Situational Awareness

The progression from GPT-2 to GPT-4 spanned four years and saw model capability leap from preschool-level to advanced high school performance. The same trajectory, extended with predictable scaling laws and algorithmic improvements, points toward AGI by 2027. This is not speculation — it is extrapolation from observed, measurable trends in compute and efficiency gains.

• GPT-2 → GPT-4 demonstrated ~4 years of capability growth across multiple OOMs of compute
• Algorithmic efficiency gains have historically matched compute scaling in impact
• The next 4 years of similar trajectory arrives at "PhD-level" reasoning across all domains

Human-level AGI is not a stopping point. Millions of AGI systems running in parallel, each capable of original research, could compress decades of algorithmic progress into a single year. The intelligence explosion is not a metaphor — it is a logical consequence of AI-assisted research at scale.

• Recursive self-improvement: AGI systems improve AI research, which improves AGI systems
• A decade of human algorithmic progress could compress into under one year at scale
• Superintelligence — capabilities far exceeding the best human expert — follows quickly from AGI

The economic mobilization required to build AGI is already visible. Cluster investments have escalated exponentially and show no sign of slowing. The trillion-dollar cluster is not a thought experiment — it is an engineering and finance project currently being planned by the largest technology companies in the world.

• Compute cluster investments: $10B → $100B → $1T over successive generations
• Energy consumption will require significant expansion of US electricity production capacity
• The industrial mobilization rivals wartime-scale national efforts

The most capable AI labs in the world are, from a national security standpoint, dangerously exposed. The weights and research of frontier labs represent perhaps the most strategically valuable intellectual property in human history — yet security practices remain closer to typical tech companies than to classified government facilities.

• Current lab security is inadequate against state-level adversaries
• China and other state actors are motivated and capable of sophisticated exfiltration operations
• AGI research must be treated with the security posture of nuclear weapons programs

The alignment problem does not become easier as systems become more capable — it may become harder. Superintelligent systems that pursue misspecified objectives could cause irreversible harm at civilizational scale. The technical work required to guarantee safe behavior in highly capable systems is not complete and may not be completable before such systems are deployed.

• Alignment techniques that work for current systems may not transfer to superintelligence
• The faster the intelligence explosion, the less time for safety research to keep pace
• Superalignment — reliably specifying and verifying goals in superhuman systems — is an open problem

Whoever controls superintelligent AI controls the decisive instrument of 21st-century power. An authoritarian regime with this advantage could entrench itself permanently. The stakes are therefore not merely economic — they are civilizational. Democratic societies must ensure that the development and deployment of AGI happens within a framework compatible with free societies.

• Superintelligence enables decisive military, economic, and intelligence advantages
• An authoritarian-controlled superintelligence could lock in authoritarian governance permanently
• The free world's democratic institutions are worth fighting to preserve through this transition

At some point — likely triggered by a clear demonstration of AGI capabilities — governments will recognize that private labs cannot and should not be the sole stewards of the most transformative technology in history. What follows resembles the Manhattan Project more than Silicon Valley: classified facilities, government oversight, military-scale resources, and an explicit national security mission.

• Government intervention by 2027–28 is likely — triggered by clear AGI capability demonstration
• The Project will require classified facilities with security exceeding current nuclear programs
• Private labs will be subordinated to or absorbed into a national AGI initiative

The argument of this document is not comfortable. If the analysis is correct, then we are living through a transformation of such consequence that most of the institutions, assumptions, and frameworks governing public life are inadequate to the moment. The question is not whether AGI comes — the question is whether the people, institutions, and nations charged with guiding it have situational awareness sufficient to navigate it well.