Cybernetics is (spoiler alert) “the scientific study of control and communication in the animal and the machine.” Simple enough, no? But if the average person is asked what “cybernetics” is, probably what they’re picturing is something like the stock images of ‘computer hacker’:
Likewise, German existentialist Martin Heidegger declared philosophy to be over and cybernetics its wholesale replacement. So there’s a way in which the term accreted heavy mystique, pretty quickly. To this day, it’s kind of amazing to think that Wiener’s book 1948 monograph Cybernetics was a phenomenon (5 printings in the first year!), given that a page can look like this:
So it would be easy to tell a story about a highly-technical expert discourse that unexpectedly fired popular imagination and spawned a bevy of vulgar interpretations - ‘deconstructed’ tiramisu or any of the hundred flavors of ‘quantum’ nonsense you can find out there.
But if that story were true, then “what is cybernetics” would just be a matter of clearing away mistaken assumptions and my job would be about as exciting as grading a mid freshman comp essay.
The real story is more like: “cybernetics was an attempt to solve technical questions and ended up raising a lot of philosophical questions.” This, unsurprisingly, means that it tends to get cleaved in half along the divide of what CP Snow called the Two Cultures.1 If you want the history of cybernetics, philosophical implications, critical accounts of its relationship with the military, capital, or more abstract notions of technics, you can go to any of the “___ Studies” departments and they can give you an avalanche of reading on it. Likewise, if you want to know about feedback, you can go to the Engineering school and register for a course on Control Theory.
But cybernetics, properly, has always been both. The Janus-face is irreducible - it’s right there in that definition we started with: “control and communication in the animal and the machine.” Branches from branches...control in the animal, control in the machine; communication in the machine, communication in the animal…it’s not obvious, at least not on a first pass, that any of these have anything to do with each other.
Crate digging this archive means threading a needle: avoiding the temptation to lapse into either math or philosophy. How does the math inform the thinking; how does the thinking call forward the math? Or, more directly: what problem needed to be solved that required this dual kind of solution?
In this short series - a sort of amuse-bouche for our eventual reading of Varela’s Principles of Biological Autonomy - we’ll try to test-drive this sort of approach. Today, we’ll consider “Behavior, Purpose, and Teleology” (1943) by Arturo Rosenblueth, Norbert Wiener, and Julian Bigelow (freely available pdf).
Doing It On Purpose
“Purpose,” in the usual sense, seems to refer to something that some behavior has. If I smack you, it might be accidental or it might be on purpose. In most normal contexts, the significance of this is to license a particular response - you really shouldn’t get all that mad if I didn’t smack you on purpose. Determining the presence or absence of purpose, in these sorts of cases, is usually pretty easy. Of course people might lie - anyone who has ever had a sibling knows that “sorry, I didn’t do it on purpose” can be a bald-faced lie, and that’s itself evidence of purpose.
But in only the first few sentences of RWB, we see something like the set-up for a magic trick, since on the one hand they aim to discuss purpose, and on the other hand, they adopt a behaviorist paradigm - which is all but specifically designed to throw out the notion of purpose. Probably best associated with the psychologist BF Skinner, ‘behaviorism’ is a way of looking at some object of interest that restricts attention exclusively to inputs and outputs - in a word, treating everything as a black box: stimulus, response. Thoughts, intentions, feelings, meanings - all of these ‘inner’ kinds of things get tossed out, in favor of this pretty simple basic picture: when you put an x in, you get a y back. Whatever the f(x) is that =y is unimportant.
It would seem, then, that “purpose” could not be anything besides some f, or some ‘internal’ modifier of x. So how do they pull the rabbit out of the hat? They slice it up.
RWB note that the particular character of their taxonomy here is arbitrary: they could have divided ‘behavior’ up differently. They further admit that each cut largely just tosses out an uninteresting class of behavior for the sake of getting to a particular destination. Rosenblueth, Wiener, and Bigelow, in other words, don’t claim to carve nature at the joints.
But what we should attend to in the way they carve things up is that “purpose” - as they characterize it - stands in for many smaller causal interactions. What RWB point out is a little like the game QWOP (https://www.foddy.net/legacy/Athletics.html). The point of the game is that it breaks a simple, everyday activity down into components and suddenly it becomes really tough.
Our experience of running with ‘purpose’ is like if the game were “press R to run.” When I run, I don’t trigger each muscle individually, tell particular impulses to fire, consciously reflect on the proprioceptive signals I receive from my arm or deliberately proportion actions in response.
