DeepMind has released a 145-page paper outlining its approach to AI safety as it attempts to build advanced systems that may one day surpass human intelligence.