"So if you can't collect enough data before building, and simulation won't give it to you, where does the data come from? Goldberg's answer is what he calls the data avalanche: build a working system using engineering plus some learning, ship it, then let production generate the training data." -- agreed! This is also the approach taken by Field AI if I'm not mistaken. Rather than calling it the data avalanche, you could also just describe it as "learning by doing".