arXiv 2606.12683 enumerates four concrete pathways from AGI to ASI

The report formalizes the post-AGI regime by extending Marcus Hutter’s Universal AI framework to systems that surpass collective human performance. It evaluates each pathway against measurable bottlenecks including data efficiency, compute scaling laws, and alignment stability. Recursive improvement receives the most formal treatment, with explicit open questions on convergence rates once an AGI can rewrite its own training loop.

Empirical anchors include current frontier model scaling curves and multi-agent experiments from DeepMind’s 2023-2025 work on population-based training. The paper notes that none of the four pathways yet shows a quantified friction threshold that would enforce a multi-year plateau. This absence of hard bottlenecks distinguishes the analysis from prior AGI timeline surveys that treated human-level performance as a singular discontinuity.

Operationally the document implies that organizations must prepare for overlapping capability jumps rather than a single deployment event. It calls for interdisciplinary tracking of scientific acceleration metrics across domains once AGI systems are widely deployed. The listed open research questions directly map to measurable indicators such as self-modification success rate and cross-agent knowledge transfer velocity.

Preparation therefore shifts from single-model alignment to monitoring distributed intelligence growth rates and their cumulative effect on technological progress.