I think this is a good observation. One thing I’ve been thinking about is probabilistically evaluating traditional software. Every un-caught exception, for exam

I think this is a good observation. One thing I’ve been thinking about is probabilistically evaluating traditional software. Every un-caught exception, for example, is usually a false negative when evaluating “did the code correctly handle the input?” Not that ML code is bug-free. But it seems like most traditional software is probabilistically correct, despite not being thought of that way.