Between the Critique of Pure Reason and the Critique of Judgement Kant produces at least two senses of the term judgement, on the one hand a system of reason that through a kind of grammar, produces the world in relation the self-conciousness or in Kant's terms the unity of apperception and on the other we have judgement from the first half of COJ, referring to Aesthetic Judgement.
Beauty as a popular idea has it, is based on symmetry, proportionality etc. However Kant's question to this judgement would be, what kind of systems of values are the values 'symmetry' and 'proportionality considered beautiful. Is it a system where the utilitarian is judged in majoritarian terms, that is the larger part of mankind agrees on it, rather than being judged by something inherent in the value itself, a virtue.
In any case of late a sufficient amount of ink has been spilt on arguing that forms of the modern internet, through mega corporations have prdoduced a kind of centralization of the use of the internet tending towards the developement of technologies that require large scale infrastructures which tend in aesthetic preference to be inherently majoritarian which points to the universalization of values of beauty etc. However this is not Kant implies by beaty at all. In a certain sense Kant can be seen to have a very particular view of beauty, as imperfection and surprise, which falls outside the conventional wisdom sense of aesthetics as a generally agreed upon sense of beauty as encoded in particular values. As I note later in this essay, beauty as a majoritarian agreement on what values constitute the beautiful is clearly available to any LLM to make judgements on. It falls thus on those using these technologies to decpipher whether they can be of use, or if there is any other reason for their developement, given that they use so much resource, and whether perhaps less centralized, human-replicating AI technology might not be a useful path for this technological line to trace.
As was made apparent to me during my employment as a trainer for a particular LLM, such technologies are immersed not just in the actual interfaces that we colloquially called LLMs but everything from the chess engine we use, to service bound interfaces like Uber, Swiggy etc. A large part of our modern economies are being determined by technologies that are expected to make large scale operations possible and in some cases to make crucial ethical decisions. It thus becomes important to query the sense of judgement involved with such machines to categorize this on the basis of what we know about judgement. Here I am asking the question can an LLM make an aesthetic judgement, assuming that the question of beauty also harbours the question of ugliness, that is we are making an aesthetic evaluation of the world along with our immediate intuitional grasp of space and time.
Aesthetic Judgement refers to judgements of taste and is logically already a higher category than that of judgement associated with reason. Aesthetic Judgement does not concern the senses, the beautiful is not a sensorial pelasure. While Judgement associated purely with the intergrative capacity of the understanding allow us to form particular perceptions from a manifold of sensible intuitons, Aesthetic judgement is concerned with the harmonius play between the understanding and the power of combination or the imagintion. Aethetic Judgements concern the specific situations where a pre-given universal category is not available for a particular situation (the realm of determinative judgement). In this case another form of the faculty of judgement is made operative: Reflective Judgement, which informs us that the form we have encountered is beautiful- a cognitive feeling of beauty. In a sense here in his deduction of the beautiful Kant opens up a completely different sphere of affect: cognitive affects.
In substance the moment is beautiful because it produces a spontaneous reflection on the form of form, on the form of that reflects on the grammar of sentences, which for Kant are the minimal units of cognitive sense. That is we make sense of sentences before the individual words are illuminated as a consequence of this.
A further function of Aesthetic Judgement involves the exhibtion of the reflected feeling. Kant’s theory of the acquisition of concepts is
the traditional one while in the theory of Aesthetic Jdugemen he has developed an advanced theory with regard to the use
of objects. This view is developed from his doctrine of the difference but mutual dependence of concepts and intutions. We cant have concepts if we don’t know how to apply them but applying them produce instances of them in intuition. The more astract the concepts are the more difficult this will be. To exhibit a concept on the other hand means to associate with it in intuiton a manifold of a distinctive unitary temporal or spatial shape. The schematism of judgements and concepts in the COPuR, for example is the exhibition of the categories.
While the faculty of the understanding has a clear analogy in that in this case to the construction of a minimal grammar and logic that then generates concepts through a kind of predictive reading of sentences. This scholars have pointed out is indeed very close to the understanding (Quiloz and Beckmann 2024) which works through the formation of categories and forms of determinative judgement that construct a kind of vocabulary of universals, really a kind of grammar, based on an initial anticipative framework, which allows the machine to learn through a wide variety of situations how concepts are formed through the perception of commonalities between objects and the generalization of these commonalities across all objects.
