Seems obvious to me that AIXI is describing a fully general learner, which is not the same as a FAI by any stretch. In particular, it's missing all of the optimizations you might gain by narrowing the scope, and it's completely unfriendly. It's a pure utility maximizer, which means it's a step *down* from a smiley-face maximizer in terms of safety - it has *no* humane values.

An AIXI solving a mathematical game would optimize. An AIXI operating in the real world would waste an awful lot of time learning basic physics, and then wirehead - if you were lucky.