account of the discovery of some system performing the function in > question? Will it do the more stringent job of accounting for the presence > of _this observed_ system among the alternatives that perform the same > function? And so forth. > >> Mathematics often has a kind of predictive power: working out the >> implications of a theory in mathematical form can lead you to places you did >> not anticipate. In my geriatric bewilderment, this is what I think >> population biology does. I still don't think it has much to say about >> design, nor has there been much attempt to explore the distribution of >> design in nature. On the whole, there is not as much of it as people have >> been led to think. But sometimes natural selection seems to do a very good >> job, as shown by comparisons between natural objects and engineered ones. > > Is it in Job that one finds “Why dost thou kick against the pricks?” If you > want God-or-man intentional metaphors, and you want to rail against > God-or-man metaphors, I can’t help you. These are your own demons. I can > try to make the case that there are worlds not inhabited by these demons, > which are not deficient thereby. > > You absolutely are right that population genetics is a weak domain for causal > inference. This is the kind of thing that Mike Lynch’s papers like to simply > elide, while he goes to beat up somebody else he has decided he doesn’t like > (or more likely in my view, has just decided he has the perch to dominate). > There are interesting things to say about why population-genetic > model-selection is a weak domain for causal inference, having to do with > sample power in relation to question classes, etc. Lots can be said, and if > one has time to engage it, I think a lot of it is interesting. The way > Machine Learning models work is going to give us a lot of lived experience to > make the strengths and limitations of these filtering-dynamic processes, and > the statistical inference over them, familiar and ultimately intuitive. > > I would claim that nobody reasonable is claiming that popgen model > identification is all we have to define or identify causation, or that it > would be sufficient if it were all we had. We do all the rest of science > too, and various parts pitch in various insights. (Yes, _that_ was a > tautology. Anybody who did claim it, I would call unreasonable. Obviously > the non-tautological insinuation is that the “reasonable” set contains most > of the biologists.) This was my point, in the previous post, that popgen > doesn’t cook your dinner and do the dishes afterward. Those tasks get done > too; they just get done through other agencies. > >> There has been a great deal of work along the lines of: “Oh wow, this thing >> in nature is designed just as well as something we might have designed >> ourselves. Holy cow, isn't that wonderful? Isn't natural selection grand?” >> But sometimes these effects are not produced by natural selection—or at >> least not entirely. > > If we decide we want to pursue these questions seriously, beyond just a kind > of idle emotion-stroking, there is plenty here that can become interesting > too: _Why_do we find some human-design project harder or easier, more > noteworthy or not? Look up the work on what makes combinatorial games > interesting to play, as cultures have evolved them over centuries. The > Mathematical Games crew at Berkeley MSRI wrote lovely stuff on this for many > years. When is a design problem for us, as future-and-counterfactual > model-makers and model-scanners, similar in kind to a discovery problem > accomplished in some non-human system? When are we recapitulating the > non-human system because certain forms of the problem are most-accessible? > When is discovery for us meaningfully-enough different from that available to > a non-human, or a non-cognitive, process, to motivate characterizing the > difference? > > There is _so_ much to do. > > Eric > > .- .-.. .-.. / ..-. --- --- - . .-. ... / .- .-. . / .-- .-. --- -. --. / ... --- -- . / .- .-. . / ..- ... . ..-. ..- .-.. FRIAM Applied Complexity Group listserv Fridays 9a-12p Friday St. Johns Cafe / Thursdays 9a-12p Zoom https://bit.ly/virtualfriam to (un)subscribe http://redfish.com/mailman/listinfo/friam_redfish.com FRIAM-COMIC http://friam-comic.blogspot.com/ archives: 5/2017 thru present https://redfish.com/pipermail/friam_redfish.com/ 1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
