Towards a Formal Scientific Epistemology
The author proposes a formal scientific epistemology, contrasting it with Bayesian approaches that assign binary truth values to propositions. Instead, the author advocates for assigning degrees of truth to models, drawing inspiration from scientific practices where new theories are constructed to explain data and make novel predictions. Garrabrant induction is presented as a significant step towards formalizing this scientific epistemology, using a market mechanism where polynomial-time algorithms act as traders setting credences for logical statements based on their predictive success. AI
IMPACT Proposes a new framework for reasoning that could influence AI alignment research and development.