We still are not good at it. We're developing our internal machine learning platform as a normal product[0] and treat questions as <i>failure demand</i>[1] to update documentation, or eliminate the friction that lead to the question.<p>The whole point of the platform is to take a lot of the burden of machine learning off of the user, so that complexity is handled by the platform to make things as simple as they possibly can be, without becoming rigid.<p>If many people ask us about something, there's an opportunity not only to document that, but to reduce ambiguity upstream so the question never gets asked again.<p>We also monitor errors and usage and ask our users and colleagues questions on their usage, or why they were not able to do something, and learn from that.<p>We also are our users and when we use the platform and find we're having trouble remembering things, we add that in. For example, when we added the functionality to deploy a model, we added a tutorial on how to invoke the model with different languages, how to make a request, etc.<p>Not the best at UI/UX because it's not our background, but we're improving.<p>- [0]: <a href="https://iko.ai" rel="nofollow">https://iko.ai</a><p>- [1]: <a href="https://vanguard-method.net/failure-demand/" rel="nofollow">https://vanguard-method.net/failure-demand/</a>