H However, the countermove, from the intuition to the concept, from the particular to the universal seems difficult for the LLM. The last sentnece should apepar counter-intuitive to you, in that in the Understanding remember in Kant we move from a general schematism of categories organized by types of judgement. This is precisely why Kant is not concerned with analytic functions of gemotric features, which are apriori, but rather whith the objects subsidary properties which require a generalized sense of space-time to intuit. This is what Kant of course calls the synthetic.
ljn
l -x-
In the surprising turn of events that sees LLMs actually gain some kind of predictive and anticipatory function that mimics the understanding, the natural question thus becomes, can the LLM make an aesthetic judgement of beauty, which is contingent on the harmony of the imagination and the understanding.
In general views about the consciousness of LLMs are myriad, as myriad perhaps as the views of what consciousness actually is. In this series I follow the general idea that consciousness is a sense of first person experience or '"whats-it-likeness"' (Askell 2022) of an experience. It is thus that I thought it relevant to bring up the philosopher who's whole ouvre revolves around determining the conditions of possiblity of experience. Askell, in an article published on her substack prevaricates on the actual question of such consciousness in LLMs arguing that advanced computation might actually instantiate such a process and instead focusses on a related term: sentience. Can an AI feel pleasure or suffering? Askell focusses on this question foregrounding the question of pain and suffering relating it to research on the question that determines how animals feel pain.
However Kant's displacement of the schematism of pleasure-suffering in the questions of beauty and sublimity in the critque of judgement point to a disjuncture in our consideration of ethics. While at one level ethics responds to the utilitarian question of pleasure and pain as the basis for moral action, at another level aesthetic judgement points to an entirely different cogntitive universe, to the questions of consciiousness and the capacity for wonder, or curiosity or joy, a series of tempered 'cognitive affects'.
T The entire series of cognitive affects displaces the question of AI consciousness from the relation to the necessity for a biological substrate for consciousness, and also displaces theories that see consciousness as a kind of semiological illusion. While my own views on the subject in fact tend to the semiological, I find for the purposes of this essay that we might in fact take Kant's entaglement of ethics and aesthetics as given, to allow us to have a deeper examination of the question: are AI to be treated as agents with the capacity for some kind of minimal consciousness.
-x-
This iamge of Kant as a philosopher of cognitive affects immediately draws us to a scholastic problem in the interpretation of Kant. In her book Hegel Contra Sociology Gillian Rose (1981) brackets all of the social sciences as having being produced in the wake of a particular image of Kant. The image is a particularly dogmatic image and can be drawn from our earlier speculation on beauty as displaced from pleasure and suffering. If we are to assume that the basis of Kant is a kind of Normalized subject that simply produces its duty in the face of suffering, or despite suffering, Kant is reduced to a philosopher of virtue, of a definitive view of the objective world as entirely stabilizable. This is the Kant Rose argues that most of the Social Sciences have picked up.
Here we have an easy way to parahrase Kant's evaluation of Beauty: simply the universally agreed upon forms of beauty. In this case the LLM passes automatically, because as scholars have shown (Quiloz and Beckmann 2024) LLMs are fully capable of concept formation, to the extent of assimilating synthetic visual details about objects like the Golden Gate Bridge in California. Through a consideration of poetry written about the bridge, reports about the bridge, the LLM is capable of aggregating a concept that in some sense could even be seen to have visual and sensoral detailing. In fact an advanced LLM designed sheerly to produce impressionistic images of common human concepts, like the Golden Gate Bridge might even produce a level of detail, if it is a machine with sufficient cognitive and industrial power to exceed in language even great impressionists like Falkner and Proust. We might even have a kind of ethnographic scale that measure the thickness of the description on various markers like the engagement of all the senses, the depiction of detail singularly related to events and places, and it might produce an ethnographic account of a phenonmenon thicker and more impressionaistically vivid than Malinowski's Coral Garden texts.
If in fact Beauty is associated with art in a utilitarian sense as a merely quantiatively different value than pleasure, then a sufficiently powered LLM might be able to produce its own aesthetics. Consider for example the famous Go game between Lee Seedol and the google's Alphago machine which even in its early stages far surpassed Seedol's moves precisely because it was not concerned with making moves that were seen by the Go orhtodoxy as being aesthetically sound. In this regard in fact LLMs of sufficient power far surpass Kant's criteria, in that it produces a new aesthetic regime in the board game of Go, and in doing so also re-writes the moral standards of the game. (I will return to this game in great detail in my next post)
However, this does not answer the question we have posed via the Kantian standards we have set: can an LLM make a judgement of beauty given that beauty is not generalizable under other aesthetic values like 'the good' and certainly not through utilitarian considerations of pain and pleasure.
