We argue that artificial intelligence systems must be able to manipulate their own internal representations automatically in order to deal with an infinitely complex and ever changing world, and to scale up to rich and complex applications. Such manipulation must go beyond the ability to change beliefs and learn new concepts in terms of old; it must be able to change the underlying syntax and semantics of the ontology. Initial progress is being made and is now urgent, due to the demands created by autonomous multi-agent systems. Understanding and automating this ability must be a major focus of artificial intelligence for the next 50 years.