He's saying, "we mathematically figured out a long time that trying to look at a computer program and predetermine what it will do before running it is a task that reduces to the halting problem."
That this is a stupid way to look at the antivirus problem is besides the point here.
You don't need to figure out what a program can do ahead of time if you limit what it can do at run time. You don't even need to let the user do anything at run time.
PS: I also just noticed that that was commic #386.
What's the difference between 'rm -rf foo', for example, and 'run-virus'?
I can't think of an especially usable uber-paranoid defensive system, although I'd definitely love to be proven wrong. Essentially, the problem isn't a totally technological one. The core of the problem is that the user can effectively be convinced to be their own attacker.