Anthropic’s new economic scenarios are out, with the Extreme case projecting 15.4% annual GDP growth by 2030. But I think even this may underestimate the upside. The model mainly captures productivity gains from automating existing cognitive work. Several potentially powerful second- and third-order feedback loops are still only partially captured. RSI: Anthropic’s Extreme case already assumes recursive self-improvement is likely. But the larger feedback loop may be even more important: stronger models accelerate AI research, which produces stronger models, which further accelerate research—not only in AI, but across science and engineering. AI creating new scientific knowledge: The bigger variable may not be doing today’s work 2x faster. It may be discovering entirely new technologies across materials, biology, energy, chips, chemistry, and manufacturing. Robotics + 3D printing: AI-designed robots, automated manufacturing, and additive manufacturing could eventually create a recursive improvement loop for embodied systems themselves—better AI designs better robots, which build better robots and more of the infrastructure behind them. That could extend automation from the cognitive economy into the physical economy far faster than most forecasts assume. Once manufacturing, logistics, construction, energy and services become increasingly automated, the productivity gains spread far beyond office work. So Anthropic’s Extreme Scenario still looks closer to a Cognitive Automation Economy than a full Recursive Innovation Economy. In a true recursive innovation economy, growth may not accelerate linearly. Each improvement creates the capacity to generate the next improvement faster, while second- and third-order effects compound across industries. Sergey Brin recently made a point I think captures the problem well: even five years out, it may be impossible to know what the world will look like. I think he’s right. No one really knows. Not even Anthropic.