Audio of my talk on “The Future of Intelligence and/or the Future of Unintelligibility” (from the Locarno Film Festival’s Long Night of Dreaming about the Future of Intelligence, Aug. 9, 2023), followed by a conversation with Film Comment Co-Deputy Editor Devika Girish, is now online on the Film Comment Podcast.
I was dealing with jet lag, and it was a late evening event, so the talk gets off to a somewhat rocky start but fairly quickly settles into a groove. Devika Girish was a great interlocutor and asked very good questions.
The following is an excerpt of my talk from the Locarno Film Festival, at the “Long Night of Dreaming about the Future of Intelligence” held August 9-10, 2023. (Animated imagery created with ModelScope Text to Video Synthesis demo, using text drawn from the talk itself.)
Thanks to Rafael Dernbach for organizing and inviting me to this event, and thanks to Francesco de Biasi and Bernadette Klausberger for help with logistics and other support. And thanks to everyone for coming out tonight. I’m really excited to be here with you, especially during this twilight hour, in this in-between space, between day and night, like some hypnagogic state between waking existence and a sleep of dreams.
For over a century this liminal space of twilight has been central to thinking and theorizing the cinema and its shadowy realm of dreams, but I think it can be equally useful for thinking about the media transitions we are experiencing today towards what I and others have called “post-cinematic” media.
In the context of a film festival, the very occurrence of which testifies to the continued persistence and liveliness of cinema today, I should clarify that “post-cinema,” as I use the term, is not meant to suggest that cinema is over or dead. Far from it.
Rather, the “post” in post-cinema points to a kind of futurity that is being integrated into, while also transforming and pointing beyond, what we have traditionally known as the cinema.
That is, a shift is taking place from cinema’s traditional modes of recording and reproducing past events to a new mode of predicting, anticipating, and shaping mediated futures—something that we see in everything from autocorrect on our phones to the use of AI to generate trippy, hypnagogic spectacles.
Tonight, I hope to use this twilight time to prime us all for a long night of dreaming, and thinking, maybe even hallucinating, about the future of intelligence. The act of priming is an act that sets the stage and prepares for a future operation.
We prime water pumps, for example, removing air from the line to ensure adequate suction and thus delivery of water from the well. We also speak of priming engines, distributing oil throughout the system to avoid damage on initial startup. Interestingly, when we move from mechanical, hydraulic, and thermodynamic systems to cybernetic and more broadly informatic ones, this notion of priming tends to be replaced by the concept of “training,” as we say of AI models.
Large language models like ChatGPT are not primed but instead trained. The implication seems to be that (dumb) mechanical systems are merely primed, prepared, for operations that are guided or supervised by human users, while AI models need to be trained, perhaps even educated, for an operation that is largely autonomous and intelligent. But let’s not forget that artificial intelligence was something of a marketing term proposed in the 1950s (Dartmouth workshop 1956) as an alternative to, and in order to compete with, the dominance of cybernetics. Clearly, AI won that competition, and so while we still speak of computer engineers, we don’t speak of computer engines in need of priming, but AI models in need of training.
In the following, I want to take a step back from this language, and the way of thinking that it primes us for, because it encodes also a specific way of imagining the future—and the future of intelligence in particular—that I think is still up for grabs, suspended in a sort of liminal twilight state. My point is not that these technologies are neutral, or that they might turn out not to affect human intelligence and agency. Rather, I am confident in saying that the future of intelligence will be significantly different from intelligence’s past. There will be some sort of redistribution, at least, if not a major transformation, in the intellective powers that exist and are exercised in the world.
I am reminded of Plato’s Phaedrus, in which Socrates recounts the mythical origins of writing, and the debate that it engendered: would this new inscription technology extend human memory by externalizing it and making it durable, or would it endanger memory by the same mechanisms? If people could write things down, so the worry went, they wouldn’t need to remember them anymore, and the exercise of active, conscious memory would suffer as a result.
Certainly, the advent of writing was a watershed moment in the history of human intelligence, and perhaps the advent of AI will be regarded similarly. This remains to be seen. In any case, we see the same polarizing tendencies: some think that AI will radically expand our powers of intelligence, while others worry that it will displace or eclipse our powers of reason. So there is a similar ambivalence, but we shouldn’t overlook a major difference, which is one of temporality (and this brings us back to the question of post-cinema).
