A central lesson from his talk for me is to look more closely how to design interactive systems so that people can exploit the technical weakness of system creatively. We will always have to deal with sensors systems, context-recognition, and learning algorithms that are not 100% perfect. I find it interesting to see this rather as a resource for design than a problem. The experience Tom reported from CYSMN (http://www.equator.ac.uk/index.php/articles/618) show nicely how people make use of GPS inaccuracies in a game.
A further point to keep in mind is that when triggering events based on context you may get boundary effects that can break the user experience. Tom gave an example of children finding virtual animals based on location. The effect was that they stopped when the saw an animal on their device – and this was at the boundary of the trigger area. This led to cases where the animal appeared and disappeared on the device and the children were puzzled about this effect. Hence one should really be careful how to put triggers – and I would expect that this is generally applicable to context-awareness not just to location-aware application.