We all ready have a well defined rigorous AGI[0] that would have god like intelligence. There are some approximations that played pacman [1] without knowing anything specific about pacman itself. i.e. it learned how to play by itself.
AIXI is an artificial 'general' intelligence. It is not hard coded to do a specific task. That neural network is useless at doing anything other than playing tic tac toe
Actually no -- it's the neural network + biases, start values that's good at playing tic tac toe. The neural network code itself is not "hard coded" for any specific task.
You could use the same neural network with different inputs to do other stuff, like recognize characters, and all kind of pattern matching / classification.
If by "null hypothesis", you mean artificial intelligence that exceeds human intelligence is impossible, that is a bold claim. Since human intelligence is nothing more than a successful evolutionary algorithm implemented in 3-dimensional cellular networks, it is inevitable that we'll be able to copy and improve the algorithm at some point in the future. Then we'll have artificial intelligence that exceeds human intelligence. Then what?
[0] http://wiki.lesswrong.com/wiki/AIXI [1] http://www.youtube.com/watch?v=yfsMHtmGDKE