Both in the technical sense and in the human sense.
Technical: People at Linus' caliber understand exactly the rules for signed/unsigned integer promotion and where underflow is defined (as wrap) and where it's undefined. Consequently he wrote perfectly-correct code for detecting the underflow and bailing out safely. Programmers at mere mortal levels of skill, however, routinely mess this up, often causing exploitable security bugs (believe me, I do code security audits as part of a real honest living). My advice for everyone (contra Linus!) is always always always use the compiler intrinsics for integer math. Feel free to decline this advice if you are a Linus level wizard (if you were, of course, you would already feel free to decline it) but if you have to wonder if you are, you probably aren't.
Linus seems to think that the kernel should only be written by folks that don't need that kind of help -- and for that I won't argue with him. It's his baby and he can chose whether to have a small number of Ã¼ber-developers or a larger number of mortals. Which goes straight to the second point:
Human: People at Linus' caliber thrive on negative feedback. At their level, positive feedback means nothing because there's nothing he can learn from someone praising his work. He wrote a kernel, he knows he's good. Meanwhile negative feedback is useful (unless trivially discountable): if the complaint is right, he'll correct something he was doing wrong; if the complaint is wrong, he'll be forced to think through why. In any event, he could never imagine why someone would sugar-coat their opinion on any matter.
So it seems like his mode of communication is meant to answer that question for the former: he wants people of his caliber that don't write ugly code using arithmetic crutches and don't care about strongly worded criticism. There's nothing invalid about that either -- maybe it's true that the best model is that Linuses work in the kernel and the rest of us go up into userland where we use crutches like memory protection and higher-level constructs :-)
 And when behavior is undefined, a smarter compiler can remove the code-path entirely -- the kernel itself was hit by such a bug where GCC legally removed a NULL check because the pointer was dereferenced before the check. See also this reference. Then there's the sad fact that people still argue against the clear language rules that say that assert( 100 + some_int > some_int ); can always be optimized away.