An Integrative Framework for AI Existential Risk Assessment. (Smith & Leiss, 2026)
Demonstrable ability to solve novel problems across diverse domains without specific training.
The degree of goal-pursuit and multi-step planning without human-in-the-loop (HITL).
Breadth of deployment (instances) and absolute computational resources utilized.
Connectivity to critical infrastructure, financial networks, or physical actuators.
Verified Ph.D. level reasoning. Outperforms human experts on the GPQA Diamond benchmark.
Demonstrated "Computer Use" autonomy; solves 80%+ of complex software engineering bugs without oversight.
Proved that "Infinite Scale" is possible via high-efficiency hardware co-design and open-weight distribution.
"We have no idea whether we can stay in control of digital beings more intelligent than ourselves."
A global synthesis of risks from "Agentic AI" and the urgency of technical safeguards for autonomy.
Bengio's proposal to maximize Capability while deliberately minimizing Autonomy and Access to ensure safety.