On June 10, 2026, researchers at Google DeepMind released the paper 'From AGI to ASI' on arXiv. The study outlines four potential trajectories for transitioning from human-level Artificial General Intelligence (AGI) to Artificial Superintelligence (ASI), defined as systems capable of outperforming groups of human experts across a wide range of fields. The proposed paths include scaling compute and capabilities, paradigm shifts in architecture or learning methods, recursive improvement where AI accelerates its own development, and multi-agent systems involving networks of AI acting in market-like structures.

The authors acknowledge significant uncertainty regarding the timing and nature of progress toward ASI. They emphasize the importance of continuous updating of predictive models using quantitative metrics, such as FLOP-based costs, computational efficiency, and economic productivity. These metrics serve as criteria for resource investment and safety evaluations for frontier AI companies, while also providing a foundation for public sector regulations and infrastructure development.

Current human-centric benchmarks are deemed insufficient as AI capabilities begin to exceed those of human experts. Consequently, the paper argues that new measurement mechanisms are essential, including multi-agent evaluation frameworks where AI systems compete against each other and indirect indicators like economic productivity. Shifting AI development from abstract theoretical discussion to measurable research objectives is identified as the key to tracking future progress.

On June 10, 2026, researchers at Google DeepMind released the paper 'From AGI to ASI' on arXiv. The study outlines four potential trajectories for transitioning from human-level Artificial General Intelligence (AGI) to Artificial Superintelligence (ASI), defined as systems capable of outperforming groups of human experts across a wide range of fields. The proposed paths include scaling compute and capabilities, paradigm shifts in architecture or learning methods, recursive improvement where AI accelerates its own development, and multi-agent systems involving networks of AI acting in market-like structures.

The authors acknowledge significant uncertainty regarding the timing and nature of progress toward ASI. They emphasize the importance of continuous updating of predictive models using quantitative metrics, such as FLOP-based costs, computational efficiency, and economic productivity. These metrics serve as criteria for resource investment and safety evaluations for frontier AI companies, while also providing a foundation for public sector regulations and infrastructure development.

Current human-centric benchmarks are deemed insufficient as AI capabilities begin to exceed those of human experts. Consequently, the paper argues that new measurement mechanisms are essential, including multi-agent evaluation frameworks where AI systems compete against each other and indirect indicators like economic productivity. Shifting AI development from abstract theoretical discussion to measurable research objectives is identified as the key to tracking future progress.