[jasonmaydie]

How would you build a business on something you do not understand well? I've worked with startups who could not get past that 80% barrier you run into with your machine learning algorithms. They didn't know if their data was bad, their algorithm was bad or whatever the issue was because they ere not using data a human could easily conceptualize.

[Eridrus]

One of the more surprising things about ML is that it can look like it's working, even when there are horrible bugs.

This makes testing even more important than usual, but testing ML systems is even harder than testing normal software, so you're in a shitty place either way.

ML ends up being this sort of bimodal level of difficulty, either it's super easy and everything works without much effort (it happens sometimes), or it doesn't work quite how well you need it to and you should probably be ready to sinks many months into it.

[dchichkov]

That. Also amplified by hype, money and thinking that you could learn the required skillset fast.