Plato’s question concerned memory and memorial technologies (which includes writing as well as, later, photography, phonography, and cinema), but if we ask the question of intelligence’s future today, it is complicated by the way that futurity itself is centrally at stake now: first by the predictive algorithms and future-oriented technologies of artificial intelligence, and second by the potential foreclosure of the future altogether via climate catastrophe, possible extinction, or worse—all of which is inextricably tied up with the technological developments that have led from hydraulic to thermodynamic to informatic systems. To ask about the future of intelligence is therefore to ask both about the futurity of intelligence as well as its environmentality—dimensions that I have sought to think together under the concept of post-cinema.
In my book Discorrelated Images, I assert that the nature of digital images does not correspond to the phenomenological assumptions on which classical film theory was built. While film theory is based on past film techniques that rely on human perception to relate frames across time, computer generated images use information to render images as moving themselves. Consequently, cinema studies and new media theory are no longer separable, and the aesthetic and epistemological consequences of shifts in technology must be accounted for in film theory and cinema studies more broadly as computer-generated images are now able to exceed our perceptual grasp. I introduce discorrelation as a conceptual tool for understanding not only the historical, but also the technological specificity, of how films are actively and affectively perceived as computer generated images. This is a kind of hyperinformatic cinema – with figures intended to overload and exceed our perceptual grasp, enabled by algorithmic processing. In the final chapter of the book, I consider how these computer-generated images have exceeded spectacle, and are arguably not for human perception at all, thus serving as harbingers of human extinction, and the end of the environment as defined by human habitation.
At least, that is what you will read about my book if you search for it on Google Books — above, I have only slightly modified and excerpted the summary included there. Note that this is not the summary provided by my publisher, even though that is what Google claims. I strongly suspect that a computer, and not a human, wrote this summary, as the text kind of makes sense and kind of doesn’t. I do indeed argue that computer-generated images exceed our perceptual grasp, that their real-time algorithmic rendering and futural or predictive dimensions put them, at least partially, outside of conscious awareness and turn them into potent vectors of subjectivation and environmental change. But I honestly don’t know what it means to say that “computer generated images use information to render images as moving themselves.” The repetition of the word images makes this sentence confusing, and the final words are ambiguous: are these supposed to be “self-moving images,” or images that, themselves, are moving? Or do the images use information to render themselves as moving images? What would that mean? The images are self-rendering? There is a multilayered problem of intelligibility involved, despite the fact that the sentences are more or less grammatical. The semantic ambiguities, the strange repetitions, and the feeling that something is just a little off are tell-tale signs of AI-generated text. This is not full-blown “hallucination,” as they say when AI just makes things up, but instead a kind of twilight recursion, suspended between the past of the training data and the future of the predictive algorithm, generating a sleepy, hypnagogic loop or a quasi-lucid, semi-waking dream.
But that summary was generated back in 2020. Since then, with GPT and other tools proliferating, we have witnessed a quantum leap in the intelligibility of AI-generated texts. In preparation for this event, I asked ChatGPT to summarize several of my books and to explain key concepts and arguments I made in them. The results were much better than what I just discussed (even though I was using the basic version that runs on GPT-3.5, not the more advanced GPT-4). Asked to explain my theory that “media are the originary correlators of experience,” the algorithm responded: “In this context, ‘originary’ suggests that media have been present from the beginning of human existence and have continuously evolved alongside our species. They are ingrained in our social and cultural development and have become integral to how we make sense of the world. […] Whether it’s language, art, writing, photography, film, or digital technology, each medium influences and organizes our experiences, constructing the framework through which we navigate reality.” That’s not bad, and it gets at what I’m calling the environmentality of media, including the medium or milieu of intelligence.
We could say, then, that artificial intelligence technology functions as a contemporary manifestation of the correlation between media and human experience. ChatGPT represents a significant leap in the relationship between humans and technology in the digital age. As a sophisticated language model, it mediates human interaction with information, communication, and even decision-making processes. ChatGPT is an intermediary that transforms the way we engage with knowledge and ideas, redefining the boundaries between human and machine. As an AI language model, ChatGPT embodies the fusion of the organic (human intelligence) and the artificial (machine intelligence). This fusion blurs the lines between human creativity and algorithmic generation, questioning traditional notions of authorship and creativity.
