Felix is essentially a Utility Monster: a thought experiment that's been addressed here before. As that family of examples shows, happiness-maximization breaks down rather spectacularly when you start considering self- or other-modification or any seriously unusual agents. You can bite that bullet, if you want, but not many people here do; fortunately, there are a few other ways you can tackle this if you're interested in a formalization of humanlike ethics. The "Value Stability and Aggregation" post linked above touches on the problem, for example, as does Eliezer's Fun Theory sequence.

You don't need any self-modifying or non-humanlike agents to run into problems related to "Torture vs. Dust Specks", though; all you need is to be maximizing over the welfare of a lot of ordinary agents. 3^^^3 is an absurdly huge number and leads you to a correspondingly counterintuitive conclusion (one which, incidentally, I'd estimate has led to more angry debate than anything else on this site), but lesser versions of the same tradeoff are quite realistic; unless you start invoking sacred vs. profane values or otherwise define the problem away, it differs only in scale from the same utilitarian calculations you make when, say, assigning chores.