Openai says that while it may seem far off, superintelligence could arrive this decade
Good argument against recursive self improvement in AIs without needing to interact with the world (and thus bottlenecking the speed)