If you one-box on Newcomb’s Problem, then it’s plausibly important for realistic (i.e. smart but non-omniscient) counterparties to be able to predict that you’d one-box.
IIRC I’ve never publicly written about Newcomb’s Problem, but Claude Opus 4.7 guessed an 80% chance that I one-box, 10% that I two-box, and 10% “other”, based on circumstantial evidence (the fact that I’m involved in the EA/rationalist scene, and I once linked to Zvi talking about MIRI’s paper on functional decision theory).
Gemini Thinking also predicted that I’d one-box, but it presented fake evidence—it hallucinated that I wrote a blog post called “A Critique of Causal Decision Theory”, and even hallucinated a (broken) link to the post.
ChatGPT predicted 70–80% chance that I’d one-box, again based on circumstantial evidence.
(I only get to run this experiment once, because it’s now public record that I’d one-box. But I’m fine with that because all three LLMs got the right answer, so surely future LLMs will also get the right answer.)
Claude and Gemini both said I probably subscribe to FDT (although Claude was uncertain). I haven’t studied enough decision theory to have a strong belief about which is correct (if any), except that CDT is not it. I do have beliefs about specific thought experiments:
You should one-box, which rules out CDT
You should pay Parfit’s Hitchhiker, which rules out CDT and possibly EDT
You should pay in the Counterfactual Mugging (although the logic is complicated so I’m less sure about this one)
It’s plausible that no as-yet-proposed decision theory is viable, a la Vibe Decision Theory.
(For the avoidance of doubt: I don’t think “choose the answer with the best vibes” is a good or even coherent decision theory, but I do think the VDT post is a clean analysis of the shortcomings of various decision theories.)
If you one-box on Newcomb’s Problem, then it’s plausibly important for realistic (i.e. smart but non-omniscient) counterparties to be able to predict that you’d one-box.
IIRC I’ve never publicly written about Newcomb’s Problem, but Claude Opus 4.7 guessed an 80% chance that I one-box, 10% that I two-box, and 10% “other”, based on circumstantial evidence (the fact that I’m involved in the EA/rationalist scene, and I once linked to Zvi talking about MIRI’s paper on functional decision theory).
Gemini Thinking also predicted that I’d one-box, but it presented fake evidence—it hallucinated that I wrote a blog post called “A Critique of Causal Decision Theory”, and even hallucinated a (broken) link to the post.
ChatGPT predicted 70–80% chance that I’d one-box, again based on circumstantial evidence.
(I only get to run this experiment once, because it’s now public record that I’d one-box. But I’m fine with that because all three LLMs got the right answer, so surely future LLMs will also get the right answer.)
Claude and Gemini both said I probably subscribe to FDT (although Claude was uncertain). I haven’t studied enough decision theory to have a strong belief about which is correct (if any), except that CDT is not it. I do have beliefs about specific thought experiments:
You should one-box, which rules out CDT
You should pay Parfit’s Hitchhiker, which rules out CDT and possibly EDT
You should pay in the Counterfactual Mugging (although the logic is complicated so I’m less sure about this one)
It’s plausible that no as-yet-proposed decision theory is viable, a la Vibe Decision Theory.
(For the avoidance of doubt: I don’t think “choose the answer with the best vibes” is a good or even coherent decision theory, but I do think the VDT post is a clean analysis of the shortcomings of various decision theories.)