This post was educational, however, I want to push back against the implicit criticism of instrumentalism and the Copenhagen interpretation. The metaphilosophical position I will use here is: to solve a philosophical question, we need to rephrase it as a question about AI design (AFAIK the full credit for this approach goes to Yudkowsky). So, suppose I am building an AI. Should the AI’s model of the world be (i) “about things that exist independently of the AI” or (ii) “about the subjective perceptions of the AI”? This depends on what kind of reward function I want my AI to have.
The standard (I call it “perceptible”) type of reward function in reinforcement learning only depends on the history of actions and observations. For such an AI the answer is (ii): it is important the AI will correctly predict the consequences of its actions, but there is no significance whatsoever that the AI’s models can be interpreted as (i). Yes, these models will still have intrinsic variables corresponding to “unobserved objects” in some sense, but there is no reason for these variables to always have an unambiguous “realist” interpretation.
Now suppose that the AI is actually designed to care about particular objects outside itself. Specifically, assume the AI uses an instrumental (or semi-instrumental) reward function. Such a function might be specified (or partially specified) using some particular ontology. Then, the AI is modeling the world as containing certain unobserved objects, at least approximately. The answer is now in between (i) and (ii). The AI’s models are about subjective perceptions of the AIs and also about some particular type of things that exist independently of the AI, namely, those things that are normatively important to it. Coming back from AIs to humans, we may conclude that, what really makes sense is our subjective perceptions + those external objects that we actually care about (e.g. other humans). But, quarks only “exist” in the sense that they are components in a useful model we created.
What does it tell us about the interpretation of quantum mechanics? Once again, consider an AI trying to discover quantum mechanics. From the AI’s perspective, what it’s looking for is a function from observation histories to distributions over the next observation. How can we construct such a function from the formalism of QM? Obviously, using Copenhagen: each observation is a measurement that causes the wavefunction of the environment to collapse. “But,” the fans of MWI will object “what about all those other Everett branches? They are still out there, right? They don’t actually vanish?!” The point is, from the AI’s perspective the question is meaningless. The AI’s design assumes the AI can record everything it observes in its memories, therefore, once an observation is made, those Everett branches will never meet again in the AI’s lifetime. “But,” another objection may go “what if someone tempers with the memories of the AI in a way that allows quantum interference between the branches? Sure it is completely impractical for humans, but it is theoretically possible, and might even be practically possible for an AI running on a quantum computer.” Alright, but tempering with the memory of the agent is effectively destroying the agent: it invalidates the fundamental assumptions of its reasoning algorithm, and any reasoning algorithm must make some fundamental assumptions of that sort. (The agent might still accurately account for the possibility of its own destruction (see “The Death of the Agent and Kamikaze Strategies”), but probably only with the help of external knowledge.)
This post was educational, however, I want to push back against the implicit criticism of instrumentalism and the Copenhagen interpretation. The metaphilosophical position I will use here is: to solve a philosophical question, we need to rephrase it as a question about AI design
Maybe the problem is that I’m not sufficiently convinced that there’s a philosophical question here. Sometimes philosophers (and even physicists) argue about things that aren’t open questions. “Do refrigerators exist, or only mental models of refrigerators?” sounds like a straightforward, testable empirical question to me, with all the evidence favoring “refrigerators exist”.
I predict I’m missing an implicit premise explaining why “I don’t currently understand where the Born rule comes from” is a bigger problem for realism than “I don’t currently understand how my refrigerator works”, or some other case where realism makes things unnecessarily hard/confusing, like infinite ethics or anthropics or somesuch.
Let’s unpack what it means to say that “refrigerators exist”. From my (instrumentalist) perspective, it means that (i) I have a predictive model of my perception, to which I assign high credence, and in which for each state of the environment I can say which refrigerators exist (where “refrigerator” is just some symbol that makes sense inside the model) and (ii) according to my belief, the current state of the environment contains at least one refrigerator with high probability.
My claim is not that quantum mechanics proves realism is wrong. My claim is that instrumentalism is the correct metaphysics regardless, and once you accept that, the Copenhagen interpretation seems quite satisfactory. Although it is also true that if you try interpreting quantum mechanics according to sufficiently strong realist desiderata you run into impossibility results like the Kochen-Specker theorem and the violation of Bell’s inequality.
