Brief notes

  • Farewell to Fart: Janelle Shane
    • char-rnn makes cute recipes, obviously wrong; larger pre-trained models make more plausible recipes
    • knitting patterns are much more susceptible to minor errors which can just eat up yarn (!)
    • if you start with crochet patterns, and use GPT-2 to generate arbitrary text, it can sound realistic, and will steer any conversation onto hats lol
    • GPT-3: no longer amusingly weird from scratch
    • folk-rnn makes ∞ Irish folk tunes, but people aren’t interested in playing them (AI to fuel buckets of unremarkable content does not respect the consumer)
    • Kate Compton: Opulent AI, AI that calls attention to its own artifice
    • use GPT-3 to complete a prompt about training a neural network to generate costumes (!)
    • horse facts can be adversarial if they’re too basic!
      • Q: How many giraffes are in the average living room? A: Two, but they won’t talk to each other!
    • text-generating algorithms are getting better at sounding cliché
  • Artificial biodiversity: Sofia Crespo
    • cf artbreeder
    • where is the beauty in a dataset? images? pixels? computational training? the NN itself?
    • “Isn’t all art made by humans an execution of reshaping of data processed by neural networks?”
    • Visual Indeterminacy in Generative Neural Art
    • Codex Seraphinianus: showcase of life, invented by an artist
    • Anna Atkins: create an impression of life itself
  • Harms from AI research — workers very susceptible to wage theft, and wages are very low (MTurk can be a race to the bottom)
  • How should researchers engage with controversion applications of AI?
    • math can be weaponized (facial recognition blocking entry to home, etc)
    • “I’m not ready to always have an alternative ready just because you’re not prepared to engage with critiques of your work” – Tawana Petty
    • “This is not a mathematical problem, so this will not have a mathematical solution, and we cannot offer more sophisticated math instead of engaging as activists”
  • Panel, anticipating / mitigating risks, both of participants and products
    • social impact: if you’re not willing to engage with communities, the risks are totally different (firsthand risk: finder/publisher controls the narrative!)
    • that the AI community has a problem with thinking about the ethical implications is very worrying
      • you never see “nursing for good” or “food for good” as often as you see “AI for good”
      • the smarter you are, the better you can justify any random decisions
    • how do we change incentives?
    • risk pyramids
    • transphobic research presented at NeurIPS last year
      • very easily spotted by trans community
      • cf the work of Nicki Washington, at the intersection of race/gender/computer science — cf Ruha Benjamin
    • HCI has value-sensitive design, as does STS
    • interdisciplinary work helps spread this knowledge
    • incentive mechanism: value papers holistically, and not condemning papers as “just dataset papers”
    • self-auditing does not work!