[throwaway1851]

I think that’s a recurrent theme with deep learning: it’s incredible, almost magical, when it’s well adapted to a specific task that fits specific parameters. But failure modes are a minefield.