Adam Bede

    M10: Operationalizing AI

    • When the technology is poorly understood + has high expectations, you leave it open to underwhelming as you don’t have clear signals and targets.
    • B/c of the speed, you will have tinkering and UX issues so you will have to mitigate the friction user’s feel
    • How to use wins to generate traction
    • Ask, what are the differences of degree and kind in this technology that we should consider when approaching a problem. Place these ?s w/in the user experience and how our teams will interact w/ and be affected by the tech.
    • Oracle vs. AI; only the latter we see as a threat.
    • The idea of AI and trust. Trust as the great lubricant and to what extent do we need to trust technology
    • How do you train/test your AI will benchmark and set the standard, “with respect to whom/what?”
    • “95% accurate” compared to what? Baselined from? How to calculate ROI from what baseline? What is your measure of performance?
      • Bias vs. noise Daniel Kahneman; send that to him, which he’s probably seen
    • How we measure AI projects is different from Software and ML model projects

    Bias is a statistical concept. If your shots group in a particular quadrant then they’re biased.

    Fairness:

    • If you’re fair concerning outcome, you might be unfair in respect to the process. (and vice versa)
    • “decreasing the accuracy and increasing the fairness”
    • Lotteries and counterfactuals