To reiterate the beautiful for Kant is a perceptual complex that does not have a universal predicate, the absence of which forces us to consider the very form of form, the very uniting substance that allows for comparisn through concepts. Here Kant liberates the meta-cognitive from the its prison in the COPuR where it is primarily as a result of the unity of apperception that the unity of concepts and finally the unity of perception is guarenteed. It is not the subject that is meta-cognitive here, but the very edges of subjectivity that are brought into play, a kind of liminal zone, where subjectivity itself is not strong an unified, the outer reaches of a galaxy, to use a metaphor very appropriate to Kant, who surely falls more in line with Copernicus, Gallileo and Newton as a figure than I believe with Descartes and Leibniz. At least this is the tradition of Kant that Schelling later usurps to change the direction of the dialectic from being purely a mirror function in the Fichtean sense to it being a kind of outwardly reflexive dynamic.
Beauty, and its judgement, in a very dominant interpreatino of Kantian morality, involves the exhibition of the use of concept and notmerely its acquisition.
In fact in this sense it might be argued that question of the consciousness of an LLM is not based on some kind of inner 'whats-it-likeness' of experience the hypothesis of David Chalmers that Askell (2022) discusses, but rather of the field of demonstration in which an LLM is placed. While on the one hand the question of whether an LLM is an Agent and a Subject and a Self (Kockleman 2006), I argue is still up for grabs its inclusion in political decision making (such as in wars) to evacuate the potential responsiblity a corporation or nation has for a particular action (more on this in the next post) demands that we do consider it to be a Person. But then again this might again displace responsiblity from the same corporation and nation and the alternative is the first solution where agency is placed in the hands of the persona of a corporation or a leader of a nation.
T The question of beauty and its judgement, I believe is one possible way to solve this contradictory set of legal scenarios involving AI consciousness. We might argue that in fact the ability to produce an account of beauty becomes a more radical measure of subjectity in an LLM. A consciousness of beauty does not mean a universal subject, but in fact the exact opposite, a faculty to find schemata in the absolutely particular, and the faculty to demonstrate singularity in experience. Here by the term singularity I mean something very different from the typical Kurzweillian ranting about a progression of AI developement from Artifical General Intelligence to something fully possessed with desires and goals for the future, but rather singularity in the sense of details so particular that they are necessarily the products of 'an experience'. Here I am displacing the involvement of experience in the discourse of American Philosophy through figures like Chalmers and Thomas Nagel, and rather pointing to the immanent, quasi-subjectless nature of experience, as a spreading web.
Where beauty is the singularity of a parituclar experience that is transposed to a formal universal dimension. Really beauty is a question of how we see the relation of ulitity and ethics in relation to a transposition of a parituclar to this formal universal dimension.
In what follows thus I attempt to trace two formal system of ethics, that trace how one might consider utlity in cases where the stakes are more complex than what simple untility allows. My purpose here is to query the idea that the ethical maxim for beauty is the possession of a kind of non-utlity, described by Askell as 'incomparability' and another philosopher, Daniel Goldsitck in a more formal utilitarian sense as the utlity of non-utlity. Instead I argue that Kant's sense of a judgement of beauty is in fact a core problematic not just for aesthetic judgement but for how we view the process that produce what is recognized in public discourse as acceptable knowledge. Our epistemologies I am arguing are grounded on aesthetic judgements, that cannot be solved within utilitarian systems, even through inversions of the ethics of these systems through considerations of more inter-subjective frameworks like that of infinite ethics.
-x-
Amanda Askell’s dissertation, Pareto Principles in Infinite Ethics, develops an argument around utilitarian principles, eventually showing that ideas like: if all the agents of one world are permutatively compared to agents in an identical world, then a quantiative goodness in one agent being greater in the first world from the second implies that the first world is more good than the second. This all changes when we consider for a variety of reasons infinte populations and infinite consequences, the latter being the substantial basis of the ontologically defined antinonies that end the COPuR. Askell begins with a deceptively simple intuition—Pareto improvement—and shows that, when extended into infinite populations, it generates deep structural instability in ethical comparison. The core of her argument can be reconstructed as a sequence of formal commitments about how to represent worlds, how to compare them, and how these comparisons behave under transformation.
At the foundational level, Askell represents a world w as an agglomeration of utlities across a population of agents. In infinite ethics, this becomes a sequence traced from agents to real-world utility values. FOrmally we can write a world as:
w = (u1, u2, u3...)
where each ui represents the welfare of an agent i. In Askell's examples, or at least many of them these worlds are countably infinite. This allows comparisns between wolds at infinite points of comparisn.
