It will be interesting to see how this works over the long run. Without doing anything else to detect guns, how are they training the learning? I assume they are doing that somewhere else and updating (like the antivirus analogy they mention) but still seems like it would be fraught with false positives and negatives for a while.