In defense of oracle ("tool") AI research

Abstract

The article defends the idea of non-self-improving AI systems designed to only answer questions without taking actions (oracle AIs) as opposed to agent AIs that additionally have the capacity to take actions, send emails, and control actuators. The author claims that oracle AIs are safer and easier to coordinate and proposes that humanity should focus on developing them rather than riskier agent AIs. The author also argues that the coordination problem (i.e., the problem of preventing powerful AIs from endangering humanity) is not a compelling argument against oracle AIs because it is harder to solve the coordination problem with agent AIs, because there will always be incentives for people to develop more powerful and less safe AGIs, and because coordination will have to be solved at some point in AI development anyway. – AI-generated abstract.