One does not simply turn the reactor all the way off
Perhaps not, but I'd expect you'd flip a switch, and then the automated systems would shut it off, or something to that effect.
Which is typical hard to parse statement which translates to: after 72 hours, the cooling water reserve evaporates off, and you get yourself regular meltdown like in any other design.
So, set it to shut down automatically after 36 hours without operator action, and it will be fine.
can release up to 10 Chernobyls of cs-137 or so
Why didn't they have that problem at Three Mile Island?
Perhaps not, but I'd expect you'd flip a switch, and then the automated systems would shut it off, or something to that effect.
You expect wrong. Nature isn't rubber padded, and technology isn't friendly magic. You flip switch, control rods go in, the chain reaction stops, the decay heat continues.
So, set it to shut down automatically after 36 hours without operator action, and it will be fine.
It is 72 hours after the full shutdown, that it melts itself down. edit: or to be pedantic, gets outside design parameters; the melt may take another couple da...
Just a reminder that some of the old threats are still around (and hence that AI is not only something that can go hideously badly, but also some thing that could help us with the other existential risks as well):
http://blog.practicalethics.ox.ac.uk/2012/03/old-threats-never-die-they-fade-away-from-our-minds-nuclear-winter/
EDIT: as should have been made clear in that post (but wasn't!), the existential risks doesn't come from the full fledged nuclear winter directly, but from the collapse of human society and fragmentation of the species into small, vulnerable subgroups, with no guarantee that they'd survive or ever climb back to a technological society.