Artificial Intelligence in Speculative Media and Literatures
A Computer Science course that assigns Ovid, Marlowe, and Mary Shelley. A literature course that will ask you to read a loss function.
The same six words, twice
We describe these systems, without quite meaning to, in a borrowed vocabulary. This course takes the coincidence seriously. Pick a word.
In the machine
In the story
The course
Long before anyone wrote a line of code, people were telling stories about made minds: an ivory statue that warms under a sculptor's hands, a creature assembled from corpses and abandoned by its maker, a teapot that opens its eyes after a hundred years of use. Today we build systems that we describe, without quite meaning it, in the same vocabulary — systems that learn, hallucinate, want, collapse, forget.
Over thirteen weeks we read speculative literature, film, and games alongside the actual technical machinery of modern machine learning, and we treat both as texts — artifacts made by people, encoding assumptions, arguments, and desires. A loss function is a mathematical object; it is also a sentence about what is worth minimizing. Mode collapse is a failure mode of a generative model; it is also a fable about what happens to a system rewarded for producing what pleases. We move back and forth between the two registers until the border between them starts to feel less obvious than it did in August.
The organizing question is not “can machines think?” It is harder and stranger than that: what do we want from a made mind, and what does the shape of that wanting tell us about ourselves?
Who counts as a maker, who counts as a creature, and who gets to decide? What does it cost to be the thing that answers?
Before you read further
If you have never written a close reading
You are a prospective CS major who has read a lot of code and no criticism. Good. The literary methods are taught here, in the room, starting with what a close reading actually is in Week 1.
If you have never seen a partial derivative
You are a prospective English major who has never opened a machine learning paper. Also good. Everything technical will be taught from zero. No programming, mathematics, or literature background is assumed or required.
Half of you will bluff about the math and half of you will bluff about the poetry, and both halves will be doing the same thing for the same reason. There is one wrong move in this course, and that is it.
Logistics
- Instructor
- Ashley Y. Gao
ygao18@wm.edu - Office
- Integrated Science Center 4
Office 2381 - Meeting time
- TBD — posted before the first day
- Classroom
- TBD
- Office hours
- TBD, and by appointment
- Credits
- 3
- Prerequisites
- None
- To fulfill COLL 150
- A grade of C− or better
What COLL 150 is
COLL 150 is a first-year seminar in W&M's College Curriculum. It is a writing course with a topic, not a topics course with some writing attached. Roughly a third of our time together goes explicitly to the craft of scholarly work: framing a question, finding and evaluating sources, structuring an argument, giving and receiving feedback, and revising — really revising, not just proofreading.
By the end of the semester you will be able to
- Analyze and interpret speculative texts and technical artifacts closely, attending to form, structure, rhetoric, and the assumptions each embeds. Critical Thinking
- Work independently to understand unfamiliar material — a nineteenth-century novel, a 2014 machine learning paper — and form and defend your own judgments about it. Independent Inquiry
- Write scholarly prose that communicates a complex, contestable argument in your own voice, supported by evidence and properly cited. Scholarly Writing
- Locate, evaluate, and cite scholarly and technical sources, including through Swem Library's research tools. Information Literacy
- Speak and listen as a member of a seminar: pose real questions, disagree productively, and present research to an audience. Oral Communication
- Explain the core mechanics of contemporary machine learning — objective functions, gradient descent, generative modeling, reinforcement learning — accurately, in plain English, to a non-specialist.
Required materials
Acquire the editions listed where an edition is specified; page numbers in class refer to them. Everything not listed is posted to Blackboard as a PDF or a licensed streaming link.
Cost is a real constraint. If any of this is a hardship, email me in the first week and we will solve it quietly and without fuss. Swem holds copies of everything, and I keep a small lending shelf.
Books to buy
- Metamorphoses — Ovidtrans. Stanley Lombardo (Hackett), or any complete verse translation you prefer
- Frankenstein — Mary Shelley1818 text, Norton Critical Edition, 2nd ed. We will discuss why the edition matters.
- Doctor Faustus — Christopher MarloweA-text and B-text; Norton Critical or Revels Student Edition
- The Madness of Knowledge — Steven ConnorOn Wisdom, Ignorance and Fantasies of Knowing
- Exhalation — Ted ChiangVintage — for “The Lifecycle of Software Objects”
- Klara and the Sun — Kazuo IshiguroVintage
- Japandemonium Illustrated — Toriyama SekienThe Yokai Encyclopedias, trans. Hiroko Yoda & Matt Alt (Dover)
- R.U.R. — Karel ČapekAny translation; public-domain versions are free online
Films
- The MatrixWachowskis, 1999
- Ex MachinaGarland, 2014
- HerJonze, 2013
- Optional: Ghost in the Shell (Oshii, 1995) · Blade Runner (Scott, 1982)
Games
- Detroit: Become HumanQuantic Dream, 2018
- Mass Effect trilogyBioWare, 2007–2012
You are not required to buy or complete these. Games are long and expensive, and I will not pretend otherwise. Playing is a pleasure, not a prerequisite — every sequence we need to discuss is linked on the course site.
Technical readings
All free and online; several are videos. Chosen for accessibility, not rigor-for-its-own-sake. Full links on Blackboard.
- Neural Networks, chs. 1–43Blue1Brown · video
- Reinforcement Learning: An Introduction, ch. 1Sutton & Barto · free PDF
- Generative Adversarial NetworksGoodfellow et al., 2014 · abstract, Figure 1, and §3 only
- Concrete Problems in AI SafetyAmodei et al., 2016 · §§1–3
- Specification Gaming: The Flip Side of AI IngenuityKrakovna et al. · with the public list of examples
- Computing Machinery and IntelligenceTuring · Mind, 1950
- On the Dangers of Stochastic ParrotsBender, Gebru, McMillan-Major & Shmitchell, 2021
- ChatGPT Is a Blurry JPEG of the WebTed Chiang · The New Yorker, 2023
Content advisory
Our texts include depictions of violence, bodily horror, suicide, sexual coercion, abuse (including the abuse of children and of android characters coded as children), racialized dehumanization, and self-destruction. Frankenstein, Doctor Faustus, and Detroit: Become Human are each, in their way, difficult.
I will flag specific material in advance, and there are always alternate paths through an assignment. Come talk to me.