Elias Schmied: FDT debates miss the metaphysical crux
Decision theory fights ignore the real issue—alternate realities, time travel, or platonism.
Elias Schmied’s LessWrong post ‘Stop doing decision theory without metaphysics’ targets the ongoing FDT debates. He uses the classic blackmail scenario: Omega predicts you’ll pay, so you get blackmailed. FDT says don’t pay because FDT agents wouldn’t have been blackmailed in the first place. But that reasoning only holds if you have time to precommit. For an AI instantly blackmailed, FDT’s recommendation requires a metaphysical foundation. Schmied identifies three possible metaphysics: alternate realities (you affect other branches), retroactive causality (your decision makes the blackmail impossible), or platonism (your decision process determines the physical world). These are the actual cruxes, yet comment threads under related posts (e.g., by Bentham’s Bulldog, with Scott Alexander and Stuart Armstrong) rarely discuss metaphysics. Instead, they argue over the meaning of ‘rationality’ or confuse precommitment with FDT. Schmied insists that the only real divergence between FDT and normal decision theory is in scenarios without time to self-modify—and that’s where metaphysics matters.
- FDT requires one of three metaphysical commitments: alternate realities, retroactive causality, or platonism.
- Current debates focus on semantic disagreements rather than these core metaphysical assumptions.
- The blackmail scenario is the key test case where FDT and standard decision theory truly differ.
Why It Matters
Shows that AI decision theory debates must engage metaphysics, not just utility comparisons.