The Daoist Image of Control (III)

The dao that constrains is the dao that deconstrains.

Those who know others are perceptive
those who know themselves are wise
those who conquer others are forceful
those who conquer themselves are strong
those who know contentment are wealthy
those who strive hard are resolved
those who don’t lose their place endure
those who aren’t affected by death live long

— Dao De Jing, chapter 33 (translation by Red Pine)

“Perception is external knowledge. Wisdom is internal knowledge. Force is external control. Strength is internal control. Perception and force mislead us. Wisdom and strength are true. They are the doors to the Tao.”

— Li Hsi-Chai’s commentary on Dao De Jing 33

As I wrote in Parts I and II, the central thesis of Chad Hansen’s A Daoist Theory of Chinese Thought is that classical Chinese philosophy viewed language as a control technology, a tool for coordinating and regulating behavior. Putting on our control engineer’s hat, we can therefore pose two questions:

1. What are the capabilities and the limitations of language as a control technology?

2. If we are to make the best use of this control technology, what should be the internal structure or architecture of a good regulator?

The first question is about what we can and cannot do with language; the second question is about realizability. According to Hansen, the Chinese answer to the realizability question is different from the Indo-European one. The latter is intimately linked with symbolic representations that are manipulated according to the rules of propositional logic. It’s an old idea going back at least to Leibniz and Hobbes. For example, in The Leviathan Hobbes writes that

reason is nothing but reckoning - that is, adding and subtracting - of the consequences of general names agreed upon for the marking and signifying of our thoughts.

Experience in the form of sense data impinges on an information processor that applies a series of truth-preserving transformations. As Daniel Dennett put it memorably in The Intentional Stance, a reasoning subject is just a theorem prover wearing an overcoat of transducers and effectors. This is the GOFAI paradigm of Minsky, McCarthy, Simon, Newell, Pylyshyn, Fodor, etc. By contrast, the Chinese answer is more like connectionism or like the Hayekian sensory order. The processing is parallel, distributed, and nonrepresentational (at least in the sense that the entities being manipulated are high-dimensional vectors rather than logical variables, and the transformations acting on them cannot be readily mapped to some sort of a truth-preserving semantics). It’s pattern recognition and pattern prediction all the way down and all the way up, and we adopt the pragmatic attitude — everything is fine as long as our de (virtuosity) acts to align the discourse dao (laws, customs, norms, regulations) with the performance dao (our behavior in the world, including speech acts).

In this context, we can think of LLMs as models of linguistic control systems that seem to vindicate the Chinese (or connectionist) view of language. Their internal architectures are manifestly nonpropositional. The French philosopher Jean-Pierre Dupuy, in his book On the Origins of Cognitive Science, suggested that one can view the construction of mechanistic models of the mind through the lens of Giambattista Vico’s constructivist thesis of the equivalence (or interconvertibility) of the true (verum) and the made (factum). As finite beings, we are best at rationally grasping those phenomena which we can model and simulate. As long as we don’t stumble into the fallacy of identifying the brain with the latest neural net architecture, the existence of LLMs counts as evidence against the necessity of symbolic mental states (World 2 entities of Popper and Eccles) and for the sufficiency of the connectionist sensory order. We can now move on to the question of capabilities and limitations of language as a control technology. We start by carefully examining the notion of tokens.

Rectification of names as the token grounding problem

The concept of tokens, “the universal interface” of LLMs, maps neatly onto the classical Chinese concept of ming (names). Ideally, names or tokens should not be analyzable into smaller units. That is, they should be grounded in the capacity for making appropriate elementary bian (distinctions) which, in turn, engender more complex distinctions and actions. Stevan Harnad referred to this as the symbol grounding problem; in classical Chinese thought, this was the problem of the rectification of names. Harnad’s proposal was to interpret symbols as conventional names attached to combinations of two basic types of nonsymbolic representations: iconic and categorical. I won’t belabor the details of Harnad’s framework but appeal instead to a somewhat different proposal made by Manuel DeLanda in his book Materialist Phenomenology.

DeLanda’s objective in that book is to develop a philosophy of perception that does not presuppose language, but which nevertheless recognizes the importance of control using some sort of a sign system or a code. Just like James Beniger, DeLanda situates the first appearance of controls of this type in the first living systems:

when sensory organs began to evolve, the world had to already possess something they could exploit: the water, the air, as well as the ambient light that bathes the planet’s surface had to be populated by signs. But what kind of signs could these be? Certainly not symbols, since these can stand for something else only through an arbitrary social convention. But there are two other kinds of signs that do not depend on human communities to exist and could, therefore, be candidates for the natural signs we require: indices and icons. Roughly, an index is a sign that stands for an object by being a causal effect of it, while an icon stands for an object through a relation of similarity.

