AI Takeoff
The First Thirty-Six Months of Recursive Self-Improvement
What changes after AI can improve the process that improves AI? A conditional future history follows the consequences for research, jobs, households, institutions, and the physical economy.
The argument
The critical threshold is a research feedback loop: systems design experiments, interpret results, and produce improvements that accelerate later AI research. Digital work can change quickly while power, chips, laboratories, and construction impose constraints. Those constraints can themselves become targets of AI-assisted improvement.
How to read this assessment
March 15, 2028 is an illustrative starting point, not a prediction of a breakthrough on that date. The chronology also assumes useful deployment, diffusion, and no civilization-ending interruption.
A question worth pursuing
Which constraints slow transformation, and which can advanced AI help remove?
Cite and share
This page provides an editorial introduction. Consult the paper for its complete evidence, assumptions, sources, and collaboration disclosures. Downloads preserve the edition retrieved on October 18, 2026.