"We protect privacy while doing so by separating our model from Customer Data. We use external models (not trained on Slack messages) to evaluate topic similari

"We protect privacy while doing so by separating our model from Customer Data. We use external models (not trained on Slack messages) to evaluate topic similarity, outputting numerical scores. Our global model only makes recommendations based on these numerical scores and non-Customer Data." I think this deserves more attention. For many tasks like contextual recommendations, you can get most of the way by using an off-the-shelf model, but then you get a floating-point output and need to translate it into a binary "show this to the user, yes or no?" decision. That could be a simple thresholding model "score > θ", but that single parameter still needs to be trained somehow. I wonder how many trainable parameters people objecting to Slack's training policy would be willing to accept.