FDT's defense: Critics attack a strawman, not the actual theory
Why most objections to Functional Decision Theory are irrelevant.
Fernand0's LessWrong post 'Criticism against "unembedded FDT" doesn't apply to FDT' clarifies a persistent confusion about Functional Decision Theory (FDT). FDT is designed for rational agents embedded in a shared universe, where predictions about the agent's actions are causally connected to the agent's internal 'code.' This embeddedness is essential: an agent cannot know for certain that it is not in a simulation, because that would require a perspective outside the entire universe—a contradiction for an embedded agent. Many standard thought experiments (e.g., Newcomb's problem) implicitly assume an unembedded agent who knows they are in the 'real' world, which violates FDT's premises. Fernand0 illustrates with a 'Bomb' scenario where Omega predicts the agent's action with near-perfect accuracy. Critics assume the agent knows they are not in a simulation, but under embedded FDT, the only remaining explanation for a mismatch is that the agent is in a counterfactual simulation (e.g., Omega's hallucination). Thus, any criticism based on unembedded assumptions does not apply to actual FDT. The post urges readers to embrace embeddedness, however difficult, to properly evaluate the theory.
- FDT only applies to agents embedded in their universe; they cannot know they are not in a simulation.
- Common criticisms assume an unembedded 'God's-eye view' that contradicts FDT's foundations.
- Fernand0 provides an embedded reformulation of the Bomb scenario to show FDT's consistent response.
Why It Matters
Clears up a fundamental misinterpretation of FDT, strengthening its relevance for AI safety and decision theory.