The Word in the Machine: an Applied Investigation of LLMs

The Word in the Machine: an Applied Investigation of LLMs through the Lens of Philosophy of Language

Convenor

Alessandro Trevisan

Alessandro is a 3rd-year PhD student at the University of Cambridge, affiliated with Cambridge Digital Humanities and the Faculty of English. His research, supervised by Dr Leonardo Impett, focuses on the study of Large Language Models (LLMs) through the lens of philosophy of language and literary theory. In particular, Alessandro uses the ‘later’ Wittgenstein’s philosophical method to explore how our concepts of language, text, reading, and writing relate to the operations of LLMs.

Alessandro came to Cambridge to read for an MPhil in Digital Humanities, developing an interest in the narrative forms and tropes deployed by LLMs in their outputs. Before that, he read for a BA in English at the University of Bristol, where he acquired a theoretical grounding in the critical framings he is now applying to the analysis of LLMs. His research is funded by a St John’s College ‘Benefactor’s Scholarship’.

Description

In this workshop, Alessandro aims to offer a semantic analysis of large language models (LLMs) informed by philosophy of language and literary theory. Central to this approach is Ludwig Wittgenstein’s notion of ‘form of life’, which holds that meaningful uses of language depend on a shared socio-cultural and physiological framework between interlocutors. This raises a provocation for collective reflection with workshop participants: how are we to make sense of language produced by machines that do not inhabit our form of life? To address this question, Alessandro will first guide participants in exploring what a string of text means for an LLM by implementing the ‘fill-mask’ task — an empirically robust, playful method for eliciting a model’s most likely substitutions for a masked ‘token’ in a prompt. He will use this exercise to propose a new theory of meaning for LLMs, in which a word’s meaning is inextricable from the specific context in which it is embedded and must be interpreted in relation to the linguistic choices it prompts the model to make at inference. The workshop will thus explore how to interpret the probability distributions over possible outputs — whether single words or longer strings — generated by an LLM: the totality of candidate outputs for a prompt, Alessandro suggests, must be read compositionally, as forming a picture of a social context the model simulates. He shall conclude the workshop by situating this mode of reading within Wittgenstein’s ‘picture theory of language’ and formalist literary theory, drawing chiefly from I. A. Richards’ work and his concept of ‘feedforward’.

This workshop is part of our Methods Fellowship programme, which develops and delivers innovative teaching in digital methods. You can read more about the programme here and view the complete series of workshops here.

Target Audience

Our CDH Methods workshops have limited places and are prioritised for students and staff at the University of Cambridge. However, if space is available, we welcome all participants who want to learn and apply digital methods and use digital tools in their research.

This session may be of particular interest to:

• PhD students in the Arts, Humanities and Social Sciences
• Early Career Researchers in the Arts, Humanities and Social Sciences

Contact CDH

If you have specific accessibility needs for this event, please get in touch. We will do our best to accommodate any requests.