Brief notes

  • Holographic crystal storage‽ And rewriteable! And self-supervised learning can help you learn the particular manufacturing artifacts of a camera and a crystal‽
  • ML can help target radiotherapy beams, taking a 73 minute process and turning it into 5 minutes
  • software that is learned from data is the AI revolution (much like Karpathy’s Software 2.0)
  • Federated Learning can be backdoored (“Attack of the tails”)
    • I’m unsure about the trust model for the interaction: I trust a company to run arbitrary programs on my device, and send arbitrary ML models to a server, but not the raw data? We’ve been able to extract patterns in raw data from ML models, so,
    • Even still, that you can misclassify a Southwest Airlines plane as a truck is Not Good
  • Big self-supervised models are strong self-supervised learners
    • distillation and self-supervision are popular and useful, and SimCLR(v2) sounds promising
  • Unsupervised translation between programming languages
    • with monolingual code!
  • Saiph Savage: A Future of work for the invisible workers in AI
    • 60% of US workers involved in invisible labor (MTurking etc, below the API) in five years?!?!?!?!?!
    • web plugins can repurpose labor platforms for good
    • twitter bots trying different messaging found direct CTAs the most “effective”
  • JAX MD for molecular dynamics
  • Worst-Case Analysis for Randomly Collected Data tries to correct for snowball sampling, much more effective than oversampling