Week 40

I didn’t have many expectations for this conference, as I was something of a ring-in, joining as part of Leonardo Impett’s group as a panel chair. But it was really quite good! On the Sunday before the conference started, we wandered around the national pavilions at the Venice Biennale, managing to catch the end of the controversial and critically well-received Austrian pavilion in time to watch the bell. I also liked the artist statement of the U.S. pavilion:

Monday, morning panel

Mazviita Chirimuuta kicked things off with “A Genealogy of Thinking as Modeling”, a tour through the same genealogy that interests some of us at the Digital Theory Lab of late, namely the epistemology of science that Ernst Cassirer develops in the mid-century. Chirimuuta chiefly drew out Cassirer’s inheritance of the idea that thinking can be compressed through a principle of economy (as in ‘economical’, with smaller expense) developed in Ernst Mach’s theory of physics: “Physics is experience, arranged in economical order” ¹1 Mach, Die Ökonomische Natur Der Physikalischen Forschung, 197.. Mach frames physics as nothing but the compression of sensorial evidence; and thus science as nothing but a minimization problem. Cassirer would critique Mach’s data realist position, which has more than a passing resemblance to Helmholtz’s physical realism.

Ben Pulver then proposed a coinvolved study of the situationists and the cyberneticians, arguing that the former’s emphasis on making the inherent contradictions in topological structures visible offered us a more powerful metaphor than ‘opening the black box’ in the common-by-now question of how we might understand latent space in machine learning, a space that (otherwise) lacks representation and thus remains essentially unexplorable. Pulver argued that Alfred Jarry’s ‘pataphysics’ shows us how we can playfully and artistically traverse latent space. I agree with Pulver that scientificizing latent space as something quasi-sensorial and thus capable of having a corresponding physics is a drive that exhibits the same totalizing drive that Chirimuuta’s talk critiqued with Cassirer (much of Anthropic’s interpretability work exemplifies this Machian hankering) and that the black box metaphor is no longer instructive (if it ever was). I will need a little more convincing, however, that the Situationist Times’ takes on topology are our best reference for a deeper critical understanding of machine learning models.

Stefan Besser offered an analysis of Andy Clark’s ‘predictive processing’ (PP) framework for understanding cognition as a “cumulative world-model.” I hadn’t heard of PP before, but as Besser presented, it struck me that it seems a distinctly post-cybernetic philosophy in the sense that its notion of error as a correcting counter-measure to top-down predictions is almost exactly (it seems to me) compatible with the cybernetic notion of error. The source of this is Norbert Wiener, i.e. ²2 Wiener, The Human Use of Human Beings.. For an excellent introduction to the philosophical stumbling points in cybernetics, I recommend ³3 Galison, “The Ontology of the Enemy”.. Besser spoke about PP as a cumulative ‘world model’—one of the conference’s core catch phrases and putative objects of study—in the sense that errors are supposed to iteratively restructure cognition in relation to some hypothesized goal or aim. If Besser’s reading of PP is accurate, we could say that it simply reinstantiates Wiener’s notion of the brain as a servomechanism. The problem, as is discussed in Galison’s piece, is that this presumes a kind of hard objectivism of aims—that our ends as human beings are fixed targets that we can therefore concretely lock in—a theory that has hardly any psychological evidence, I would have thought. (It occurs to me now that the Hegelian humanism that Jensen Suther develops in his new book ⁴4 Suther, True Materialism: Hegelian Marxism and the Modernist Struggle for Freedom. might be productively counterpoised with the cybernetic/PP presumption.)

