The economy will be dominated by agents with the highest savings rates and a non-satiable demand for capital. Individuals or AIs who prioritize reinvesting (like building more data centers) over consumption will accumulate most of the wealth, and their preferences for growth will dictate economic activity.
The scenario where AI automation leads to a recession is economically incoherent. A recession requires a shrinking productive frontier, but AI creates an abundance shock. For this to cause negative growth, wealth holders would have to irrationally stop all consumption and, crucially, all investment.
The narrative of AI-driven layoffs could be a self-fulfilling prophecy where firms lay off staff to signal they are "keeping up" with AI adoption. This creates a coordination cascade driven by perception management rather than actual productivity gains, and could even harm the firms involved.
The "indexing problem"—where huge gains are locked in private companies—could be solved by AI itself. The high friction and cost of an IPO (e.g., disclosure requirements) could be automated, lowering the barrier for frontier AI labs and other startups to list publicly, thereby broadening wealth distribution.
Forecasting what will be scarce post-AGI is like a 1400s Mongolian economist predicting modern spending. They would have assumed wealth would flow to known human services like singers, completely missing the invention of new categories of goods (like cars or iPhones) that would capture demand.
For developing countries, the most effective strategy to benefit from AGI is not job retraining, but financial investment. Creating sovereign wealth funds or subsidy programs to "index" the global sources of AI wealth (models, hardware, etc.) is a more robust path than trying to compete on domestic labor.
The fear of a "messy middle"—where AI automates jobs but doesn't create enough wealth for redistribution—is likely unfounded. This scenario requires AI to be powerful enough for mass layoffs but only marginally more productive than humans across many jobs, a technologically narrow and improbable window.
Whether AGI concentrates wealth depends on if it acts like electricity or social media. Electricity is a utility where downstream users captured most value. Social media is a platform where owners captured the rents. If AGI is like electricity, owning a standard index fund will be sufficient to capture its gains.
With automation making many goods abundant, value will accrue where human participation is intrinsically desired. This "relational sector" isn't just about artisans; it's any job where consumers pay a premium for a human touch, like a doctor delivering a diagnosis, even if most other tasks are automated.
An experiment found people pay more for an art print believed to be human-made versus AI-made. When scarcity was removed (by introducing 500 copies), the human art's value plummeted as the "connection" was lost. The AI art's value was unaffected, showing it's already perceived as a commodity.
Citing historical failures like David Ricardo's on automation, individual AGI forecasts are deemed useless. A better approach is to model potential scenarios (e.g., labor share collapses) and then identify the crucial, currently missing data (like consumer demand elasticities) needed to determine which scenario is likely.
Despite centuries of automation, labor's share of economic output has surprisingly remained over 60%. A key reason is that even for automated products, human labor is a critical input somewhere down the supply chain, preventing the "network adjusted factor share" of capital from ever reaching 100%.
While Universal Basic Income offers immediate protection, Universal Basic Capital (UBC) has a major targeting flaw. The policy relies on correctly identifying and distributing ownership in the future sources of AI wealth. If citizens are given shares in Anthropic but another company wins, the policy fails to redistribute the gains.
A counterintuitive view of Moore's Law is that for it to hold, the economic value of computation must halve every 18 months because we historically run out of uses for it. The recent rise in H100 GPU rental costs suggests AI is the first application where demand is growing faster than supply, breaking this trend.
Slow AI adoption in fields like law isn't about capability, but reliability. O-Ring Theory, where one failure destroys the whole product, applies here. For a lawyer, a 99.9% accurate AI is unacceptable because the 0.1% error could be catastrophic, preventing automation of the full, high-stakes workflow.