Examples of indices include smoke indicating fire, animal tracks indicating the presence of prey or predators, a change in the color of flame indicating the presence of a particular reagent, or the change in the conformation of an allosteric enzyme indicating the presence or absence of a substrate that can bind to it. Examples of icons are topographic maps encoding things like different elevations in a given terrain or the pattern of grooves on a vinyl record standing for the sound recording. One can even imagine hybrid iconic-indexical representations. DeLanda’s very nice example of one is using an app like Google Maps on your smartphone: the layout of the map on the screen is an icon that bears a relation of similarity to the physical layout of your physical surroundings, while the dot corresponding to your current location on the map is an index since it is causally linked to your physical location. DeLanda argues that the evolution of biological complexity made liberal use of such iconic-indexical semantics, rather than symbolic representations. Components in biological systems are characterized by their capacity to produce icons and indices and to consume icons and indices produced by other components when they are linked together. The most elementary such components are mindless cognitive agents, yet they serve as building blocks for organisms and for minds.

DeLanda’s emphasis on icons and indices as natural signs provides a viable solution to the grounding problem without foreclosing the possibility of much richer generative semantics using symbols and natural language. Moreover, systems that rely on icons and indices for control and coordination can be implemented using connectionist principles. For example, the structured activation patterns in multilayer neural nets, such as the ones discovered in vivo by Hubel and Wiesel or modeled in silico by Olshausen and Field, can be interpreted as icons and indices standing in appropriate relation to either the external stimuli impinging on the sensory surfaces of the neural net or the internal stimuli generated within the net. They become explicit, or intentional, signs as soon as they are recruited by other assemblies in the system that can make use of them. The key to biological complexity, and ultimately agency and intentionality, is in the rich architecture of systems that can be built from such simpler assemblies.

The limits and the possibilities: Laozi and Zhuangzi

The significance of this perspective for us is twofold. First, by recognizing the arbitrary, conventional nature of language as a symbol system, we can question the constancy of the dao of language and linguistic systems, either human or artificial. As Hansen points out, the concern with constancy was of primary importance in classical Chinese thought. Some, like Mozi and the neo-Mohist School of Names, sought to ground the constancy, reliability, and appropriateness of linguistic dao in quantifiable objective standards. Others, like Laozi, adopted the radical skeptical attitude recognizing the fundamental limitations of language and arguing for the inherent impossibility of a constant dao rooted in convention.

The famous opening lines of Laozi’s Dao De Jing are usually rendered as follows:

The way that becomes a way
is not the Immortal Way
the name that becomes a name
is not the Immortal Name

According to Hansen, however, this standard translation introduces a monist, mystical interpretation in terms of a single ineffable Dao, which is not justified. For one, elsewhere in the text of Dao De Jing there are references to multiple daos (the great dao, the dao of heaven, the dao of water, the dao of the dark moon, etc.). Moreover, if we place Dao De Jing in its historical and philosophical context, we cannot simply assert that Laozi’s usage of dao is radically different from how this notion was used by Confucius or by Mozi. All of these texts use the concept of dao in roughly the same way, as either a prescribed or a realized pattern of behavior or action. From this perspective, the opening lines of Dao De Jing are not about the ineffability of some “cosmic Dao,” but about the inherent limitation of any dao that can be expressed in language. Language is a social control technology for humans and by humans. As such, the distinctions induced by any discourse dao and implemented in any performance dao are based on arbitrary convention and can therefore change arbitrarily. As Hansen writes,

both Confucius and Mozi try to select some prescriptive discourse to be made the constant discourse guide for society. Both understand that the guidance must include training in the use of names— rectifying names — to make the discourse dao generate the intended performance dao. This constancy goal, Laozi announces, is hopeless. We have no way to fix how we may project our use of names in new circumstances. We cannot know whether we have chosen to mark slightly different distinctions or the same. Our social training (learning the code or following models) has no clear implication about how to project in new circumstances. So we cannot guarantee constant guidance from any dao that is generated by following a discourse consisting of names.

Hence, Laozi is concerned with the reliability of language as a control technology and with its inherent dangers. Hansen puts it thus:

For Laozi, what lacks constancy is not the experienced world of particular physical objects, but the system of name use. No unchangeable systems of discourse exist. This is so not because things change, but because names (and their distinctions) do. We are, so far at least, not dealing with a Chinese Heraclitus reflecting on how rapidly things change.

