What if he's inverting his every thought, in the hope that he'll be right?

They're not standing in the way though -- they're standing inside the AC vent pipe (or wherever that is), and creepily looking at us through the blinds.

Also since last layer is likely a dense/fullyconnect layer, you should be able to start out with original weights (padded for new output depth).

Why? If you're just adding a category, why not use what you've got and just train the last layer. You mean in case the new category has some unique low-level feature that early layers of network need to extract?

You must have never spent a ridiculous amount of time learning and building on some api, just for some committee or company board to decide its time to "end-of-life" it (i.e. can the whole thing). It changes you, and your future api choices.