Pragmatic FDT, and predictors as game theory

2026-07-03 · Source: AI Alignment Forum · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Stuart Armstrong introduces "Pragmatic FDT" (p-FDT), a variant of Functional Decision Theory designed to address theoretical pitfalls identified by critics like Will MacAskill and Bentham's Bulldog. Published on July 3rd, 2026, the post outlines p-FDT's four-step decision process: computing a Causal Decision Theory (CDT) baseline, searching for "likely-true" exploitable isomorphisms between the agent's decision process and external world elements, evaluating these for expected utility advantage, and adopting the best or defaulting to CDT. This approach avoids abstract definitions of algorithmic identity by focusing on practical, verifiable isomorphisms. Armstrong further argues that counterfactual predictors fundamentally shift decision theory into game theory, explaining why FDT's seemingly irrational outcomes, such as in the Bomb thought experiment or refusing a blackmailer, are actually game-theoretic policy costs. He also notes that CDT inherently struggles to account for reliable self-predictors.

Key takeaway

For AI Scientists designing advanced decision agents, you should consider implementing a pragmatic FDT approach that prioritizes verifiable, exploitable isomorphisms over abstract algorithmic identity. This allows your agents to effectively navigate scenarios with reliable predictors, which often become game-theoretic interactions. Be prepared for your agent's optimal policy to incur costs in specific, low-probability branches, as this is a common trade-off in game theory, not a flaw in rationality.

Key insights

Pragmatic FDT reframes decision-making by identifying exploitable isomorphisms between an agent's process and the world.

Principles

• Algorithmic identity can be established via "likely-true" isomorphisms.
• Counterfactual predictors transform decision theory into game theory.
• Decision policies may incur costs in specific "branches".

Method

p-FDT involves a four-step process: compute CDT baseline, search for exploitable isomorphisms, evaluate their expected utility advantage, then adopt the best or default to CDT.

In practice

• Identify isomorphisms using identity, code, or behavioral equivalence.
• Use black-box testing to validate predictor claims.
• Consider game theory when facing counterfactual predictors.

Topics

• Functional Decision Theory
• Pragmatic FDT (p-FDT)
• Game Theory
• Causal Decision Theory
• Predictor Models
• Algorithmic Isomorphisms

Best for: AI Scientist, Research Scientist

Related on AIssential

• 49 - Caspar Oesterheld on Program Equilibrium — Program equilibrium enables cooperation in games by allowing AI agents to analyze and simulate each other's source code.
• Playing Games With TheoryCoder | ARC Prize @ MIT — Theory-based reinforcement learning with hierarchical abstraction offers scalable, human-like efficiency for complex sequential decision problems.
• Enhancing Decision-Making with Large Language Models through Multi-Agent Fictitious Play — Iterative fictitious play with LLM agents resolves "stance entanglement" in complex decision-making by decomposing recursive anticipation into sequential best responses.
• Unveiling the Structure of Do-Calculus Reasoning via Derivation Graphs — Derivation graphs simplify do-calculus reasoning, revealing equivalent causal expressions and enabling more efficient estimation.
• Games people — and machines — play: Untangling strategic reasoning to advance AI — AI can achieve superhuman strategic reasoning in complex games with imperfect information through advanced game theory and optimization.
• GT-HarmBench: Benchmarking AI Safety Risks Through the Lens of Game Theory — AI models struggle with multi-agent coordination, but game-theoretic interventions can improve socially optimal outcomes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Alignment Forum.