I unfortunately missed much of Aline Guillermet’s presentation after having to slip out, but returned to attention for Moritz Kriegleder’s attempt to map the types of consciousness science in circulation today. He noted that consciousness science is premised on the assumption that subjective experience can be (exhaustively) described in algorithmic terms—the exact point Chirimuuta takes issue with above. One interesting axis along which Kriegleder wanted to differentiate consciousness science was the extent to which research had an investment in computational functionalism as opposed to what he calls bare functionalism. At the end of the talk, I asked Kriegleder a question about a lead-in phrase he offhandedly used to justify consciousness science before going on to stratify it: “physicists don’t necessarily have a clear definition of what energy is, but they can still work with it.” The suggestion being: consciousness scientists can still work with consciousness, even though they don’t have a clear, good, or coherent definition of their object of study. I think that this is a clever turn of phrase; but it also seems to me that consciousness is a fundamentally different object than energy in some important ways. For one, we don’t have agreed-upon ways to measure consciousness. (Kriegleder noted in response to my question that energy measurement was/is also a contested activity.) It would be interesting to get into the details of this analogy to see whether it holds real water, but I for one am quite suspicious of the rhetorical move of likening consciousness science—or any computing science, for that matter—to the ‘hard’ sciences such as physics and chemistry, as I think it can lend unwarranted legitimacy to digital technical activities when not seriously thought through.

Monday, afternoon panel

Ellen Charlesworth and Amira Moeding opened representation for MVC with a presentation on Winograd’s Ph.D. thesis, SHRDLU. I unfortunately neglected to note the core argument, but they compared SHRDLU to Nvidia’s COMPASS, “a framework for cross-embodiment mobility”, and noted the latter’s association with the Platonic Representation Hypothesis ⁵5 Huh et al., “The Platonic Representation Hypothesis”.. I didn’t know about this hypothesis, but it seems to be the source of the latent idea that is driving a lot of ‘world model’ research, namely that reality isn’t fundamentally a co-production of irreconcilable perspectives that partake in it (what we could call the dialectical position), but instead can have a canonical, Platonic representation (“a shared statistical model of reality, akin to Plato’s concept of an ideal reality”). The representation of the world is not general or generalizable, but fundamentally task-specific, a fact that seems to be reflected in how algorithmic models seem to be able to do better on benchmarks when they are not presumed to derive from some universal and universalizing world model.

The rest of the panel consisted of talks by Lars Pinkwart, Emilie K. Sunde (new assistant professor at Durham with Louise Amoore’s initiative), and Hannah Peuker. Peuker took up the idea of the world model with respect to what she called ‘European subjectivism’, citing a new book by Yuk Hui that I didn’t know about, which apparently claims that technology is the central question of post-European philosophy. I need to read this book, as I’m curious as to why a philosophy of technology (which Yuk Hui is known for writing about) is necessarily post-European given that Hui also wrote a book this year about how we should return to Kant to think about machines and their bearing on philosophy. The lineage of technologically sensitive European thinkers from Leibniz, through Dedekind, Hilbert, Heidegger, Canguilhem, Stiegler, etc. also makes me wonder if Yuk Hui positions post-Europe as an aspiration or a conceit.

Monday, evening panel

I chaired this panel and so struggled to keep detailed notes while prioritizing time-keeping. But I was attuned enough to remember that Charlotte Merchant gave a wonderfully detailed talk called ‘Resolution Without Realism: Doubt, Critical Limits, and the Kilometer Scale’—detailed, because she surfaced some of the more recent literature on climate modeling, a space about which I know very little.

Tuesday, morning panel

More MVC representation from Alexandra Gilliams and Eryk Salvaggio, paired with incredible talks and a short film (video?) from Silvia Cipelletti and Jacqueline Rowley. Gilliams digested Benjamin Bratton’s and Antikythera’s political program of ‘planetary-scale computation’, which describes itself as “pragmatic to the point of being speculative.” As Alexandra had described to me a week or two ago, Bratton moves between academic and art/design venues in an attempt to restructure computation towards more emancipatory ends, theoretically and practically. I’m sympathetic to this idea, as I do not believe that academic publication is necessarily the most meaningful (read: political) contribution that one can make to the world. But I am also somewhat skeptical of the disbelief in academic work that many academics—maybe particularly those who dabble in art and other worlds—seem to exude these days. As I said to someone while walking to see the Biennale on Sunday evening before the conference, if one actually believed that academic work is ultimately inconsequential, then the logical political position that should follow is an abandonment (or even abolishing) of the University. Some do call for this, and I agree with aspects of the critique: the University as an institution is drowning under reckless and relentless waves of neoliberalization after neoliberalization, rendering tertiary education as valuable if and only if it directly places its students in high-paying jobs, and framing academic freedom as a fantasy rather than an ethical imperative and commitment. Yet I do believe that the production of thought (proxied, perhaps, by the production of books and essays that are called ‘theory’) can have consequential and maybe even essential political effects. This belief is why I want to write, because I see academic space as a rare environment in which one can create something that resists the logics that otherwise condemn our activity as alienated and unfree (capital, colonialism, racism, sexism, pick your poison). Bratton, for his part, seems to both believe in the political content of thought at times (he wrote a 500-page book on sovereignty and software) but also to refuse it, preferring ‘art and design practice’ as a more efficacious political activity. To be clear, this is not meant as a polemic against art or design. It is rather a question about whether academic work, specifically, can effect political change.

