Hayek's Abstract Logic 9000

The use of knowledge in Searle's Chinese Room.

Friedrich von Hayek, in the time free from arguing that social democracies are a slippery slope to serfdom and laying the groundwork for the neoliberal consensus, came up with one of the first explicitly connectionist approaches to cognition. His 1952 book The Sensory Order proposed a framework for theoretical psychology based on the idea that the mind is a massive, interconnected system of pattern classifiers jointly possessing distributed knowledge that allows the organism to interact effectively with its environment. These ideas had clear affinity with his view of markets as distributed information processors, where no single individual or institution has complete knowledge of everything and yet spontaneous order arises in the entire system through the interactions of individuals and their local, idiosyncratic perspectives and objectives.

In a way, Hayek’s ideas anticipated the so-called “system reply” to John Searle’s Chinese Room thought experiment: Even if we grant that the person confined to the Chinese Room lacks “genuine understanding” of Chinese, it would be much harder to argue this about the entire system comprising (1) the person, (2) the rulebook for mapping queries in Chinese to appropriate responses in Chinese, and (3) the environment that supplies those queries and evaluates the responses, insofar as it can maintain effective and coherent communication. Even if the system knowledge of Chinese (knowing-how, in the Rylean sense) could not be localized, it is there in a distributed form. Searle claimed that he could get around this by asking us to imagine that the person could memorize the rulebook, exit the room, and then just go on to simulate the system with its putative understanding of Chinese still without any trace of genuine understanding. (For example, he wouldn’t have a way of associating words in Chinese to objects in the world.) But this counter-argument falls apart: the Chinese Room with the person inside and the person who has memorized the rulebook are two different systems because their sensory interfaces are very different.

However, we should go a bit deeper into how Hayek thought about the fundamental role of rules and abstraction. While all of these ideas can be found in The Sensory Order, it is better to turn to Hayek’s paper entitled “The primacy of the abstract,” which he presented at the 1968 Alpbach Symposium organized by John Raymond Smythies and Arthur Koestler. The symposium, whose theme was Beyond Reductionism: New Perspective in the Life Sciences, brought together a diverse group of people — in addition to Smythies, Koestler, and Hayek, it featured talks by Paul Weiss, Ludwig von Bertalanffy, Jean Piaget, Jerome Bruner, C.H. Waddington, and even Viktor Frankl. The main thesis of Hayek’s paper is that

the primary characteristic of an organism is a capacity to govern its actions by rules which determine the properties of its particular movements; that in this sense its actions must be governed by abstract categories long before it experiences conscious mental processes, and that what we call mind is essentially a system of such rules conjointly determining particular actions. In the sphere of action what I have called “the primacy of the abstract” would then merely mean that the dispositions for a kind of action possessing certain properties comes first and the particular action is determined by the superimposition of many such dispositions.

This is a very sophisticated view, positing a layered cognitive architecture making use of abstraction and virtualization based on pattern classification. What Hayek calls abstractions are essentially labels for behavioral strategies (“dispositions”) that codify the organism’s tendencies for taking various actions that are determined partly by the organism and partly by the environment. These routines involve adaptive sensing, trial-and-error, internal simulation, and various feedback loops. Hayek gives the example of a lion about to attack its prey, where the particular movements of the lion will be determined by various external attributes (of the prey, the terrain, the ground, etc.) and internal variables (the lion’s proprioception, state of alertness, and physical fitness). So, in Hayek’s terminology, rules are abstractions — they abstract away the whole infrastructure supporting and facilitating a given kind of action in response to a given kind of stimulus. In other words, the mind projects itself onto the world, and the actual state of affairs is determined by the specifics of the coupling of the organism and the world. (In support of this view, Hayek cites Hermann Helmholtz, the Gestalt psychologists, J.J. Gibson, and Maurice Merleau-Ponty, among others.)

He goes on to say that

[w]hat this amounts to is that all the “knowledge” of the external world which such an organism possesses consists in the action patterns which the stimuli tend to evoke, or, with special reference to the human mind, that what we call knowledge is primarily a system of rules of action assisted and modified by rules indicating equivalences or differences or various combinations of stimuli. This, I believe, is the limited truth contained in behaviourism: that in the last resort all sensory experience, perceptions, images, concepts, etc., derive their particular qualitative properties from the rules of action which they put into operation, and that it is meaningless to speak of perceiving or thinking except as a function of an acting organism in which the differentiation of the stimuli manifests itself in the differences of the dispositions to act which they evoke.

Let us now get back to the Chinese Room. Hayek’s point about the primacy of the abstract obviously applies here: The person confined to the room does not receive an unstructured stream of sensory impressions. Being sequences of Chinese characters, they are highly structured. Moreover, the operator of the room can use all of his senses, memory, and other cognitive abilities. Thus, just as Hayek says, the abstract precedes the concrete in the sense that there is already a kernel of basic rules and sensory strategies standing ready to be deployed. The operator can recognize, classify, and act upon observed patterns in the external inputs, the instructions in the rulebook, and the external outputs prescribed by the rulebook. On the basis of this experience, he will hone his repertoire of different types of action, and (assuming sufficient variety in the queries coming from the outside) this repertoire will evolve and grow richer. With time, the inhabitant of the Chinese room will be able to make predictions about which sequences of characters are likely to follow certain other sequences, and will be able to anticipate, to plan, and thus to interact with the environment more effectively. This evidently constitutes a form of understanding, and can we even argue that this is not how we should think about understanding and knowledge in general?

Indeed, this is exactly what Hayek says:

the organism responds to—and thereby, as I like to call it, “classifies”—the various effects on it of events in the external world. This is the limited extent in which it can be said that these action patterns are built up by “experience”. It seems to me that the organism first develops new potentialities for actions and that only afterwards does experience select and confirm those which are useful as adaptations to typical characteristics of its environment. There will thus be gradually developed by natural selection a repertory of action types adapted to standard features of the environment. Organisms become capable of ever greater varieties of actions, and learn to select among them, as a result of some assisting the preservation of the individual or the species, while other possible actions come to be similarly inhibited or confined to some special constellations of external conditions.

Frank Rosenblatt, the inventor of the perceptron, acknowledged Hayek’s influence on his thinking. We also can recognize the Hayekian evolutionary metaphor in Gerald Edelman’s neural Darwinism. The closing paragraph of “The Primacy of the Abstract” will surely resonate with those of us who appreciate John Doyle’s thinking about layered architectures:

the processes I have been considering occur not just on two but on many superimposed layers, that therefore, for instance, I ought to have talked not only of changes in the dispositions to act, but also of changes in the dispositions to change dispositions, and so on. We need a conception of tiers of networks with the highest tier as complex as the lower ones. What I have called abstraction is after all nothing but such a mechanism which designates a large class of events from which particular events are then selected according as they belong also to various other “abstract” classes.

Do large language models possess knowledge and understanding in this Hayekian sense? They are obviously extremely complex “systems of … rules conjointly determining particular actions.” They are getting better at classifying and categorizing their inputs, and their context-conditional probabilities jointly determine the repertoire of outputs. Models like Claude Code produce programs, i.e., outputs that are themselves functions with some arguments left free for the environment to fill in. More sophisticated capabilities for perception and action are not far behind, although for that we need to finally go beyond souped-up chatbots. As for things like intelligence and consciousness, I’ll leave you with this quote from Hayek’s response to Viktor Frankl’s question about self-awareness and self-reflection in the discussion section:

If, as I believe it to be the case, the mind can be interpreted as a classifying machine, this would imply that the mind can never classify (and therefore never explain) another mind of the same degree of complexity.