We have lost control/rev II

Photo by Natasa Grabovac on Unsplash

BRT has talked copiously about AI and the inherent dangers said tech has as its open ended and evolving at rates far beyond the kin of man thanks to the marriage of analog neural nets to real time genetic algorithms. Seems yours truly is not alone about this given just how fast AI is gaining ever greater smarts 24/7. 

We argue that total containment is, in principle, impossible, due to fundamental limits inherent to comput-ing itself. Assuming that a superintelligence will contain a program that includes all the programsthat can be executed by a universal Turing machine on input potentially as complex as the state ofthe world, strict containment requires simulations of such a program, something theoretically (andpractically) impossible.“Machines take me by surprise with great frequency.  This is largely because I do notdo sufficient calculation to decide what to expect them to do.”Alan Turing (1950), Computing Machinery and Intelligence, Mind, 59, 433-460

The timing of the new debate about the dangers of superintelligence is not arbitrary.  It coincideswith recent demonstrations of human-level control in classic arcade games via deep reinforcementlearning (Mnih et al., 2015).  The key feature of this achievement is that the AI uses purely un-supervised reinforcement learning – it does not require the provision of correct input/output pairsor any correction of suboptimal choices, and it is motivated by the maximization of some notionof reward in an on-line fashion.  This points to the possibility of machines that aim at maximizingtheir own survival using external stimuli, without the need for human programmers to endow themwith particular representations of the world. In principle, these representations may be difficult forhumans to understand and scrutinize.

8 steps to control AI ... we hope.

Such an “ethics engineering” approach has been very useful in the design of systems that make autonomous  decisions  on  behalf  of  human  beings.   However,  their  scope is  not  suitable  for  the problem of controlling superintelligence (Yudkowsky, 2007).

In other words, we don't know how neural net driven AI works and ... real time evolving code must be written by code, not humans, this is why we have lost control.