My claim is that instrumentalism is the correct metaphysics regardless
What does it mean for instrumentalism to be the correct metaphysics? Normally, I’d interpret “the correct metaphysics” as saying something basic about reality or the universe. (Or, if you’re an instrumentalist and you say “X is the correct metaphysics”, I’d assume you were saying “it’s useful to have a model that treats X as a basic fact about reality or the universe”, which also doesn’t make sense to me if X is “instrumentalism”.)
Although it is also true that if you try interpreting quantum mechanics according to sufficiently strong realist desiderata
Well, sufficiently specific realist desiderata. Adding hidden variables to QM doesn’t make the theory any more realist, the way we’re using “realist” here.
What does it mean for instrumentalism to be the correct metaphysics? Normally, I’d interpret “the correct metaphysics” as saying something basic about reality or the universe. (Or, if you’re an instrumentalist and you say “X is the correct metaphysics”, I’d assume you were saying “it’s useful to have a model that treats X as a basic fact about reality or the universe”, which also doesn’t make sense to me if X is “instrumentalism”.)
Like I said before, it means that instrumentalism is the point of view that is the most useful for designing AI or answering questions about AI. According to the “Yudkowskian computationalist” metaphilosophical view, this also makes it the most useful for rationality in general.
Adding hidden variables to QM doesn’t make the theory any more realist, the way we’re using “realist” here.
I imagined “realist” to mean something like “the universe can be described in a way independent of the choice of observer, and the perceptions of any given observer can be decoded from the history of the universe in this description, s.t. different observers have compatible observations”. Adding hidden variables does make QM more realist in this sense, for example the de Broglie-Bohm interpretation is realist (but it only makes sense if you assume all observer perceptions can be reduced to configuration variables, which seems false and disqualifies it). MWI fails to be entirely realist because you have to either make the decoding of observer perceptions stochastic (and thereby admit that your description of the universe is incomplete) or to postulate, for each “normal” observer Alice, a huge ensemble of different observers that correspond to versions of Alice in different Everett branches (and thereby lose the physical meaning of quantum probabilities and with it essentially all the predictive power of the theory).
Like I said before, it means that instrumentalism is the point of view that is the most useful for designing AI or answering questions about AI. According to the “Yudkowskian computationalist” metaphilosophical view, this also makes it the most useful for rationality in general.
Except that if you are the kind of rationalist who cares about what is really real, you should reject instrumentalism immediately.
This post was educational, however, I want to push back against the implicit criticism of instrumentalism and the Copenhagen interpretation. The metaphilosophical position I will use here is: to solve a philosophical question, we need to rephrase it as a question about AI design (AFAIK the full credit for this approach goes to Yudkowsky). So, suppose I am building an AI. Should the AI’s model of the world be (i) “about things that exist independently of the AI” or (ii) “about the subjective perceptions of the AI”? This depends on what kind of reward function I want my AI to have.
The standard (I call it “perceptible”) type of reward function in reinforcement learning only depends on the history of actions and observations. For such an AI the answer is (ii): it is important the AI will correctly predict the consequences of its actions, but there is no significance whatsoever that the AI’s models can be interpreted as (i). Yes, these models will still have intrinsic variables corresponding to “unobserved objects” in some sense, but there is no reason for these variables to always have an unambiguous “realist” interpretation.
Now suppose that the AI is actually designed to care about particular objects outside itself. Specifically, assume the AI uses an instrumental (or semi-instrumental) reward function. Such a function might be specified (or partially specified) using some particular ontology. Then, the AI is modeling the world as containing certain unobserved objects, at least approximately. The answer is now in between (i) and (ii). The AI’s models are about subjective perceptions of the AIs and also about some particular type of things that exist independently of the AI, namely, those things that are normatively important to it. Coming back from AIs to humans, we may conclude that, what really makes sense is our subjective perceptions + those external objects that we actually care about (e.g. other humans). But, quarks only “exist” in the sense that they are components in a useful model we created.
