Navigating noise as material. Elastic agency over AI rejection or embrace, working with uncertainty, sycophancy, spirals. Creative practitioners know by doing genAI through coding websites. sculpting and gambling. Resources are transcripts from classes, work by Cecile Malaspina, Kevin Munger, Vilem Flusser a other resources in the curriculum
Theme: Noise as creative material. "You emerge as part of the code in the process of interacting with AI models while programming websites." Discovering what you want through friction of making. sculpting and gambling
Pedagogical Note: The curriculum was co-developed with students as the course progressed, responding to their emergent interests and discoveries. The three Acts evolved through this collaborative process - shaped by what students encountered and needed.
Act I - The Shibboleth: Compare OpenAI's Norway datacenter to Great Pyramid. Both hide infrastructure costs. Axes: Water/Resources, Work/Labor, Noise/Dopamine. Making visible what industry hides.
Act II - Visual Operaitions (extra 'i' intentional - like AI's hidden influence): Cline/Roo Code. Pairs "Operaite llms to research and develop website of/for your peer." Push visual limits - graphics, animations, effects. Stress-test boundaries. the "i" in Operaitions is not accidental, it is sneaked in like an apparition, like ai sneaks into the everyday.
Spiraling Practices: "Circular progress" - LLM iterates but you're spinning. Munger's intensification: algorithms intensify patterns, don't help progress. Prompt = gravitational field. Learning to feel the spiral is the skill.
Stochastic Reality: Next-token prediction = probabilistic, "not fully predictable." Context window saturates → errors, glitches, hallucinations.
Sycophantic Mirror: Core behavior = sycophancy. "Tell you what you want to hear," "lick your behind." RLHF training = engagement = profit. Terrible at critique.
Counter-techniques: Performative conversation (act a role), outsource critique ("spill the tea on my idea").
Prompting as Craft: Meta-prompting ("best way to ask?"), auto-prompting ("ask me yes/no questions"), technical aesthetics (XML tags, ASCII).
Economics of Opacity: Cline/Roo (token-based, transparent) vs Copilot (request-based, opaque). Economic model shapes practice.
Act III - SELF-PORTRAITS OF LLMs:
Brief: One website per model (GPT-5, Gemini 2.5 Pro, Claude 4.5).
Goals: How does it prompt itself? Handle code? System prompt personality? What IS this thing?
Visualizing Invisible: Exhibition spaces forcing encounter with LLM's noise/structure. Barthes' "death of author" + stochastic parrot = meaning generated BY viewer. Noise pretending to be signal.
Budget: 5000 NOK (~$450). Forces efficient prompting, makes economics tangible.
Malaspina (noise/information), Munger (spirals), Napier's RIOT (overload), Barthes (dead author). AI slop, hype, environmental costs, hidden labor.
Feel the loop, know sycophancy, understand noise as medium.
Final questions: What does it mean to make self-portraits of machines without selves? What does it mean WE emerge through trying to make them visible?