Situational Awareness: The Decade Ahead
Leopold Aschenbrenner. Situational Awareness: The Decade Ahead, 2024-06-04. About 165 pages in five chapters.
Aschenbrenner had worked on OpenAI’s Superalignment team until he was fired in April 2024. He then wrote this essay series, arguing that AGI by 2027 is “strikingly plausible” and that the US government would soon take over its development. It became one of the most discussed timeline documents, especially in Washington, and he went on to found an AI-focused investment fund of the same name.
The argument in brief
- Count the orders of magnitude. Progress comes from compute and algorithms, each improving about threefold a year, plus “unhobbling”, which turns chatbots into agents that use tools and work for long stretches. GPT-2 to GPT-4 was a leap from preschooler to smart high-schooler. Another leap of the same size by 2027 gives systems that can do the work of an AI researcher. “Trust the trendline.”
- Then an intelligence explosion. Millions of automated AI researchers compress a decade of progress into a year or less, reaching Superintelligence by the end of the decade. See Intelligence explosion.
- A trillion-dollar buildout. By the end of the decade, single training clusters cost more than a trillion dollars and draw power equal to over 20% of US electricity production. See Compute governance.
- Lock down the labs. Lab security is far too weak, and China can steal the key secrets and model weights.
- Superalignment is unsolved, and an intelligence explosion would leave little time to solve it.
- The free world must prevail. Superintelligence gives a decisive military advantage, so the US must stay ahead of China.
- “The Project.” By 2027–28 the US government will run a national AGI effort, much as it ran the Manhattan Project.
Responses
- It feeds a dangerous race. Gideon Futerman’s Against Aschenbrenner (EA Forum, 2024) argues that the race-with-China framing makes catastrophe more likely, not less. See Can the US and China agree to limit AI.
- The obstacles are bigger than the trendlines. Lawfare’s AI Timelines and National Security: The Obstacles to AGI by 2027 argues that power, chips, data and the step from benchmarks to real work all slow the curve.
- Zvi Mowshowitz collected and commented on its key passages in June 2024.
How it has held up
Reviews in 2026 (EA Forum, Philipp Dubach) agree that the technology and investment forecasts have held up, while the political ones, on lab security and a government-run “Project”, have not, and AGI by 2027 is still open.
Related: AGI timelines, AI 2027.