<div> Misaligned or Undersupplied?  </div> <div> Pure Mathematics in the Age of LLMs </div>

Source abstract

Twenty-five Fields medalists have declared artificial intelligence and mathematics misaligned. I observe that the wait between a mathematical result and its use was in large part transmission: a result reached a use only when someone along the way knew both the problem and the theorem, and a language model can now be that person at every use. In a minimal model of reach, adoption cost, and lag, the value of a marginal result rises by a multiple and the wedge between social and private returns widens. Before language models could match and implement, elementary economics already said underinvestment was likely. If language models remove a substantial part of a delay measured in decades, it becomes very hard to argue that the present level could be optimal, and the observed allocation is moving the other way.