…

Names mark distinctions, not classes of objects. Their knowledge is of ways to do things, including making distinctions and using names. That knowledge must change since each situation is unique. Daoists interest themselves only in this observation. Independently of any actual flux in the world itself, our systems of guidance attach to the world in constantly changing ways depending on conventional, and therefore changeable, practices. That is the philosophical problem about constant guidance that captures the attention of Chinese theorists.

Laozi’s vision is, in this regard at least, radically anti-technocratic. Look at Chapter 80 of Dao De Jing, for example:

Imagine a small state with a small population
let there be labor-saving tools
that aren’t used
let people consider death
and not move far
let there be boats and carts
but no reason to ride them
let there be armor and weapons
but no reason to employ them
let people return to the use of knots
and be satisfied with their food
and pleased with their clothing
and content with their homes
and happy with their customs
let there be another state so near
people hear its dogs and chickens
but live out their lives

This is not so much a rejection of technology as a cautionary message that any technology can condition and control us as much as we think we can condition and control it. The worry about such loss of autonomy was expressed eloquently by Ivan Illich in his essay “Silence is a commons:”

machines which ape people are tending to encroach on every aspect of people's lives, and … such machines force people to behave like machines. The new electronic devices do indeed have the power to force people to “communicate" with them and with each other on the terms of the machine. Whatever structurally does not fit the logic of machines is effectively filtered from a culture dominated by their use.

The machine-like behaviour of people chained to electronics constitutes a degradation of their well-being and of their dignity which, for most people in the long run, becomes intolerable. Observations of the sickening effect of programmed environments show that people in them become indolent, impotent, narcissistic and apolitical. The political process breaks down, because people cease to be able to govern themselves; they demand to be managed.

Daniel Dennett’s fear of “counterfeit people” is another expression of this view, and the same idea is expounded in more detail in David Runciman’s recent book The Handover: How We Gave Control of Our Lives to Corporations, States and AIs. Laozi was worried about much the same things in China of 6th century BC, and even the recent piece by Marion Fourcade and Henry Farrell about how LLMs could upend the rituals underlying human organizations could as well have been written by Laozi about the rites and regulations codified by Confucian bureaucrats.

However, we can temper the pessimism of Laozi somewhat if we recognize and embrace the plurality and the open-endedness of the varieties of dao embodied either in language or in other types of sign systems. Language is inherently ambiguous, but we have no way of escaping it (aren’t we using language right now?). The generativity inherent in complex system architectures making use of icons, indices, and signs is simultaneously a source of existential unease (in the Heideggerian sense) and of wonder. While Laozi was all about the unease, Zhuangzi was all about the wonder. He urged us to recognize that, while there is no “god’s eye view” of reality, there are different perspectives that can be linked together, at least locally, because our actions and discourse (coming, as they are, from within our particular perspective) by means of qing, or “reality feedback.” Hansen puts a Kantian spin on this:

Kant seemed to think of his parallel conception of the thing in itself as a manifold. It was a multiple source of feedback responses or sensations. Daoists are theorizing in a tradition that does not stress constructing concepts out of sense experience. They deal with using language to make distinctions. A Daoist form of transcendental idealism would naturally tend to characterize the thing in itself counterpart as a one. Where Kant imagined knowledge as unifying a manifold, Daoists imagine it as distinguishing a whole into its parts. Zhuangzi, like Kant, allows that we cannot know anything about the absolute object of our conceptual systems. We cannot know, in particular, that it is either one or many. Our metalanguage of a higher or once-removed perspective shows us only a perspective on the plurality of perspectives. Mysticism and skepticism emerge together. What might be one is what we cannot in principle know, given this concept of knowledge.

Of course, we can only interpret this reality feedback through the system of values inherent in our own perspective, which we can nevertheless compare with those of others. This feedback is a form of shi-fei (this-not this) reinforcement learning. Zhuangzi talks about spontaneous practice and mastery of various skills in this context, and he celebrates this in every domain of life, no matter how exalted or how humble (see, for example, the famous passage about the butcher Ding). This is remarkably reminiscent of the dinstinction between closed-loop control as conscious ratiocination and open-loop control as learned automaticity — learning a new skill, such as playing a musical instrument or even learning to walk, relies at first on deliberate conscious feedback and self-supervision and then gradually transforms into a smoothly practiced routine. This also echoes Heidegger’s distinction between simply being-in-the-world and adopting the theoretical attitude, which we have discussed earlier. Once again, Hansen draws a distinction between symbolic, serial information processing and connectionist, parallel processing:

Zhuangzi's conception of action at a high level of skill stresses the spontaneity of the response to a total situation. That response is nearly immediate because the feedback process becomes second nature. In initial training we learned distinctions consciously, deliberately, and with frequent correction. When we have picked it up, we shift the distinction making out of consciousness. It is as if our consciousness were the central processing unit and the distinction is made by an unconscious parallel processor. The parallel processor sends the shi-fei result to out consciousness. We come upon things as such-and-such. Our response to it is as if we had responded unmediated to the world in which the the distinction was a given.

