My Empire of Sunshine and Roses

Open original

George Hotz argues that you can only measure your impact on the world by asking how history would differ without you, and he doubts AI will collapse into a single dominant player. He prefers a future of many independent power centers with fair, linear-cost markets, where no central planner decides the outcome.

  • Real impact means counterfactuals: ask how history would look if you hadn't existed, not how much credit you collect.
  • AI scaling hits diminishing returns, since paying 10x for 2x speed or 100x for 3x speed breaks down quickly, which undercuts the singleton scenario.
  • The key thing to watch is whether offense or defense gains the edge in security and cryptography, a question he admits he can't yet answer.
  • He favors millions of independent power centers that cooperate or impose small costs on each other, rather than monopolies or vertical integration.

The author, who writes as geohot, opens with a framing claim: changing the world can only be measured counterfactually, asking how history would have differed without you. Credit assignment is hard enough on its own, he argues, so the better test is what would have happened otherwise. He then sketches two views of history, a hard determinism where the outcome was fixed and nobody can really move anything, and a "Great Man" view versus "Causes and Forces" view that he suggests mostly comes down to how far back you look. He leans toward thinking AI's trajectory is already set by deep structural variables, like the balance of offense and defense and the pull between centralization and decentralization, which he calls features of a "technological optimization landscape." He still says "we're here," meaning the choices matter in practice even if the map is fixed. Next he gets specific about where he thinks the industry's stories are wrong. He watched an OpenAI ad showing an office, while the real work happens in a datacenter thousands of miles away, and he asks why the office would persist at all. On Anthropic he is skeptical of the singleton scenario that science fiction has trained people to expect, pointing out that nature has not collapsed into one dominant actor and that most AI systems have exponentially rising resource costs for linear gains. His concrete example is agent swarms: is spending ten times as much to finish work twice as fast worth it, and is a hundred times the spend for a threefold speedup sensible? He thinks returns break down at some point, and he is more interested in how attacker versus defender advantage is shifting, especially in security and cryptography, than in dramatic bad-outcome scenarios. He admits he cannot tell which side, the individual or the state, that shift favors. His positive vision is modest and structural. The minimum he considers acceptable is fair markets that resist monopolies, internal models, and vertical integration, with linear spending for linear compute, and producing more than you consume. His ideal is a world of millions of independent power centers that cooperate when they choose and impose small costs on each other when they don't, with no central planner deciding what the result should look like. He ends by framing that decentralized, emergent order as the real American Dream.

Ask about this article