In retrospect, the most important paper in the economics and geopolitics of AI might be Stanford's Alpaca (03/2023). It was the first proof that you could distill OpenAI’s text-davinci-003 into a 7B model for <$600. This was a very surprising result. But it wasn't fully appreciated at the time what this would mean. Now this phenomenon is why Washington and Beijing are putting export controls on AI models, splintering the global market for intelligence. They don't want their adversaries distilling their models. But it's almost impossible to stop, because it doesn't take that many customer accounts (either smuggled or in a third country) to be able to pull enough traces to distill someone else's model. Distillation is like gravity. It pulls every capability advantage down toward everyone else until eventually the landscape is flattened, whether the labs like it or not. It was not obvious distillation would work so well. But it does. Like the gravitational constant, Alpaca proved that we live in a universe with a good distillation constant, and therefore monopolies on intelligence would be temporary. As a result, Intelligence is getting cheaper and cheaper. This is good.