Annotated · critical & intersectional · project and mechanism primers
The camp's tools show you the mechanism. These readings ask what the mechanism is doing in the world — who builds it, whose labor and data it runs on, who pays when it fails, and what it does to how we see each other. A compact primer section follows the camp arc from token embeddings to shared image/text representations to point tracking across video.
The paper behind the camp's central move: a language model is a fluent predictor, not an understander — and scale has costs in data, labor, and environment. Pairs directly with the Prediction ≠ understanding bridge and the Tokenizer.
The best single metaphor for what Session 1 demonstrates: lossy compression of the training text, returned with confident fluency. Short enough to assign between Saturdays.
Written by ELIZA's own creator after watching people confide in his pattern-matcher. The original argument that the question is not what computers can do but what we ought to delegate to them. The camp's ELIZA Simulator is this book made playable.
Use it in Session 1 to ask why words are nearby: shared meaning, similar contexts, or a learned cultural association? Treat the projection as a simplified embedding snapshot; modern transformer token representations also change with context.
A direct primer for Session 2’s three-caption question: images and text are mapped into representations that can be compared. This explains one influential form of prompt-image conditioning without implying image search, human understanding, or that every generator uses CLIP.
The clean Session 3 bridge from PoseNet: select a point in one frame, then estimate its correspondence through later frames. Tracking analyzes video that already exists; a generator faces the harder inverse problem of creating later frames while maintaining those relations.
How "neutral" ranking and retrieval encode racism and sexism — the search-engine ancestor of every default the camp's Default Test surfaces.
Names the "New Jim Code": discrimination that hides inside technical neutrality. The sharpest companion to the Default is a design decision bridge.
AI as an extractive industry — lithium, labor, data, classification power. The wide-angle lens behind the camp's Access Tiers conversation.
The pre-generative classic on opaque models scoring teachers, loans, and sentences. Useful precisely because none of it is about chatbots — the pattern precedes the hype.
Three case studies of automated decision systems aimed at poor families. For the "what does the human decide next?" half of the camp's loop, this is the stakes.
The invisible human labor — labeling, filtering, moderating — inside every "automated" system. Answers the cohort's labor questions with reporting, not vibes.
Reported from inside the frontier labs and from the data-labor economies they depend on. The most current map of who profits and who is extracted from.
How the default-female, deferential voice assistant got that way — the report that named the pattern and the title comes from Siri's old reply to abuse. For the cohort's questions about gender and how models are personified; pairs with Klara and the Sun and the ELIZA Simulator.
For the opposed-but-curious pathway specifically: an argument that refusal and restructuring are legitimate technical positions, not failures to adapt.
The shortest route into the human-built lexical layer: word senses are grouped into synsets and connected by semantic relationships. Read it before asking whether a category is “inside” an image—it is not.
The original account of building an image database on WordNet’s hierarchy. Pair its scale and engineering ambition with the Session 2 questions: who selected the categories, gathered and annotated the images, defined correctness, and was represented or excluded?
A readable explanation of PoseNet’s input and output: an RGB image becomes 17 estimated body keypoints, x/y positions, and confidence scores. It is a recognition/inference example—not identity recognition and not image generation—and it creates the bridge from one frame to motion across time.
What's actually inside image training sets — the categories, the politics, the people photographed without consent. Read it the week of Session 2, after the Dataset Balance Simulator.
Not about AI at all — about how images carry assumptions about gender, class, and power. Fifty years later it reads like a manual for interrogating image-model defaults.
From the Gender Shades audits to the "coded gaze": what face-analysis systems fail to see, and what it took to make the failures count. A model-behavior investigation, book-length.
The imitation game, in Turing's own surprisingly readable prose — including the objections he anticipated. Better than every summary of it.
An artificial friend narrates what she sees — and what she can't. The most precise fiction we have about machine perception, personification, and the gendered warmth we project onto assistants. Pairs with ELIZA and What does the machine see?
What would it actually take to raise a digital being — in time, care, and labor? The slow, unglamorous counter-story to instant intelligence.
Murderbot would rather watch its shows than be personified, thank you. A short, funny on-ramp to personhood questions — and an easy recommendation for students.
A century of attempts to automate teaching, and why they keep promising the same things. Inoculation for the next edtech pitch you sit through.
Why bias in technical systems is structural, not a bug to patch — written to be teachable, with classroom-ready examples across race, gender, and ability.