Situational Awareness: The Decade Ahead (situational-awareness.ai)

In "Situational Awareness: The Decade Ahead," Leopold Aschenbrenner argues that we’re entering an accelerated AGI race: continuing trends in compute, algorithmic efficiency and system “unhobbling” make human-level AGI plausibly reachable by 2027, with a subsequent rapid push to superintelligence before the decade’s end. The essay traces GPT-2→GPT-4 as a four-year qualitative jump and quantifies progress in roughly 0.5 orders-of-magnitude (OOM) per year for both compute and algorithmic gains. Crucially, he warns that hundreds of millions of automated AGIs could automate AI research, compressing what would normally be multiple OOMs of progress into months and triggering an intelligence explosion. The implications are industrial, technical and geopolitical. Trillions of dollars and massive power buildouts (U.S. electricity up by tens of percent) will be poured into GPU farms, datacenters and supply chains; national security actors will mobilize (“The Project”) and government-led AGI efforts are likely by 2027–28. Aschenbrenner stresses urgent risks: labs are insufficiently secured against state actors, model secrets and weights are vulnerable, and reliably aligning systems far smarter than humans remains an unsolved but critical problem. His bottom line: the AI community must treat infrastructure, opsec and alignment as first‑order priorities now or face rapid, high-stakes shifts in economic and military power.