Zing Data reposted this
Interesting article from Isabelle Bousquette @ the The Wall Street Journal -- though I think one element missing was some of the more clever ways to reduce the 'handholding' for ai assistants but still gain signal that improves the system or solves the 'cold start' problem of not having enough context for a reliable set of outputs. We spend a lot of time on this at Zing Data doing things like: * Learning context from the questions people have asked historically (without AI)...this is effectively training data if they are asked by knowledgable colleagues and skips 'starting from scratch' * Determining how data relates to each other far beyond join keys but also in semantic and vector representations * Providing things like SeeAhead that provide easy, visual ways to disambiguate ( https://lnkd.in/g_R2crut ) * Creating mechanisms to deterministically do things (drag+drop, etc.) that the model can learn from Much like Google's Pagerank learns about more authoritative sources, or Facebook's Newsfeed (which our team worked on!) learns what topics you spend more time on, you can similarly learn from the history of the data, and colleague's actions to build a much more robust and reliable system. https://lnkd.in/gWkTfGxc