"Prosaic AI alignment": 10 years later
Looking back a decade after the publication of Paul Christiano's foundational text on prosaic AI alignment, the author argues that while we have successfully built powerful systems using reinforcement learning and prediction tasks without solving alignment, the core problem remains a difficult and unsolved challenge. The piece asserts that prosaic AI alignment should remain a primary focus for researchers because it is tractable today and shares crucial similarities with whatever future AGI architecture might emerge. Open article →