After 2016, I really did feel like we needed to make a substantially larger investment. It's actually a much more complicated set of philosophical rather than technical questions. We use transparency reports every quarter and the metric that we track is for each of those 20 types of harmful content. How much of that content are we taking down before someone even has to report it to us? So how effective is our AI at doing this? But that basically creates this big question, which is, okay, now we need to really be careful about how proactive we set the AI.