WEEK 6 LECTURE SUMMARY - Generative AI Foundations, Environmental Impact, and Local Workflows

Formative Course Feedback

  • A formative survey is available on the Stream site under the "please help" section in Week 6.
  • Provides mid-semester feedback regarding learning impact, enjoyment, assessment clarity, and course experience prior to final summative evaluations.

Historical Context and Generative AI

  • Historical technology disruptions:
    • 1440s: Gutenberg's movable type shifted media production from hand scribes to mass printing and distribution.
    • Industrial Revolution: Power looms mechanized textile production and restructured factory labor.
    • 1990s: The World Wide Web digitized and globalized information access.
    • 2020s: Generative AI introduced automated cognitive synthesis.
  • Technology adoption speed: The Internet required 23 years23\,\text{years} to reach widespread uptake, whereas generative AI reached equivalent adoption in 3 years3\,\text{years}.
  • Generation AI: Cohorts born starting in 2025 inherit generative AI as integrated infrastructure.
  • Historical Luddite parallel: Skilled textile artisans protested wage suppression and craft descaling caused by mechanized looms—a historic precedent for evaluating human labor and craft against synthetic tools.

Environmental Footprint of Generative AI

  • Energy Metrics:
    • Single AI image generation: 0.6 Wh0.6\,\text{Wh}.
    • One hour of video streaming (81 Wh81\,\text{Wh}): Equivalent to 130 AI images.
    • One hour of PC gaming (400 Wh400\,\text{Wh}): Equivalent to 650 AI images.
    • One passenger flight (>1000 km>1000\,\text{km}): Consumes 300 kWh300\,\text{kWh}, equivalent to approximately 500,000 AI images.
    • Boiling 1 L1\,\text{L} jug of water (2–3 minutes2\text{--}3\,\text{minutes}): 100–120 Wh100\text{--}120\,\text{Wh}.
    • Training a flagship LLM (e.g., GPT-4): Requires 90–100 days in a data center using 50–62 GWh50\text{--}62\,\text{GWh} (equivalent to power for 5,000 US homes for 1 year).
  • Water Footprint Metrics:
    • Data center cooling and grid evaporation: Consumes 3 L3\,\text{L} of water per \t\text{kWh} (or kWh\text{kWh}).
    • Single AI image generation: Consumes 2 mL2\,\text{mL} of water.
    • Single beef hamburger patty: Embeds 240 L240\,\text{L} of virtual water (equivalent to over 1,000,000 AI images).
  • Generative Video Scaling:
    • Generative video requires frame-by-frame image blending (e.g., 25 frames per second).
    • One second of AI video draws 0.7–77 Wh0.7\text{--}77\,\text{Wh} of energy and up to 230 mL230\,\text{mL} of water.

Intellectual Property, LoRa Models, and Ethics

  • Black Box Models vs. LoRa Fine-Tuning:
    • Base black box models train on broad, unverified internet data.
    • Low-Rank Adaptation (LoRa) adapters are compact fine-tuned models (~100 MB100\,\text{MB}) trained on 15–30 custom artworks to capture specific visual styles locally.
  • Copyright & IP Rules:
    • Text prompts entered into closed black box models yield outputs without enforceable copyright.
    • Combining custom dataset creation, LoRa fine-tuning, manual arrangement, and post-editing creates legally protectable intellectual property.
  • Alignment & Ethics Mandate:
    • Alignment incorporates human taste, ethics, and cultural context into models. Unaligned base models reflect dominant web demographics while underrepresenting minority voices.
    • Core practice principles: Prioritize intentionality over raw automation, maintain material awareness of energy/water costs, and preserve craft, attribution, and data consent.

Tool Ecosystem and Local Execution

  • Closed-Source Black Box Ecosystem:
    • Text/Image/Layout: ChatGPT, DALL-E, Midjourney, Adobe Firefly, Canva Magic Studio.
    • Video Generation: Runway, Pika, Luma Dream Machine, Sora, Google Veo 3.
    • Audio/Voice: Suno, Udio, Stable Audio, ElevenLabs, Resemble AI.
    • Limitations: High subscription costs, lack of training transparency, privacy issues, and continuous feeding of external algorithms.
  • Open-Source Local Execution:
    • ComfyUI: Modular open-source software running locally on Apple M-chip or Windows hardware.
    • Benefits: Eliminates cloud compute dependency, keeps data secure, enables local fine-tuning, and logs generation seeds for workflow tracking and reproduction.

Practical Applications in Creative Practice

  • Style Transfer: Training custom LoRa models locally to apply personal visual styles across new image generations offline.
  • Computational 3D Modeling: Writing Python scripts via AI assistance in Blender to model surface curves matching light waveforms, exporting STL files for 3D ceramic printing to physically split the light spectrum.
  • Collaborative Tools: Building RGB color mixing interfaces that output programmatic spreadsheets for remote lighting setups (e.g., installations in UAE).
  • Hybrid Workflows: Combining physical photography, generative fill in Photoshop, and spatial lab software to create large-scale environments.
  • Custom Technical GPTs: Training specialized assistant models on software technical documentation (e.g., Vectorworks) for context-specific design troubleshooting.

Questions & Discussion

  • Prompt on Internet Connection Sound: Identified as dial-up modem audio representing early data transfer noise.
  • Prompt on Hardware Costs: Discussion confirming recent market price increases in RAM and memory hardware driven by enterprise AI compute demand.
  • Clarification on Luddites: Confirmation that historical Luddites were skilled artisans opposing wage suppression and craft descaling from mechanized factory looms rather than opponents of technology itself.