Google DeepMind's 60-page paper outlines four pathways from AGI to ASI: scaling, new algorithms, recursive self-improvement, and multi-agent collectives.