Questions surrounding the future of AI are creeping up on us. For many, it seems like a far out sci-fi dream that couldn’t possibly be realised in the near future. But for many — mainly those who are in the know — this distant sci-fi vision is seemingly becoming more and more realistic, within a shorter and shorter time frame. Nick Bostrom addresses many of the issues surrounding the future of AI and the infamous ‘control problem’ in his book ‘Superintelligence’. Firstly, what he makes clear is that Superintelligence is not an ‘if’, it is a ‘when’. A lot of thought must be given to how a Superintelligent AI can help humanity, not lead to its destruction — by ensuring it doesn’t stray from a ‘good’ path. The issue of Superintelligence, as a faculty of intelligence that can barely be fathomed by human intelligence, leads to a myriad of complex philosophical and scientific problems. These largely surround the fact that an unprecedented level of intelligence can most likely traverse the safety measures or booby-traps that humans lay in its way to ensure it remains safe. Similarly, there is an added level of complexity in thinking about the issue as we cannot anthropomorphise Superintelligent Will, due to the inevitable divergence between the goals, desires and motivations of an Intelligent product derived from billions of years of evolution (us) and an intelligent product that lacks such an extensive developmental process. Bostrom’s ‘Instrumental convergence thesis’ posits that an intelligent agent with unbounded but apparently harmless goals can act in surprisingly harmful ways in pursuit of intermediary goals. This leads to what Bostrom calls ‘Perverse Instantiation’; the smarter and more capable an AI is, the more likely it will be able to find an unintended shortcut that maximally satisfies the goals programmed into it. Basically, this is the implementation of a possibly benign final goal through deleterious methods unforeseen by humans. A notable example he discusses is the ‘paperclip’ scenario, where a Superintelligent AI may be tasked by a company to maximise the production of paperclips at their factory and proceeds by first converting the Earth and then increasingly large chunks of the observable universe into paperclips. Simply put, ensuring AI does what humans initially intended it to do is not easy. In order to make AI safe we must look at a variety of methods. Bostrom discusses ‘Capability Control’ and ‘Motivation Selection’ as primary methods, with various ways of carrying out such ideas - for example forms of ‘boxing’ — preventing access to outside world to some degree, ‘stunting’ — limiting access to information, ‘tripwires’ to shutdown AI if a certain negative threshold is reached and ‘domesticity’ — limiting the scope of its goals and ambitions (to name a few). While talking about making AI safe, we often ex ante presume that the introduction of AI will only increase the existential risks that humanity faces. However, it is interesting to note that this may not be entirely true. While the introduction of machine Superintelligence will create a substantial existential risk, it will reduce many other current existential risks. Risks from nature: asteroids, super volcanoes and pandemics will virtually be eliminated as Superintelligence could easily employ countermeasures or at least demote the threat to the non-existential category. Similarly, anthropogenic risks: accidents related to new technology, Climate Change, wars, technology races and undesirable nationalistic competition will be reduced as Superintelligent AI will prevent human fallibility from leading to its own extinction. Whilst the future of AI is not certain and there is plenty we don’t know — we must begin to divert focus and resources now to ensuring we can solve the AI related safety and control problems, before we are graced by its presence.
Comments
Top
New
Community
No posts