Start your day with intelligence. Get The OODA Daily Pulse.
Published on the DeepMind Institute platform on November 19, 2026 (drawing on DeepMind’s From AGI to ASI research led by Tim Genewein, Shane Legg, and colleagues), the essay moves beyond abstract speculation to classify how artificial general intelligence (AGI) could transition toward artificial superintelligence (ASI) treating AGI as a transition point rather than a fixed finish line. The analysis maps overlapping routes toward ASI, centered on continuing model, data, and test-time compute scaling, pursuing technological and algorithmic paradigm shifts beyond today’s transformer/next-token architectures, enabling recursive self-improvement where AI accelerates AI research, and scaling multi-agent coordination (“group agents” or agent collectives) capable of solving problems beyond any single system. While current transformer and reasoning/test-time approaches can execute rich in-context algorithms, the essay notes gaps compared with humans in continual learning, world modeling, decision-making, sample efficiency, and inventing genuinely new concepts/abstractions rather than recombining existing ones leaving debate over whether current paradigms can scale into ASI or require deeper breakthroughs.