newsfilter.io
Interview, Fireside Chat

Highlights: Carl Shulman on the economy and national security after AGI

80,000 HoursCarl Shulman, Rob Wiblin
Watch on YouTube →

The Economics and Feasibility of Robot Nannies and AI Labor

  • Carl Schulman argues that "robot nannies" will likely become the norm in a mature AI future, dispelling the nostalgia that humans must provide primary childcare.
    • Core Advantages of AI Care:
      • AI systems offer continuous, 24/7 availability and individual attention, which are proven to improve educational and developmental outcomes compared to traditional classrooms.
      • Superior performance metrics: AI can be wittier, funnier, and more empathetic, optimized by data from millions of child interactions.
      • Safety and reliability: AI eliminates risks of abuse, laziness, or human error, with parameters set by parents to manage child satisfaction and bonding dynamics.
      • Confidentiality: Unlike human staff, AI can be programmed to "forget" or manage confidential information without risk of leakage.
      • Cost efficiency: A single AI unit can provide round-the-clock service for a fraction of the cost of a human team required to cover 24/7 shifts.
    • Market Dynamics:
      • In an economy with relatively equal incomes, human labor is too expensive to compete with AI for tasks where humans have lower intrinsic performance.
      • The "human taste" argument fails because no one would voluntarily work 24/7 as a low-paid nanny to afford similar care for their own children when AI offers superior service for near-zero marginal cost.

Energy Budgets and the Scale of Cognitive Output

  • Current Energy Constraints: Human energy consumption is ~10¹³ watts, while solar energy hitting the atmosphere is ~2×10¹⁷ watts (20,000 times current usage).
  • Per-Capita Energy Potential on Earth:
    • Harvesting 5–10% of solar input could sustain 1 million watts per person.
    • Given a human brain uses ~20 watts and a body ~100 watts, this energy budget could support 50,000 "human brain equivalents" of AI cognitive labor and 10,000 scaled robots per person.
  • Economic Value of AI Compute:
    • One hour of human brain-equivalent compute operating 8,760 hours/year at $100/hour generates ~$1 million in value.
    • A single human's energy budget on Earth could theoretically sustain >$50 billion worth of skilled cognitive labor at current high-end wages.
    • Off-world expansion (solar system) offers ~2 billion times more energy, potentially increasing total cognitive capacity by a factor of 100 million to 1 billion.
  • Industrial Implications:
    • A society with trillions of AI minds and distributed robotic bodies would possess physical, industrial, and military capabilities orders of magnitude larger than today's.

Reconciling Low Historical Growth with Future Projections

  • The "Boring Present" Objection: Skeptics note that despite decades of IT advancement, economic growth rates have been flat or declining, seemingly contradicting the anticipation of explosive AGI-driven growth.
  • The Diminishing Returns of Compute:
    • Historically, doubling the compute power has only yielded ~1/5th the revenue increase of the computer industry because early computers solved only a tiny fraction of economic tasks (e.g., arithmetic).
    • Unlike human populations where doubling labor often doubles output (due to complementary capital and scale), compute has faced diminishing returns until recently.
  • Future Trajectory:
    • As AI reaches a threshold to robustly replace humans, the share of the economy computable will rise.
    • Expected revenue impact per doubling of compute is projected to jump from 1/5th to 1x or higher as AI substitutes for broad human labor rather than just niche calculations.
    • Moore's Law has already reduced the cost of computation by billions of times, meaning current additions of compute (e.g., H100s) are still a negligible fraction of total human cognitive capacity.

Addressing Objections on Intelligence and Construction

  • Declining Returns on Intelligence:
    • Skepticism suggests that making AI 100x smarter than humans yields negligible improvements in chaotic forecasting (e.g., weather) or complex engineering advice.
    • Rebuttal: This argument only limits the depth of specific predictions, not the breadth of duplication.
      • Just as Japan's larger population supports a larger GDP than Singapore, a trillion AI instances can process correspondingly more tasks even if each instance isn't infinitely smarter.
      • Human wage data shows that increased education/capability yields significant (2x–16x) returns, suggesting non-linear gains are possible up to very high thresholds.
  • Construction and Regulatory Bottlenecks:
    • Skeptics question how physical construction could scale 100x–1000x given historical regulatory delays (NIMBYism).
    • Historical Precedent: Physical production per capita has already grown massively compared to pre-industrial eras.
    • Regulatory Arbitrage:
      • High payouts from AI-driven industrial output create incentives for jurisdictions (e.g., UAE, China) to become "regulatory havens."
      • Governments may relax zoning and permitting to capture massive tax revenue or military strategic advantages.
      • AI can reduce governance costs via "unlimited cheap lawyers" and AI bureaucrats, simplifying complex regulatory navigation.

Income Distribution and Wealth in the Post-AI Era

  • Employment Concerns: While humans may lose the ability to earn wages via labor, Schulman argues poverty is unlikely to persist.
  • Redistribution Mechanisms:
    • Existing government redistribution models (e.g., 50% of GDP in some nations) applied to a trillion-fold increase in total output would result in billionaire-level consumption for all citizens.
    • Universal Basic Income (UBI) or pension plans would become economically feasible at levels far exceeding current elite wealth.
  • Resource Distribution:
    • Inequality Risks: Unequal land distribution (e.g., Singapore vs. Canada) creates baseline inequality.
    • Mitigation via Global Resources:
      • Most Earth's resources (offshore, atmospheric energy limits) are global in nature; allocation could be based on population or heat-emission quotas.
      • Off-world resources are so vast that even equitable distribution among existing nation-states would eliminate absolute poverty, hunger, and lack of access to medicine.
    • Asset Inflation: Old capital goods may lose value relative to natural resource entitlements, shifting the wealth basis from physical ownership to access to AI/energy.