Though RWB use as an example the phenomenon of hand tremors in attempting to drink from a glass here, a more detailed and incisive version is offered in Wiener’s 1948 book. In Chapter 4, “Feedback and Oscillation,” Wiener notes that ataxia may be a consequence of two opposed malfunctions. In one case, damage to the spinal cord may prevent the central nervous system from receiving proprioceptive information about the position of their legs, meaning that a person would have to physically look at them to see where they are. In essence, the normal act of walking becomes a real-life game of QWOP. In the other case, an injury to the cerebellum leaves incoming proprioceptive information untouched but damages the relays that proportion muscular response: in trying to pick something up, a person overshoots the target, then attempts to adjust and overcompensates, resulting in oscillation.
The significance of these two examples is that in both cases, purpose is thwarted - a person’s purpose is to walk or drink from a glass - but it’s thwarted specifically because a circular flow of information gets interrupted. This is the critical point: it is not that one first hits upon some purposive action, and then the central nervous system sends the message downstream to the muscles. The physiological picture, rather, is that acting with purpose requires many streams of information, traveling to and from the central nervous system. And it’s all interconnected and interdependent: to do something requires a lot of countersignatures, all saying “yes, you’re doing the thing.”
It is worth pausing here to carefully evaluate what RWB’s claim is when they place ‘purpose’ downstream of black-boxed behavior and upstream from ‘feedback’. It’s not to say that ‘purpose’ stands in for a longer chain, as a kind of abbreviation the way that we might use f(x) to represent an infinite series we would prefer to not write out explicitly. But it’s also not to say that the mental state of ‘purposefully moving’ is a whole greater than the sum of its causal parts.
What they are saying is that when part of this network fails, then purpose (specifically as some sort of inner state) is present, but superfluous. If you are not able to walk because you lack proprioceptive information, “having ‘walking’ as your purpose” won’t help. Either the signals that enable walking get where they need to go, or they don’t. So purpose cannot be a stand-in for the causal interactions and outputs, nor can it be their sum, even if it seems like “purpose” is the software that allows us to interact with the musculoskeletal ‘hardware’. There simply is no ‘software’ layer. The whole isn't greater than the parts, or even quite the sum of the parts - the whole is the parts, and the parts are feedback loops, there and back again.
So RWB have made a definitional move via diairesis2. But having made 'purpose' equal to 'the topology of feedback loops', asking after behavior - in humans, animals or machines - no longer is a matter of 'mental stuff.' The focus of inquiry then must be the specific degree of success or failure in that network. Compensatory action is now what matters and a qualitative matter - purpose, intent, goal - has become a quantitative one - to what degree is that goal attained?
Let’s Be Discrete
A formal answer is not given by RWB in this piece, but what matters is that their philosophical analysis naturally raises the question of formalization. If purpose is feedback, then a workable model isn’t just clever engineering - it’s vindication of the thesis.
Here’s a very simple version that already exists:
This is called a Discrete Proportional Controller. It’s an iterated map, so it’s basically the same as a normal y= equation, but we replace x with xn and y with xn+1. Because we can start by letting n=0, it’s effectively “do this thing, then take whatever you get and do it again”, giving us x1,x2,… and so on. k and T are parameters that we choose values for: k is the gain and T the target.
In a physical realization, we don’t choose x0 - it will be the initial value of whatever quantity we wish to control; T is then the quantity we would like x to approach. k, you can think of like a correction factor. When we put in some number for x0, we first subtract it from our T: this gives the distance from our target, essentially “how much we have to correct for”. Then, of course, we multiply that by k and add it to the value. Note that - after scaling by k - we might end up with a positive or a negative number. That’s the beauty: if that term turns out positive, we know we’re undershooting the target; if it’s negative, we’re over. This means that adding it to xn results in an xn+1 which is (respectively) bigger or smaller than we started with. Try putting in T=10 and k=.5, with x0=0. You’ll see that we get x1=5, x2=7.5, x3=8.75,… We step halfway to 10 from wherever we started each time.
Note, however, that everything depends on k. You can see a few possibilities in the charts shown below. If it’s chosen well - if we pick 0 < k < 1 - the x values will converge monotonically over some number of iterations (top-left in the figure below). If we pick less well, we actually work against ourselves: this is the bottom-right case, where instead of converging, the values diverge like a snowball rolling downhill. The other two cases - top-right and bottom-left - are the ataxia patients, stumbling as they try to walk by eyesight alone or trembling uncontrollably as they go to drink from a glass.
There’s two interesting things here, though:
In this simple model, there’s nothing that dictates a particular choice of k. That it’s called “gain” is actually kind of clarifying: if you’ve ever dialed in a tone on a guitar amp, you know that there’s a point at which cranking the gain on your amp too high makes feedback - all we’re doing here is hanging some numbers on it! The system - the controller itself, as constituted - can’t calibrate its own gain. Finding the sweet spot comes from outside.
Everything depends on the reliability of measurement. That is: unless we actually know what xn+1 is with some level of confidence after correction has been applied, we won’t be able to dial in k properly. When you’re just working all this out on paper, this isn’t a problem - but in a real physical system, a whole host of other concerns come up: are we measuring right? what’s the latency? is there drift? does the correction actually take effect?
