Training is a much, much different issue. There’s a huge difference between an API telling Sense that something is on and a user telling Sense that something is on. The API doesn’t ‘train’ Sense, but tells Sense every single time that a device is on (in this case, an LED bulb). Training is a much different issue that involves you telling Sense that something is on enough times and in enough contexts for Sense to be able to do so on its own. That’s not an impossible feat, but it would be incredibly long winded (much more than just turning a device in and off ten times in a row). We would love to incorporate a feature like this. Why wouldn’t we? It would make the device detection process a million times easier. It’s just not feasible though. 10 manually IDed on/off cycles really doesn’t give us that much data.
I’ve written about this recently here: Why can't you train Sense? - #4