⚖️AI Alignment Forum•Mar 13, 2026Stalecollected in 10m

Operationalizing FDT

Post LinkedIn

⚖️Read original on AI Alignment Forum

#decision-theory #logical-causality #do-operatorfdtfdt parfit-hitchhiker ai-alignment-forum

💡Formalizes FDT's logical do-operator—essential for building predictor-proof AI agents.

⚡ 30-Second TL;DR

What Changed

Defines logical do-operator via 2x2 table with cut/forget options for logical causal graphs.

Why It Matters

Advances FDT formalization, aiding robust AI agent design in predictor scenarios. Helps alignment researchers implement decision theories without commitment hacks.

What To Do Next

Implement option 2 logical do-operator in your FDT agent simulator for hitchhiker tests.

Who should care:Researchers & Academics

Key Points

•Defines logical do-operator via 2x2 table with cut/forget options for logical causal graphs.
•Option 2 (cut incoming, forget downstream, condition) enables FDT to pay in Parfit's hitchhiker.
•Distinguishes logical causality from physical, requiring forget step for observed downstream nodes.
•Explains FDT vs EDT/CDT, addressing anthropic updating and need for logical counterfactuals.

🧠 Deep Insight

Background and context from public sources — not the original article. 8 sources cited.

🔑 Enhanced Key Takeaways

•FDT was formally introduced by Eliezer Yudkowsky and Nate Soares as a successor to Timeless Decision Theory (TDT), outperforming CDT and EDT by treating decisions as outputs of a fixed mathematical function[3].
•ACDT, a related acausal approach, extends CDT by adding potential logical links from the decision node to other nodes in causal graphs, enabling one-boxing in Newcomb's problem through empirical learning of graph structures[1].
•FDT is characterized as a meta-causal theory emphasizing subjunctive dependence via source code correlations that resist confounding by choice, avoiding risks like dynamic updating exploited by predictors[6].
•Critiques highlight FDT's vulnerability in adversarial settings, such as XOR blackmail, where predictors might manipulate agents into switching to exploitable decision theories like EDT[6].

🔮 Future ImplicationsAI analysis grounded in cited sources

FDT agents will dominate in logically correlated multi-agent environments by 2030

FDT's use of subjunctive dependence on shared decision functions enables cooperation without communication, outperforming CDT/EDT in Newcomb-like scenarios as predictors improve[3][6].

Operationalized logical do-operators will standardize FDT implementations in AI by 2028

Defining do-operators via causal graphs with cut/forget mechanisms resolves Parfit's hitchhiker, providing implementable algorithms for robust acausal trade[1][2].

⏳ Timeline

2017-04

Yudkowsky and Soares publish Functional Decision Theory on LessWrong, replacing TDT[3]

2018-01

Soares and Levinstein release 'Cheating Death in Damascus,' showcasing FDT counterexamples to EDT/CDT[8]

2020-10

LessWrong post dissolves FDT confusions, framing it as meta-causal theory[6]

2021-05

AI Alignment Forum introduces ACDT using causal graphs for acausal links[1]

2022-03

Coester publishes Tickle Defense analysis with logical causal graphs for Newcomb problems[2]

📎 Sources (8)

Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.

⚖️Read original article on AI Alignment Forum

📰

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Related Updates

Same topic

Explore #decision-theory

Same product