>why should anyone work on one [an AI]?
Because we can.
> i understand that conventional ai will become increasingly important
>and useful, of course, but by not allowing programs to modify their
We'd certainly program in broad goals, but in a complex world we'd never know all their ramifications so AI's would never stop surprising us. Also, despite our best efforts their goals would undoubtedly contain contradictions, just as our own personal goal system does. In addition, I'm certain an AI would soon start modifying its goal system on its own. As you say we'd probably try to stop them from doing this but it would be a futile effort, you just can't outsmart something more intelligent than yourself.
John K Clark firstname.lastname@example.org