To get the best performance out of HHVM, you have to let it "warm up" (JIT the functions you're using), and compile the bytecode ahead of time. Otherwise it assumes you could include new files it hasn't seen before, and has to skip out on some optimizations to accommodate that. So he can't claim he's faster when he's running it like this: @exec('hhvm ' . __FILE__ . " $test");

It's also worth noting that Facebook is actually pursuing this same approach with Hack (making the language more optimizable). And they've tried the AOT approach, and it turned out to be a huge pain for several reasons.