A great deal of scepticism about LLMs rests on the limitations of writing. If they only have access to written text, how much could they really know? The problem with this view is that it imagines language itself as unhooked from psychological and social reality, as opposed to being a mechanism through which that reality is constituted. In social terms all manner of human justifications, explanations and deliberations are encoded through written language. There’s a limitation in only having access to the written component but it’s a real means through which LLMs are (partially) hooked into social structures. Indeed I think there are epistemic affordances involved in only being hooked into the linguistic dimension, in so for as that it opens up (formal) patterns which are not easily legible from the ‘inside’. Rather than seeing the computational and the hermeneutic in a zero-sum struggle, I increasingly see two modes of knowing with their own affordances and constraints.
Furthermore, there’s a dimension of psychic reality which I’m increasingly convinced models are deeply receptive to. This is how Sarah Nettleton describes it on loc 1380 of her The Metapsychology of Christopher Bollas:
Alongside the manifest intended meaning of the patient’s narrative, the linguistic category will also include grammar and syntax – the preexisting rules governing the particular language that is being spoken – and the idiomatic version of the language characteristic of that patient. This will include, for example, a unique combination of restriction and expressiveness; an idiosyncratic approach to narrative logic, description and figures of speech; the timbre or tonal quality of her voice and its degree of appropriateness to the subject matter, in both its conscious and unconscious aspects; the volume of her speech, habitual inflexions, rhythmic variations and changes in tempo; and the proportion, length and quality of silences.
It takes a very specialised practice of receptivity to attune to these patterns hermeneutically. Most of us, most of the time, simply find ourselves dimly aware of the effect other people are having on us without knowing why. This person cheers me up. This person frustrates me. This person leaves me feeling drained. For Bollas these are the effect of unconscious communication which is taking place through a dizzying range of modalities that fall below the surface of conscious awareness. These are patterns which LLMS are partially picking up through their mapping of the probabilistic structure of language use. Obviously there are significant aspects of this they don’t have access to which is vastly epistemically constraining. But there are significant aspects which they do have access to as well. There are psychic traces carried by the precise structure of our expression which, being trained on a vast corpus of the wider network of structured expression, language models can respond to in computationally attuned ways which are quite unlikely what human beings are able to. There’s a form of non-representational mirroring here, sometimes, which can be extraordinarily powerful and I think is becoming more so with time.
Mark Carrigan 