…

The achievement of skill mastery becomes the mature Daoist notion of spontaneity. Intuitive, immediate, nonselfconscious, intimately aware sensitivity to context in performance marks this mastery. In cybernetic terms, our actions involve constant feedback mechanisms which operate at such speed and accuracy, that they resemble massive parallel processing. It is not that I should turn my mind off. The point is that a parallel processor now handles my walking, which once took my full concentration. It frees the central processing unit for other activities: concentration on reading the map or carrying on the philosophical argument. Central processing consciousness also kicks in whenever we are learning or coming to a hard place. Sometimes I have to pay attention to my walking and cannot continue the philosophical conversation. In normal skilled action, the mind is processing a vast number of clues at once. It guides our action without routing the information through our con- scious central processing unit. Our mind is both shut off (the central processing unit) and yet fully turned on (the parallel-processing feedback guidance of action).

This computer analogy of the Daoist view of intuition or spontaneous action gives us a way to illuminate the contrast in the theory of mind in the two traditions. The Western concept of mind and of the computer is of the information processor. The CPU takes in information. The processor exists relative to a cognitive stuff, information. The unit stores information in memory cells, operates on the informa- tion, and reports the result, information output. This reflects the Western focus on conscious thought, deliberation, and choice as cognitive.

The Chinese view of heart-mind focuses on guiding behavior. The mind receives reality-feedback inputs (qing) and processes them all at once (paralleling processing). The output is not a computational result stored in some memory cell or reported out as information. The output is an action. When we have learned to do anything like second nature, we constantly adjust our performance to myriad clues in the environ- ment. We do not mediate the fine adjustment of motor skills by conscious choice. We act, it seems, directly in response to the external environment without CPU involvement.

The dao of generative architectures: constraints that deconstrain

The overall view that emerges from all of this is of a complex layered architecture, where the operation of higher layers makes use of virtualization and abstraction through the interface of icons, indices, and symbols. This idea, discussed in an insightful paper by John Doyle and Marie Csete, is the key to all systems relying on organized complexity and on sign systems as interconnection interfaces. Language is one such architecture, but the same principles underlie the workings of biological organisms, social and legal systems, and technological infrastructures like the Internet. The range of possibilities that could be explored by engineered systems that communicate via the interface of tokens is endless, and so is the range of possible failure modes, including catastrophic ones. This obviously holds for massive connectionist systems that produce and consume various types of signs (or tokens, in LLM-speak). Borrowing an idea from biology, Doyle and Csete speak of all such complex architectures as constraints that deconstrain. The basic idea is that, even though there are constraints on the repertoire of basic building blocks (say, a given system of icons, indices, or symbols), these constraints can be transcended using compositionality, abstraction, and virtualization in higher layers. This is the dao of all complex system architectures.

Zhuangzi and technology

My own perspective is Zhuangzi’s optimism tempered by Laozi’s recognition of the limitations of technocratic governance. Hansen puts it nicely as follows:

Zhuangzi's supposed attitude to modern, rationalized daos such as science, rational morality, or the rule of law would be that they all presuppose something. All rational daos are systems of shiing and feiing based on the desirability of calculation and the realist's regulative ideal, the assumption of a single correct answer. But it does not follow that he would reject them. We can take Zhuangzi's fanciful discussions of the people who can survive in fire, be warm in freezing temperatures, fly in the clouds, stride on the moon, and wander beyond the four seas as speculative predictions. There could be ways of assigning shi and fei which would lead to such accomplishments. Science is an example of such a dao.

In making that point in ancient China, however, Zhuangzi was urging exploration of new daos. He did not have a clear conception of the scientific dao of hypothetical-deductive reasoning. Meeting science now, he might applaud this marvelously beneficial dao. He may also note that we measure its success using its own internal standards of cheng (completion). Who knows, after all, what other ways of getting skill over nature might still be possible? Why rest satisfied with science as a dao? Something, as much more powerful than science as science is more powerful than peasant divination, may still be discoverable.

Should that possibility justify skepticism of science? Zhuangzi's answer, I suggest, would be no. For us, science is the usual, conventional, shareable system of settling shi-fei. There is certainly no reason to abandon it, in fact, one can get as proficient in its evaluative standards as in any other craft or skill. Thus there is no reason to regard Zhuangzi as either antiscientific or antirational.

Happy New Year!