Google DeepMind researchers have published a formal conceptual framework analyzing the transition from human-level AGI to artificial superintelligence (ASI), mapping four technological development paths against six key structural constraints. Rather than speculating about a sudden “intelligence explosion,” the paper grounds the analysis in physical, theoretical, and economic limits — arguing that bottlenecks like the Data Wall, the Embodied Bottleneck (real-world experimentation speed), and the Abstraction Barrier (discovering genuinely new concepts from raw data) will shape a gradual, multi-path continuum rather than a singularity.

The framework builds on the Legg-Hutter universal intelligence measure and synthesizes landmark theories including Sutton’s Bitter Lesson, Bostrom’s instrumental convergence, and formal AI alignment work on corrigibility and interruptibility. The authors call for a shift from speculative debate to a rigorous research agenda — prioritizing un-saturating benchmarks based on algorithmic compression, quantitative multi-agent scaling laws, and high-fidelity physical simulators to safely navigate the post-AGI era.

From AGI to ASI: Pathways and Bottlenecks on the Road to Superintelligence (arXiv)