(I wrote this before seeing timtyler's post.)

If there is a rule that says 'optimize X for X seconds' why would an AGI make a difference between 'optimize X' and 'for X seconds'?

I does seem like you misinterpreted the argument, but one possible failure there is if the most effective way to maximize paperclips within the time period is to build paperclip-making Von Neumann machines. If it designs the machines from scratch, it won't build a time limit into them because that won't increase the production of paperclips within the period of time it cares about.