Situational Awareness

Situational Awareness is a forecasting framework that maps the likely trajectory of advanced artificial intelligence. Published in June 2024 by Leopold Aschenbrenner under the AI Futures Project, it offers a structured argument for when and how human-level and superhuman machine intelligence might arrive.

The framework rests on a single gating variable: the supply of training compute, measured in H100-equivalent chips, which it treats as the binding constraint on near-term AI progress. From this base, it projects decade-by-decade capability growth from 2024 onward, branching into scenario families that include AGI, superintelligence, and the national-security consequences of concentrated AI power. The argument leans on a claimed pattern of stubborn exponential growth in compute and algorithmic efficiency over recent decades, treating that trend as a basis for extrapolation rather than a guarantee.

The framework has been praised for making its assumptions explicit and tying forecasts to a measurable quantity rather than to vague intuition. Critics have argued that the compute-centric view underweights data quality, energy constraints, hardware bottlenecks outside accelerators, alignment difficulty, and the possibility that capability gains stall before reaching human-level generality. Some readers also note that the national-security framing reads as advocacy rather than analysis, and that the projection timelines depend on extrapolations that have historically broken in other technologies.

Open questions include whether compute remains the dominant constraint once current scaling laws saturate, whether superintelligence scenarios should be modeled as a single phase or as a long transition, and how the framework should be revised if major-lab progress slows or accelerates relative to its 2024 baseline. The specific date estimates for AGI and the broader superintelligence timeline remain contested across the AI research community.