Rob Wiblin argues artificial intelligence systems from iterated optimization for narrow objectives have uninterpretable connections and generate erroneous outputs outside training distributions, offering little reason to expect full human capabilities · Digg