[mlthoughts2018]

One caveat though is that in many situations, there is a lot of bureaucracy and overhead involved with collecting and interpreting user feedback. User feedback is often inconsistent and contradictory, influenced by mistakes in survey design, influenced by time-varying factors. And once it is collected, a lot of different stakeholders inside the company will have conflicting incentives about what the feedback is supposed to mean for priorities and decisions.

In some cases, if someone has a very highly developed sense of aesthetics and design, or can extrapolate from previous user feedback scenarios, then they might be able to come up with designs that users will love even if users wouldn't have thought of it that way ahead of time, or if more generic feedback would never have collated down into actionable decisions that empowered that particular designer.

As I've gone on in my career, I tend to see the value of this more and more, and lose faith that feedback collection mechanisms won't be politically subverted.

One example: I'd say a lot of the feedback about the principal lines of work for TensorFlow from the recent dev summit is wrong-headed, and they should just let Francois Chollet's design aesthetic motivate what gets worked on and how it gets designed, for a while.