I'm looking forward to a system that only describes what it sees (and what can be drawn from it), rather than doing stuff like trying to explain non-visual jokes (that, inevitably, it doesn't get). As I understand, that would require training new models from scratch, on training data that might not yet exist – but unlike the majority of the LLM GenAI hypey nonsense, this is actually a real, useful tool.
One really smart thing they do is localize the volunteers as much as possible, so often if the question is something like "Am I at the corner of X & Y" there is a good change you (the volunteer) know the area, etc.
Things I have learned on the "volunteer" side: keeping track of canned food is often a problem, people worry about their outfits a fair bit, and parks are sometime difficult because there are lots of alternate paths/routes.