Gen AI - Week 3 - Google Slides
How Do We Talk to AI?
This week we move on from the more technically intensive readings to think about the question: how do we talk to machines?
This (no less demanding) question is, on the one hand, the topic of a thousand social media posts promising to teach you everything you need to know about the new (and dubiously named) discipline of “prompt engineering”. On the other, it also opens up a more profound space for exploring what it means to talk at all – with machines, with each other, with ourselves.
Accordingly I’ve selected an array of readings this week that aim to stretch our thinking about “talk”, as a performance (Austin 1962), as a defiant act of truth (Foucault 2020), and as user interaction (Weizenbaum 1966, Mollick & Mollick 2023, Magee et al. 2024).
I want first to reflect a little on J.L. Austin’s seminal text, “How to Do Things With Words”. This book – a series of lectures – instigated attention to the “pragmatics” of language, and inspired, among others, Searle’s Chinese room thought experiment. Again I don’t suggest reading this end-to-end, but perhaps focus on getting some conceptual understanding of Austin’s distinctions between:
- performative / constative utterances
- locutionary / illocutionary / perlocutionary speech acts
Perhaps the question for us is “How to Get AI to Do Things with Words”.
Then, with Ethan and Lilach Mollick’s paper, we will examine how different prompts can be used to play roles on the pedagogical “stage”. I’ve also assigned a paper I have developed recently with colleagues titled “The Drama Machine”. It is long, and I don’t suggest you read all of it, but rather focus on the principal idea: that LLMs have internalized vast amounts of information about how people perform with language, and this internalization can simulate quite complex character enactments.
Both papers are in some ways throwbacks to the early days of AI, when psychoanalysis and dramaturgy were more prominent theoretical rubrics. As if on cue, we look at a classic AI paper, Weizenbaum’s 1966 presentation of ELIZA: a system that simulated a certain therapeutic approach, to surprising effect given the technical limits of the time.
Finally, for those with theoretical inclinations, I’ve included Michel Foucault’s work on “Discourse and Truth and Parrhesia”. Perhaps more of a placeholder for discussions in future weeks, Foucault’s analysis of parrhesia in classical (Greek and Roman) discourse holds special relevance for how we view AI’s social role: as truth-teller, as performer, as stochastic parrot, as assistant, and so on.
I realise several of these texts also skate across traditions and language not all of you may be familiar with, and I’d encourage people supplement their readings with queries to AI itself. For example:
- Tell me about Austin’s different kinds of speech acts, and how
these relate to LLMs.
- How could Foucault’s discussion of parrhesia relate to contemporary
discussions about AI?
For those wanting a (slightly surreal) rendition of “The Drama Machine” paper rendered by AI, quite dramatically:
https://notebooklm.google.com/notebook/60047a4e-2674-4d6c-b896-67e5f3166580/audio
I’d of course encourage people who enjoy podcast enactments to explore other texts – as some of you have been doing.
Readings:
Austin, J. L. (1975). How to Do Things with Words. Harvard University Press. https://archive.org/embed/HowToDoThingsWithWordsAUSTIN
Foucault, M. (2020). “Discourse and Truth” and “Parresia”. University of Chicago Press. https://i-share-uiu.primo.exlibrisgroup.com/permalink/01CARLI_UIU/q00vor/cdi_gale_lrcgauss_A609143295
Magee, L., Arora, V., Gollings, G., & Lam-Saw, N. (2024). The Drama Machine: Simulating Character Development with LLM Agents. arXiv preprint arXiv:2408.01725. https://arxiv.org/abs/2408.01725
Notebook version: https://notebooklm.google.com/notebook/60047a4e-2674-4d6c-b896-67e5f3166580/audio
Mollick, E. R., & Mollick, L. (2023). Assigning AI: Seven Approaches for Students, with Prompts. SSRN. http://dx.doi.org/10.2139/ssrn.4475995
Weizenbaum, J. (1966). ELIZA – A Computer Program for the Study of Natural Language Communication Between Man and Machine. Communications of the ACM, 9(1), 36-45. https://dl.acm.org/doi/10.1145/365153.365168
Assignment(s):