> I asked Hinton if he believed an A.I. could be controlled. Ugh, that question just goes back to the boring pie-in-the-sky stuff. The interesting question i

> I asked Hinton if he believed an A.I. could be controlled. Ugh, that question just goes back to the boring pie-in-the-sky stuff. The interesting question is if the automated processes we build on these tools can be secured. That's the terror the next generation has to live with. NNs can be so opaque it might be hard to tell if someone has tinkered with the input data/retrained it/overridden it. I shudder to think of all the legal implications. Like, in the above example a trout murderer could just flip some bits and nobody would know there are no more trout because the system is still reporting that there are lots of trout and the manual checks have gone.