Cosine capital: Large language models and the embedding of all things

This article describes the emergence of a novel form of capital—which I call “cosine capital”—that finds objectified form in the “embedding” structures of large language models. In the past decade, massive neural network architectures transformed computational approaches to language in the form of large language models. This approach to modeling language is now being adapted to nearly any sequential data structure imaginable in both academia and industry. While these technologies have been hailed as revolutionary, I situate them within a continuous technological and philosophical lineage that runs directly back to the origins of cybernetics and information science, in particular Claude Shannon’s noisy channel model of communication. I imagine this noisy channel as a sort of “diagram of power,” arguing that a similar process of “enclosure” that commodified the bit as the foundational unit of information is now taking place with embeddings, objectifying them as fungible commodities across an increasing range of societal domains. I compare this cosine capital to Fourcade and Healy’s recent notion of “eigencapital,” suggesting that the particular technical features of embeddings—specifically, their inherently relational nature—challenge the eigencapital model and instead represent a fundamentally novel form of abstraction with strong implications for the future of capitalism and technology.