What does it tell us about the interpretation of quantum mechanics? Once again, consider an AI trying to discover quantum mechanics. From the AI’s perspective, what it’s looking for is a function from observation histories to distributions over the next observation. How can we construct such a function from the formalism of QM? Obviously, using Copenhagen: each observation is a measurement that causes the wavefunction of the environment to collapse. “But,” the fans of MWI will object “what about all those other Everett branches? They are still out there, right? They don’t actually vanish?!” The point is, from the AI’s perspective the question is meaningless. The AI’s design assumes the AI can record everything it observes in its memories, therefore, once an observation is made, those Everett branches will never meet again in the AI’s lifetime. “But,” another objection may go “what if someone tempers with the memories of the AI in a way that allows quantum interference between the branches? Sure it is completely impractical for humans, but it is theoretically possible, and might even be practically possible for an AI running on a quantum computer.” Alright, but tempering with the memory of the agent is effectively destroying the agent: it invalidates the fundamental assumptions of its reasoning algorithm, and any reasoning algorithm must make some fundamental assumptions of that sort. (The agent might still accurately account for the possibility of its own destruction (see “The Death of the Agent and Kamikaze Strategies”), but probably only with the help of external knowledge.)
Maybe the problem is that I’m not sufficiently convinced that there’s a philosophical question here. Sometimes philosophers (and even physicists) argue about things that aren’t open questions. “Do refrigerators exist, or only mental models of refrigerators?” sounds like a straightforward, testable empirical question to me, with all the evidence favoring “refrigerators exist”.
I predict I’m missing an implicit premise explaining why “I don’t currently understand where the Born rule comes from” is a bigger problem for realism than “I don’t currently understand how my refrigerator works”, or some other case where realism makes things unnecessarily hard/confusing, like infinite ethics or anthropics or somesuch.
Let’s unpack what it means to say that “refrigerators exist”. From my (instrumentalist) perspective, it means that (i) I have a predictive model of my perception, to which I assign high credence, and in which for each state of the environment I can say which refrigerators exist (where “refrigerator” is just some symbol that makes sense inside the model) and (ii) according to my belief, the current state of the environment contains at least one refrigerator with high probability.
My claim is not that quantum mechanics proves realism is wrong. My claim is that instrumentalism is the correct metaphysics regardless, and once you accept that, the Copenhagen interpretation seems quite satisfactory. Although it is also true that if you try interpreting quantum mechanics according to sufficiently strong realist desiderata you run into impossibility results like the Kochen-Specker theorem and the violation of Bell’s inequality.
What does it mean for instrumentalism to be the correct metaphysics? Normally, I’d interpret “the correct metaphysics” as saying something basic about reality or the universe. (Or, if you’re an instrumentalist and you say “X is the correct metaphysics”, I’d assume you were saying “it’s useful to have a model that treats X as a basic fact about reality or the universe”, which also doesn’t make sense to me if X is “instrumentalism”.)
Well, sufficiently specific realist desiderata. Adding hidden variables to QM doesn’t make the theory any more realist, the way we’re using “realist” here.
Like I said before, it means that instrumentalism is the point of view that is the most useful for designing AI or answering questions about AI. According to the “Yudkowskian computationalist” metaphilosophical view, this also makes it the most useful for rationality in general.
I imagined “realist” to mean something like “the universe can be described in a way independent of the choice of observer, and the perceptions of any given observer can be decoded from the history of the universe in this description, s.t. different observers have compatible observations”. Adding hidden variables does make QM more realist in this sense, for example the de Broglie-Bohm interpretation is realist (but it only makes sense if you assume all observer perceptions can be reduced to configuration variables, which seems false and disqualifies it). MWI fails to be entirely realist because you have to either make the decoding of observer perceptions stochastic (and thereby admit that your description of the universe is incomplete) or to postulate, for each “normal” observer Alice, a huge ensemble of different observers that correspond to versions of Alice in different Everett branches (and thereby lose the physical meaning of quantum probabilities and with it essentially all the predictive power of the theory).
Except that if you are the kind of rationalist who cares about what is really real, you should reject instrumentalism immediately.