1. The Horizon
“It is tempting to call large language models, and ChatGPT in particular, hyperagents. This is not to say that such models are excessively agentive or so large and powerful as to be mysterious and unknowable. Nor is it just another way to fixate on their size. It is to stress that, whatever their actual capacities, they generate far more hype than other agents.
Perhaps only Jesus, Barbie, Obama, and T-Rex come close.”
The media has devoted article after article to large language models, such as OpenAI’s ChatGPT, and the incredibly realistic, often eerie, and sometimes horrific conversations they can generate. A study claims that GPT-4, the “generative pretrained transformer” underlying the latest version of ChatGPT, can already pass its freshman year at Harvard. Another predicts that 340 million full-time jobs will be lost to artificial intelligence—based on the ability of generative AI to create content that is indistinguishable from human work. One article informs us that recent advances in artificial intelligence have made noninvasive mind-reading possible. Another instructs us how to turn our chatbot into a life coach. We learn that tech leaders have signed a letter calling for a “pause” in AI research and the creation of even larger language models because they fear the long-term negative effects. Only a few weeks later, the AI boom is generating optimism in the tech sector as “stocks soar.”
And all that is already yesterday’s news.
If language has long been an emblem of the human species, talking machines seem to be harbingers of some kind of technological singularity. Indeed, if the brilliance—or at least eloquence—of large language models is any indication, we seem to be poised at the threshold of general AI, a form of artificial intelligence that will not only meet, if not surpass, human intelligence but also maybe even replace humans altogether.
Messiah for some, Armageddon for others.
It is not just the recent acceleration of the verbal powers of such agents that is staggering, it is also their scale. They can contain trillions of parameters, require months of training, and have the entire corpus of the written word at their disposal. It is therefore tempting to call large language models, and ChatGPT in particular, hyperagents. This is not to say that such models are excessively agentive or so large and powerful as to be mysterious and unknowable. Nor is it just another way to fixate on their size. It is to stress that, whatever their actual capacities, they generate far more hype than other agents.
Perhaps only Jesus, Barbie, Obama, and T-Rex come close.
What follows is a spirited attempt to cram large language models into a relatively small text. I focus on the semiotic processes—or meaningful practices—that mediate the emergent relations among three kinds of actors: human agents, like you and me; machinic agents, such as ChatGPT; and corporate agents, be they states, tech companies, or institutions more generally. I pay particular attention to the coupling, and hence co-mediation, of the guiding principles of such agents. I thereby offer a critical genealogy of the highly contested relations among human values, machinic parameters, and corporate powers. My goal is not so much to see past our current social and technological horizon as to offer a theory of the reasons for and effects of the horizon itself.
Chapter 2, “Human Semiosis,” offers an account of meaning that can span the distance between human interpretation and machinic calculation. Chapter 3, “Machine Semiosis,” is a gentle introduction to the inner workings of large language models. Chapter 4, “Pretraining and Fine-Tuning,” zooms in even closer, focusing on two key processes underlying the machinic mediation of meaning. Chapter 5, “Labor and Discipline,” argues that language models are the objects and agents of disciplinary regimes and constitute a novel—and problematic—mode of production. Chapter 6, “Parrot Power,” discusses the various kinds of generativity that underlie language models and shows their relation to linguistic competence, labor power, and corporate profits. Chapter 7, “Language Without Mind or World,” discusses the limitations of large language models and users’ limited awareness of those limits. Chapter 8, “Metasemiosis and Monsters,” analyzes the mediatization of human-machine interaction and shows its relation to a range of enclosures, horizons, and scales. Chapter 9, “On Interpretation,” takes up the distinction between humanistic and machinic interpretation and discusses strategies for interpreting texts that were generated by machines. Finally, chapter 10, “The Problem with Alignment,” discusses the promises and pitfalls of aligning machinic parameters with human values, as well as the necessity of dealigning them with corporate interests.
Insofar as large language models, and machinic intelligence more generally, are a central target of speculative capital, they constitute a fast-moving topic. So rather than fetishize bleeding-edge developments, which usually only adds to the hype, I focus on bread-and-butter issues—especially in the lead up to ChatGPT. I write for a broad audience, and so for people willing to learn a little bit of math and work their way through a limited amount of formalism for the sake of a deeper understanding. I argue, implicitly, that critical theorists need to have a detailed understanding of the media they are analyzing—or else they are the equivalent of cavemen critiquing calculators. And I relate machine learning, language models, and natural language processing to meaning, and the great (post)humanist interpretive tradition, with a particular focus on ideas coming out of anthropology, critical theory, and pragmatism.
Given the stakes, as well as the hype, I will begin cautiously and work my way toward the horizon slowly.