The only problem, though, is that everything I just said about ChatGPT was written by ChatGPT, which I asked to speculate, on the basis of my books, about what I would say about large language model AIs. The impersonation is competent, and even clarifying, as it brings out implications of my previous thinking in transferring them to the new case. Significantly, it points the way out of the impasse I described earlier with reference to Plato’s Phaedrus: AI will neither simply empower nor simply imperil human intelligence but will fundamentally alter it by transforming the parameters or environment of its operation.
The fact that ChatGPT could write this text, and that I could speak it aloud without any noticeable change in my voice, style, or even logical commitments, offers a perfect example of the aforementioned leap in the intelligibility of AI-generated contents. Intelligibility is of course not the same as intelligence, but neither is it easily separated from the latter. Nevertheless, or as a result, I want to suggest that perhaps the future of intelligence depends on the survival of unintelligibility. This can be taken in several ways. Generally, noise is a necessary condition, substrate, or environment for the construction of signals, messages, or meanings. Without the background of unintelligible noise, meaningful figures could hardly stand out as, well, meaningful. In the face of the increasingly pervasive—and increasingly intelligible—AI-generated text circulating on the Internet (and beyond), Matthew Kirschenbaum speaks of a coming Textpocalypse: “a tsunami of text swept into a self-perpetuating cataract of content that makes it functionally impossible to reliably communicate in any digital setting.” Kirschenbaum observes: “It is easy now to imagine a setup wherein machines could prompt other machines to put out text ad infinitum, flooding the internet with synthetic text devoid of human agency or intent: gray goo, but for the written word.”
Universal intelligibility, in effect, threatens intelligence, for if all text (or other media) becomes intelligible, how can we intelligently discriminate, and how can we cultivate intelligence? Cultivating intelligence, in such an environment, requires exposure to the unintelligible, that which resists intellective parsing: e.g. glitches, errors, and aesthetic deformations that both expose the computational infrastructures and emphasize our own situated, embodied processing. Such embodied processing precedes and resists capture by higher-order cognition. The body is not dumb; it has its own sort of intelligence, which is modified by way of interfacing with computation and its own sub-intellective processes. In this interface, a microtemporal collision takes place that, for better or for worse, transforms us and our powers of intelligence. If I emphasize the necessary role of unintelligibility, this is not (just) about protecting ourselves from being duped and dumbed by all-too-intelligible deepfakes or the textpocalypse, for example; it is also about recognizing and caring for the grounds of intelligence itself, both now and in the future.
And here is where art comes in. Some of the most intelligent contemporary AI-powered or algorithmic art actively resists easy and uncomplicated intelligibility, instead foregrounding unintelligibility as a necessary substrate or condition of possibility. Remix artist Mark Amerika’s playful/philosophical use of GPT for self-exploration (or “critique” in a quasi-Kantian sense) is a good example; in his book My Life as an Artifical Creative Intelligence, coauthored with GPT-2, and in the larger project of which it is a part, language operates beyond intention as the algorithm learns from the artist, and the artist from the algorithm, increasingly blurring the lines that nevertheless reveal themselves as seamful cracks in digital systems and human subjectivities alike. The self-deconstructive performance reveals the machinic substrate even of human meaning. In her forthcoming book Malicious Deceivers, theater and performance scholar Ioana Jucan offers another example, focusing on the question of intelligibility in Annie Dorsen’s algorithmic theater. For example, Dorsen’s play A Piece of Work (2013) uses Markov chains and other algorithms to perform real-time analyses of Shakespeare’s Hamlet and generate a new play, different in each performance, in which human and machinic actors interface on stage, often getting caught in unintelligible loops that disrupt conventions of theatrical and psychological/semantic coherence alike.
Moreover, a wide range of AI-generated visual art foregrounds embodied encounters that point to the limits of intellect as the ground of intelligence: as I have discussed in a recent essay in Outland magazine, artists like Refik Anadol channel the sublime as a pre- or post-intellecitve mode of aesthetic encounter with algorithms; Ian Cheng uses AI to create self-playing videogame scenarios that, because they offer not point of interface, leave the viewer feeling sidelined and disoriented; and Jon Rafman channels cringe and the uncomfortable underbellies of online life, using diffusion models like Midjourney or DALL-E 2 to illustrate weird copypasta tales from the Internet that point us toward a visual equivalent of the gray goo that Kirschenbaum identifies with the textpocalypse. These examples are wildly divergent in their aesthetic and political concerns, but they are all united, I contend, in a shared understanding of environmentality and noise as a condition of perceptual engagement; they offer important challenges to intelligibility that might help us to navigate the future of intelligence.
