With each day I feel like I know less about modern-day ML/DL 😂 but I'm excited to talk about what I think is the biggest failure mode in many low-code (eg AutoML) systems: they aren't accessible to new groups of ppl bc they require the same knowledge (albeit fewer lines of code)

My opinion is that since ML is very human-centered (e.g., solving human-created problems), it's kinda a pipe dream to abstract away human thought. I'm more bullish on tools to help humans identify problem formulations, discriminative features, etc

and thus a good number of future challenges are in data management & visualizations, how do you make it really easy for non-experts to plug and play data sources? Instead of "cleaning" for them, how do you show them what needs to be cleaned & why? In my experience we benefit

when non-ML experts know more about the end-to-end ML process, and when they hear about the steps they are able to provide more insight (e.g. "actually those values shouldn't be cleaned bc [x] or [y] reasons").

So when the AutoML solution is a TF plugin, only ppl who know TF & why they need such a solution can use it. I'm excited about work that makes ML accessible for non-ML folks, and hopeful that more AutoML research trends in this direction