I think the thing @jacob_cannell is imagining is not plain CDT, EDT, or FDT, and writing out what it is he’s imagining in the language of https://arxiv.org/abs/2307.10987 would clarify. I suspect the RL thing he’s imagining is some mix of CDT and EDT depending on the amount of experience the agent has with a context. He’d have to clarify. I bring this up because I anticipate any language model having the correct response to that example scenario, because it has experience with those dynamics in previous language, but it’ll be vulnerable to tweaked versions of that, and yet also behave CDTishly in some scenarios. these decision theories are “pure”, approximation-free models, and so approximation learning systems behave differently sometimes.
I think the thing @jacob_cannell is imagining is not plain CDT, EDT, or FDT, and writing out what it is he’s imagining in the language of https://arxiv.org/abs/2307.10987 would clarify. I suspect the RL thing he’s imagining is some mix of CDT and EDT depending on the amount of experience the agent has with a context. He’d have to clarify. I bring this up because I anticipate any language model having the correct response to that example scenario, because it has experience with those dynamics in previous language, but it’ll be vulnerable to tweaked versions of that, and yet also behave CDTishly in some scenarios. these decision theories are “pure”, approximation-free models, and so approximation learning systems behave differently sometimes.