The fact the author can think of one particular way to design an AI that would result it in falling over and dying once it figures out how to modify itself does not imply that all possible intelligences would behave similarly.
I was actually surprised a bit to see that the author was somewhat familiar with Eliezer Yudkowsky's writings on the topic (he cited http://lesswrong.com/lw/wp/what_i_think_if_not_why/), because the line of thought doesn't seem to incorporate a real understanding of what he's said on the topic (which, to be fair, is a huge body of work...).
Most of EY's "Friendly AI" worries are rooted in this idea that when considering the entire universe of algorithms that could be described as intelligent or self improving, we need to be exceptionally careful not to assume that more than a negligible percentage of them share anything in common with human intelligence, because for the most part, they won't, unless they're carefully and explicitly designed to do so.
Here, the author assumes that the AI is simply trying to optimize some internal measure of happiness, with complete disregard for the meaning of that measure. This is an incredibly naive view of how deeply important and carefully constructed any optimization target would have to be in any self-improving intelligent machine; it's literally the core of the entire problem of friendly AI, and to trivialize it by assuming that such an AI would ever even consider rewriting its "happiness button" to be always-on is to miss the entire difficulty of the problem.
Hell, it's even the core of the problem of non-friendly AI, because it doesn't even require human level intelligence to realize that if you rewrote your own code so that you were always thrilled with the result, that's the easiest way to increase "utility". Any self-rewriting algorithm that's capable of real self improvement has to, by design, be able to consider the likelihood that changes to its objective function will end up with negative expected value.
None of which is to say this isn't a valid concern, by any stretch. But it's not a proof of universality; in fact, getting around this type of problem is exactly what any real AI designer must contend with. It's an issue that's very well known, and it's certainly not well-accepted as an insurmountable hurdle.
I was actually surprised a bit to see that the author was somewhat familiar with Eliezer Yudkowsky's writings on the topic (he cited http://lesswrong.com/lw/wp/what_i_think_if_not_why/), because the line of thought doesn't seem to incorporate a real understanding of what he's said on the topic (which, to be fair, is a huge body of work...).
Most of EY's "Friendly AI" worries are rooted in this idea that when considering the entire universe of algorithms that could be described as intelligent or self improving, we need to be exceptionally careful not to assume that more than a negligible percentage of them share anything in common with human intelligence, because for the most part, they won't, unless they're carefully and explicitly designed to do so.
Here, the author assumes that the AI is simply trying to optimize some internal measure of happiness, with complete disregard for the meaning of that measure. This is an incredibly naive view of how deeply important and carefully constructed any optimization target would have to be in any self-improving intelligent machine; it's literally the core of the entire problem of friendly AI, and to trivialize it by assuming that such an AI would ever even consider rewriting its "happiness button" to be always-on is to miss the entire difficulty of the problem.
Hell, it's even the core of the problem of non-friendly AI, because it doesn't even require human level intelligence to realize that if you rewrote your own code so that you were always thrilled with the result, that's the easiest way to increase "utility". Any self-rewriting algorithm that's capable of real self improvement has to, by design, be able to consider the likelihood that changes to its objective function will end up with negative expected value.
None of which is to say this isn't a valid concern, by any stretch. But it's not a proof of universality; in fact, getting around this type of problem is exactly what any real AI designer must contend with. It's an issue that's very well known, and it's certainly not well-accepted as an insurmountable hurdle.