But... it did work pretty well for nuclear weapons. In the early 1950s, the US did not preemptively strike the USSR despite it being game theoretic optimal [1].

But... it did work pretty well for nuclear weapons. In the early 1950s, the US did not preemptively strike the USSR despite it being game theoretic optimal [1]. Then while there were some crises, we successfully passed a series of international treaties that placed limits on the arms race between the US and USSR (SALT I and II), and even began to dismantle nuclear weapons stockpiles (START I and II, INF). We successfully went from a peak of over 60,000 nuclear warheads deployed to 10,000. Regarding the "barrier to entry" for AI, this is not really true. Training a frontier AI model takes an enormous amount of compute using chips that have a supply chain with several chokepoints, like EUV machines and HBM which are only made by a few firms. [1]: https://cs.stanford.edu/people/eroberts/courses/soco/projects/1998-99/game-theory/neumann.html#:~:text=%22If%20you%20say%20why%20not%20bomb%20them%20tomorrow%2C%20I%20say%20why%20not%20today%3F%20If%20you%20say%20today%20at%205%20o%27clock%2C%20I%20say%20why%20not%20one%20o%27clock%3F%22 https://cs.stanford.edu/people/eroberts/courses/soco/project... [2]: https://www.atomicarchive.com/resources/treaties/index.html https://www.atomicarchive.com/resources/treaties/index.html