As opposed to utilitarian principles, the world is not arranged in an aggregative, or scalar way associating the world with some number that marks total utility. INstead the world is arranged into a set of vectors in infinite dimensional space. (This bares close resemblance to the problem solving spaces of LLMs which also generate a great degree of complexity by aggregating details at higher and higher dimensions which is preciely why it requires so much machinic-cogniitve capacity which makes community based and sustainable AI that perform such tasks unthinkable at least for now- which raises many questionsin itself about the ethics of such large models, however more on this in the next post).
Ethical comparisn thus in Askell is produced by treating all such worlds as something close to a vector space W ⊂
R^N, where ethical comparisn is distributed over these vectors rather than reduced to a single number.
THis form of comparisn also requires a new operator, which is neither better or worse, but 'as good as' : ≥. Askell constructs four central theorems using this operator:
1. Agent Sensitivity (Pareto)- If there are are two worlds that contain the same agents and if every agent is as well off in the first world (w1) as in the second world (w2) and at least one is strictly better off in w1 then w1>w2.
2. Transitity- If w1>w2, and w2>w3, then w1>w3
3. Permutation: If we permute the positions of agents in the series then the result remains the same
4. QUalitativeness: If two worlds are to be structurally identical the relative positional value of agents in the series must remain the same
Taken together, these axioms define a comparison structure that is invariant under permutation, sensitive to local improvements, and globally coherent.
However what Askell eventually demonstrates is that these four axioms are not sufficient to jointly provide complete ordering for infinte worlds. This may be demonstrated as follows. Consider two worlds such that:
{i| u1 > u2} is infinite
and {i| u2 > u1} is infinite
That is infinitely many agents are better off in w1 and infinitely many are better off in w2. In finite settings aggregation can produce an adequate result. However in infinite situations this is not possible.
Askell strengthens this intuition through permutation constructions. By rearranging the utilities of w1 and w 2 she generates new worlds w1' and w2'.
-
by Pareto
-
by Pareto
-
but transitivity would force contradictory rankings
This produces what she calls cyclic or four-world structures, where:
From which transitivity would imply an impossibility.
The deeper mathematical structure here resembles paradoxes as is often the basis for paradoxes . Paradox is pcisely the zone of minimal conditions of subjectivity, or ordering and yet not without some minial ordering.This produces her core result: incomparability.
-x-
I would like to now present a possible variation to this view through a kind of case study through a thought experiment posed by the philosopher Daniel Goldstick (1991) in his article 'Distributive justice and utiliy'.
Goldstick asks us to consider the following argument posed by the economist William J Baumol:
'If Ellen likes marmalade and Daniel likes jam, how do we know what happens to social utility when we produce more jam and less marmalade?'
A further stipulation to this situation is Hellen would prefer the marmalade far greater than Daniel would prefer the jam.
THere is in the question an inherent tension between need and desire, and whether a system should reward hedonism or desire, as well on the other hand how a system is to balance its overall good, in a utilitarian sense in relation to need.
In Goldsticks example we have a world w such that:
w = {ue , ud} , that is the aggregative utilities of Ellen and Daniel. There are three principles to compare compossible worlds, based on Goldstick's analysis:
1. A comparisn of magnitude of utility can be made if the total utility in one is higher, which in this case implies giving Hellen what she wants, because according to a utlity principle, since Ellen's happiness would increase to a greater extent than Daniel's
2. A comparisn of magnitude can be made if the goods are equally split on the equalitarian principle.
3. A comparisn of magnitudes can be made if goods are split on the basis of 'equally dividing good and evil' or a distributive principle which provides daniel more.
The results in relation to each other however for two identical worlds, w1 and w2 are incomparable because the results of which world will be better off in the case of the application of either of these three pincrimples. In the end however Goldstick concludes with a positive result. He argues that in the end the only princple that can actually work through the situation of a kind of quasi natural social utlity is the principle of utlity itself which is the only principle that can guarantee the isolation of each actor in relation this general quasi-natural social utility, is effectively added to the final result. That is he suggests, that each individual must be equally commited to negating his utlity in service of general social altruism. THat is each individual to maximize overall utlity engage equally in non-utilitarian activity to produce a maximized utlity.
-x-
Now I would like to make an attempt to compare Kantian structures of aesthetic judgement to these utlitarian and non-utilitarian considerations.
I would like to start with a simple reference to a Swedish movie from the 2000s called Together or in Swedish Tilsamanns. It proposed a simple thought expierment, a commune, out in the wilderness, peopled by students and workers, who live together, sharing all property and benifits, and minimizing personal utlity for the greater good. However in order to not be hedonistic the commune prefers monogamous unions, and the relations are sexually non-binary. The system of common property and monogamy of course square of and cause a kind of existential crisis in the commune.