Eryk Salvaggio gave what I found a very moving talk on the dysmorphic hand, reading images of his partner’s hand that he had reinterpreted using generative image modeling through Lacan’s mirror stage, and in turn offering a robust critique of the ‘world model’ production of pixelated subjectivity, the mainstream algorithmic imaginary, through a careful engagement with psychoanalytic theory. I can only tease a few phrases that I jotted down from the paper, such as “we are already alien”, “is Salvaggio stuck in 2022?”, “the engineering of misrecognition”, “the other half of our cybernetic subject”, and “it is inexplicable that anything should respond to desire”, as I really recommend this piece to the reader once it gets published. My breath got caught in my throat when he ended with a staunch defense of the subject’s opacity against computation’s facile attempts to reduce it to rubble as nothing but data: “[This AI-image hand is] held up, as if to say, ‘stop searching for the world you see within me’”.

In the Q&A, Silvia Cipelletti also made a wonderful comment about how “the more we try to get something that is average, the more we have to give away.” She was speaking in relation to the experience of prompting an AI model to produce a scene for a film, I believe, but this phrase highlights the singularity of what it means to be a subject, rather than just a one-dimensional man, as AI would have us. Cipelletti argued that “the environment is an invitation”, not a directive or demand, a suggestion that I took up in a comment I posed to the (straggling) attendees in the open discussion that closed the conference, remarking that “we make our own abstractions, but we do not make them as we please”.

Keynote: Giuseppe Longo

The conference keynote was Giuseppe Longo, a mathematician who we read at the start of the year in the Math(s), Philosophy, History reading group I run online. His talk took up what he sees as the main aporia of mathematical thought, namely the difficulty of dividing between the continuous and the discrete. His history of navigating this tension began with the invention of writing, which discretized language into signs that could be inscribed in tablets. The invention of writing is not to be confused with the invention of the alphabet, which discretized “the continuous song of language” into not just signs, but recombinable words and letters. If I have him right, Longo’s claim is that the automaticity of reading—I understood this to mean our ability to glide over words and sentences without cognizing them—only becomes possible in Latinate languages. It is then only with Turing’s machine—the computer—that the automaticity of writing also becomes possible.

I have to think further on how/whether these points actually hang together, but I find them very evocative. Another claim Longo made is that there is an epistemological/cognitive difference between Gödel’s/Turing’s formulation of incompleteness, which he describes as coding letters by numbers, and Alonzo Church’s lambda calculus, which by contrast is coding numbers by letters. Again: evocative! Yet hard for me to parse at first glance. He also emphasized the importance of the Church-Rosser theorem as the basis for why “everything [in digital computing] boils down to imperative logic on zeroes and ones”, which in turn leads to ‘The New Imperative Pythagoreanism’, summarized by the dictum: I write, I program, I control. Longo also delectably offered a kind of redemption for AI trailblazers such as LeCun and Hinton, noting that they are excellent mathematicians who do very beautiful work, and that it’s a shame they have “a really miserable vision of mathematics and cognition”.

There is much to digest here, and I will be reading Longo’s work more closely in the not-too-distant future to chew on it all some more, I decided a few minutes into the keynote.