Operationalizing FDT

A new research effort aims to operationalize Functional Decision Theory (FDT), a framework for AI decision-making that emphasizes logical causality over physical causality. The core innovation is defining a 'logical do-operator' that allows AI agents to reason about counterfactuals in logical causal graphs. Unlike traditional causal decision theory's do-operator, which cuts incoming connections to action nodes, the logical version must handle scenarios where downstream logical facts can be observed before decisions are made—as in Parfit's hitchhiker problem where an agent's algorithm determines whether they get rescued.

The research presents four possible definitions for this logical do-operator in a 2x2 matrix based on whether to cut incoming connections and whether to forget downstream nodes. Options 1 and 3 (which don't forget downstream nodes) fail to produce the correct FDT behavior in classic problems. The debate centers on whether to choose option 2 (cut connections AND forget downstream nodes) or option 4 (just forget downstream nodes), both of which enable FDT agents to 'automatically' make optimal decisions in logical dependency scenarios without requiring additional commitment mechanisms.