In context of boundaries/membranes, I’ve been thinking of how a lot of spurious proof issues involve a search that’s greedy. It can be as trivial as an agent A that encounters “Because A()=X” as an answer to “Why should I do X?”, which is clearly an irrelevant point to consider (pebblesorters don’t pursue primeness morally-because they are pebblesorters, even as they pursue it approximately causally-because they are pebblesorters).
Perhaps a reasoner should be bounded in an even stronger sense than limits on how hard it’s allowed to search, by instead working in an environment of carefully curated statements, with some epistemic membrane deciding which statements to channel (when) for further consideration. Anything else would expose the reasoner to its crashspace, things it’s not ready to either see or willfully ignore (there’s chicken rule for ignoring claims about the current action, but there are other claims that might also need to be ignored, and it’s less clear how to do that).
(This is also my model of how pseudokindness could work, offering bespoke-curated options intended as not being from the crash space, without directly optimizing the patient.)
In context of boundaries/membranes, I’ve been thinking of how a lot of spurious proof issues involve a search that’s greedy. It can be as trivial as an agent A that encounters “Because A()=X” as an answer to “Why should I do X?”, which is clearly an irrelevant point to consider (pebblesorters don’t pursue primeness morally-because they are pebblesorters, even as they pursue it approximately causally-because they are pebblesorters).
Perhaps a reasoner should be bounded in an even stronger sense than limits on how hard it’s allowed to search, by instead working in an environment of carefully curated statements, with some epistemic membrane deciding which statements to channel (when) for further consideration. Anything else would expose the reasoner to its crash space, things it’s not ready to either see or willfully ignore (there’s chicken rule for ignoring claims about the current action, but there are other claims that might also need to be ignored, and it’s less clear how to do that).
(This is also my model of how pseudokindness could work, offering bespoke-curated options intended as not being from the crash space, without directly optimizing the patient.)