Huh, I did not realize how close our views are. Thank you for posting this. Did you have problems reading the Sequences where the Correspondence Dualism (reality vs models) is implicit and relied upon?
I would add a couple of points.
One is AIXI-like: the “naive” Correspondence Theory of Truth presumes (an?) external reality which we strive to understand and therefore affect, like a perturbation theory. Sort of like in Special Relativity we postulate that the spacetime is flat and unchanging. However, in many situations “reality” is determined by how we model it, like the market sentiment determines market behavior, what Eliezer called the anti-inductive property. This reminds me of General Relativity (or at least its Initial Value formulation), where there is no spacetime until you build it out of matter. Sometimes, when the interaction between spacetime and matter is weak, you can use the background calculations, such as the parametric post-Newtonian expansion, or Quantum Field Theory on a curved spacetime background. But this approach breaks down when the interactions are strong and the perturbations do not converge, resulting in non-nonsensical results, such as the Hawking radiation blowing up at the black hole horizon. But, as long as you stay away from the highly nonlinear cases, it makes sense to use the concept of External Reality and the Correspondence Theory of Truth.
Another is the question “what are you predicting, if there is no reality?” Where do the inputs you use to build your models and estimate their success come from? You call it
a brute, basic constraint on effective action.
But this is just a poetic way of saying the same thing: there are stimuli coming from outside of our hierarchy of models, at least down at the bottom level. These are definitely affected by our actions, which are mostly interactions within the hierarchy of models, but at some point leak outside of it and steer future stimuli in the desired direction (if the model hierarchy is successful in its goals). So, either you assume this invisible feedback loop and call it reality, or you bite another bullet and assume that there are models all the way down. Given that we have not yet found any barrier to building our hierarchy of models, I tend toward the latter, but I am not super-convinced that this is a better meta-model than the former.
Huh, I did not realize how close our views are. Thank you for posting this. Did you have problems reading the Sequences where the Correspondence Dualism (reality vs models) is implicit and relied upon?
I would add a couple of points.
One is AIXI-like: the “naive” Correspondence Theory of Truth presumes (an?) external reality which we strive to understand and therefore affect, like a perturbation theory. Sort of like in Special Relativity we postulate that the spacetime is flat and unchanging. However, in many situations “reality” is determined by how we model it, like the market sentiment determines market behavior, what Eliezer called the anti-inductive property. This reminds me of General Relativity (or at least its Initial Value formulation), where there is no spacetime until you build it out of matter. Sometimes, when the interaction between spacetime and matter is weak, you can use the background calculations, such as the parametric post-Newtonian expansion, or Quantum Field Theory on a curved spacetime background. But this approach breaks down when the interactions are strong and the perturbations do not converge, resulting in non-nonsensical results, such as the Hawking radiation blowing up at the black hole horizon. But, as long as you stay away from the highly nonlinear cases, it makes sense to use the concept of External Reality and the Correspondence Theory of Truth.
Another is the question “what are you predicting, if there is no reality?” Where do the inputs you use to build your models and estimate their success come from? You call it
But this is just a poetic way of saying the same thing: there are stimuli coming from outside of our hierarchy of models, at least down at the bottom level. These are definitely affected by our actions, which are mostly interactions within the hierarchy of models, but at some point leak outside of it and steer future stimuli in the desired direction (if the model hierarchy is successful in its goals). So, either you assume this invisible feedback loop and call it reality, or you bite another bullet and assume that there are models all the way down. Given that we have not yet found any barrier to building our hierarchy of models, I tend toward the latter, but I am not super-convinced that this is a better meta-model than the former.