Languaging and large language models: a message from radical linguistics

Abstract The paper takes a radical linguistic perspective on languaging and LLMs. Just as people draw on languaging alone and together, we make/construe aggregates of punctuation, font, alphanumerics, etc. In the paper, first, I use case studies to track human use of aggregated patternings (APs) based on careful coordinating, stance-taking, attending and use of judgements. Later, I contrast this with how LLMs generate APs that set off epistemic and communicative (“epicommunicative”) effects. As LLMs map AP input onto vectors in a high-dimensional space, “training” sets off derived statistical mappings. Although using data (not “language"), the resulting AP fields prime human feelings/claims/belief/expertise (“seemings”). The face value of LLM output enables immediate, constructive and emotive responding. Given the “normalcy” of an LLM’s statistically generated output, I hypothesize that human brains must also use something like vectorial weightings. Indeed, without a parallel means, machine-generated AP fields would be less effective or beguiling. Hence, the radical linguist’s message is that the epicommunicative, or familiar, draws on the normalcy of (artificial) “text.” We use, not “language models”, but structured information that affects human expectations, consistency, habits and judging. Rather than focus on novelty, I stress that speaking or writing draw on predicative acts which co-occur with predicative renderings, orinformed responding based on a life history. In this discussion of skilled linguistic action (Cowley, Lang Sci 84, 2021), I extend Wittgenstein's famous description of rendering a text aloud. Accordingly, I stress how affect, practices and languaging (i.e. attending, stance-taking and rendering explicit) enable pre-conceptual feelings, belief, expertise etc. As a result, priming sets off epicommunicative effects to which, crucially, human judgements can add value (or meaning).