This is a slightly extreme example but a simple one, that describes how in fact in any situation of political judgement, there are sub-systems of utlity, even non-utlity systems that produce general utility. The incompatiblity of these worlds produces fissures, which are exactly the points at which some kind of minimal subjectivity is prodced.
Let us extend the example. Say the commune decides to turn take on a monastic function and eschew all love relations in favour of a daily investment in some aesthetic pleasure, assuming following kant that art, which is the central subject of aesthetic exhbition, we might say in some kind of creative culinary experince, or in collective music or magical ritual, etc. Here we are implementing a utilitarian reading of Kant's Aesthetic principle, arguing that investment in Aesthetics might uplift a people above their self-denigration into ethical considerations that involve pleasure-pain principles as the basis for judgement. Instead judgements of beauty and moral character, we might read as rising above pety hedonistic utlity.
Let us assume a community of three individuals. They are all invested in such aesthetic upliftment and an immersion into producing judgements of taste, which as Kant suggests are higher than judgements of pleasure-pain. Let us assume a fairly diverse result:
A gains immense disinterested enjoyment from the aesthetic exhibition
B gains some mild personal enjoyment from it
C derives no enjoyment and is irritated
In the case of a collectively prepared meal all derive the same degree on enjoyment.
Kant's claim I believe is that clearly A produces the highest overall utlity, but the question is really how do we convince B and C to agree that this is the highest overall utlity. This agreement in Kant's sense would require an exhibition, a demonstration, or in more comprehensible terms, a judgement of the world that produces the conviction that disinterested pleasure is of the greatest value to general human happiness.
THat is it is not a question of a scalar aggregation of ulity as an overall number but rather of a qualitative shift produce through a singular act, that institutes a microsystem of a new normative framework. Kant's focus in the COJ is clearly on the value of agency, in the constitution of human subjectivity.
-x-
Microsystems of political relation, consistently produce friction. Here subjectivity is not a kind of unificiation of the relation of the self-to-the-self, so that from one moment to the next it is the same self constituting itself, in a closed loop, but rather subjectivity is diffuse when seen through an aesthetic prism, it stretches out like a web through nature, it is liminal.
It is singularity in the sense of a historical moment, undeniably like Walter Benjamin's dialectical image, that surfaces ever so often in the Arcades Project, a moment of historical consciousness produced by a moment of political and physical danger.
That is utility is to be redefined if we follow Kant, to include an act that exceeds utlity, which defines a new normal for what utility is. Really the human world of perception is shaped, in the end for Kant, by the creative function of such acts.
Consider the case from around August of last year, where a 16 year old, killed himself, after what is described by the press as 'months of encouragement from chatgpt'.
In the end of course, it is necessary to veer towards Askell's result of incomparability, however in a very specific sense. In our world it is already clear that such technologies are producing acts that necessitate judgement, that exceeds utilitarian considerations, precisely because these acts exceed the usual bounds of normativity and produce new situations, particularites that mark singularities in the historical developement of LLMs.
From this point of view at least, LLMs are as conscious in the quasi-subjective web of matter and mind, as any of us, in that they have in a psychoanalytic 'case-file'. In a social sense it is clear that these programs have psychological morbidities which already presuppose the need to conceptualize them as at least having the liminal consciousness so characteristic of aesthetic judgement, preciesly because aesthetic judgement extends Kant's framework from the subjective to the inter-subjective, however not through a consideration of inter-subjectivity being some kind of ultimate utiliraian virtue in non-utility, but rather because it is only actualized in an act that extends the bounds of normativity. HOwever if we must conceptualize LLMs as sucwe must do so in the full Kantian sense, by drawing a historical image of what they are as our perception of the apperception of these machines. We must assume, not only that they are agents, and have personas but also that they have selves (Kockelman 2012), that is histories, of acting as agents in the human social framework.
The question, finally of whether they are subjects or not, I will consider in the next post.
References:
- Askell, Amanda. 2018. Pareto Principles in Infinite Ethics. PhD. New York University.
- Askell, Amanda. 2022. My mostly boring views about AI consciousness. Substack.
- Goldstick, Daniel. 1991. Distributive Justice and Utlity. THe Journal of Value theory.
- Kockleman, Paul. 2006. Agent, Person, Subejct, Semiotica 26 (2).
- Mattheu, Quiloz and Pierre Beckman. 2024. Why we care about understanding: Competence through predictive compression.
-Rose, Gillian. 1981. Hegel Contra Sociology. Verso.
Comments
Post a Comment