These two things aren’t just technical challenges, though, because they show how subtle epistemological matters are implicit - almost encoded - in the system: dialing in k properly is your job as operator, but your success depends entirely on how accurate the xn+1 data is. Whether you are reading it off of a thermometer, a speedometer, or some other similar kind of gadget, there is a critical link between “what should I do,” “what do I know,” and, completing the triad, “why - and how much - can I trust what I know?” The parts of purpose are feedback loops, remember, but here, what closes the loop? All that it can give you is snapshots, but it needs a history - and that comes from somewhere else. Whence, probably, the infamous IBM slide:
Since this article was written, of course, control theory has blossomed. There are many different techniques and models that can be used to regulate and optimize machine behavior. And yet, nobody would go so far as to say that RWB conclusively answered “what is purpose.”
So why did the two faces of Janus come apart? Why did the glue fail? It’s not a matter of “final answers in engineering but not in philosophy,” I don’t think. There is no question as to whether systems can converge on goals - that they do is obvious whether you do a lot of math or none at all. The controller converges on whatever T we pick - whether T=10 or =10,000,000; likewise, RWB picked their own T and then converged on it. We shouldn’t read this as “cheating” or anything like that.
The question, then, must be “where does T come from?” If you cover one of cybernetics’ eyes and look only through the engineering part, this question is trivial: one simply writes “let T be such-and-such.” To now switch - and look at ‘purpose’ philosophically - one sees God, Atman, Capital, the Good, Truth, Reason, the Self - any number of transcendencies. It’s not that any of these are ‘easy’ or ‘solved,’ but there are options and then you can move into a different problem space entirely.
But to take a synoptic view here, what then? If you attempt to reduce the notion of purpose, is there a way to avoid the conclusion that you have your own purpose? The answer is, in this case, a function of the question and the question a function of the questioner; the RWB project of wholesale reduction unearths an incontrovertibly real kernel which refuses reduction.
In this way, the actual dialectical challenge here is whether you, as an asker of a question, are then capable of subjecting yourself to the answer you find. So if there is a mistake here (and I don’t think it’s even necessarily a mistake that RWB make) it is the equation of goal-seeking with goal-setting. It is this disidentity that emerges only on a specifically cybernetic, specifically stereoptic view. For all that has been written about cybernetics as a techno-military-capital adjunct, I will tentatively here suggest that this view overemphasizes one of the two faces - the ethical was built-in from the beginning because the philosophy was irreducible…whether the gain got turned down on the ethical over time is a different question, of course.
NEXT TIME on FVFC:
“By gawd…is that Heinz von Foerster’s music??”
“literary intellectuals at one pole - at the other scientists, and as the most representative, the physical scientists. Between the two a gulf of mutual incomprehension - sometimes (particularly among the young) hostility and dislike, but most of all lack of understanding.”
Diairesis is the Greek term for ‘division,’ the classical philosophical instrument for sorting a kind into its species. It is interesting to consider - we can’t get into it here - RWB’s use of diairesis to justify their redefinition of purpose against Deleuze’s reading of diairesis (in Chapter 1 of Difference and Repetition and the appendix to The Logic of Sense) as not a dialectically sound method of sorting, but a rhetorical performance and a motivated tool of selection - in other words, downward-moving division is not the ‘bad’ counterpart to upward-moving Platonic dialectic, but the suppressed truth of Platonic dialectic itself. RWB’s deployment here is - by their own admission - a selective strategy.
Looking forward to reading Principles of Biological Autonomy together :) In his Der Spiegel interview, Heidegger introduces Cybernetics so:
---
Heidegger: Philosophy [today] dissolves into individual sciences: psychology, logic, political science.
Spiegel: And what now takes the place of philosophy?
Heidegger: Cybernetics.
---
While he's clearly not a fan, Heidegger identifies the need for a discipline that coordinates between the sciences that have dissolved into specializations. If anything the dissolution is even more extreme today, and in that sense, there's an even more urgent need for a new general science of connection and coordination between the sciences, replacing metaphysics.
I met Varela once as a grad student and his intellectual progeny many more times, and I always read his journey through Madhyamika, Cybernetics, Autopoiesis, Embodied Cognition etc as an attempt to unify not just the two cultures of C.P Snow, but also the many cultures of the world.
However, I don't think it was possible while Varela was alive. Two decades too early IMHO; we needed the extreme atomization of science of which generative AI is both the symptom and cause and the possibilities of unification that open up, now that the end state of the previous regime is clear in outline.
Perhaps the study of cybernetics is an attempt at understanding the microcosmic feedback-loop structures underlying all human and environmental systems which then allow for their commodification and control.