I had always assumed that these things were created as sort of abstract ideals, things you could program an AI to use (I find it no coincidence that all three of these concepts come from AI researchers/theorists to some degree) or something you could compare humans to, but not something that humans can actually use in real life.
But having read the original superrationality essays, I realize that Hofstadter makes no mention of using this in an AI framework and instead thinks about humans using it. And in HPMoR, Eliezer has two eleven-year old humans using a bare-bones version of TDT to cooperate (I forget the chapter this occurs in), and in the TDT paper, Eliezer still makes no mention of AIs but instead talks about “causal decision theorists” and “evidential decision theorists” as though they were just people walking around with opinions about decision theory, not the platonic formalized abstraction of decision theories. (I don’t think he uses the phrase “timeless decision theorists”.)
I think part of the rejection people have to these decision theories might be from how impossible they are to actually implement in humans. To get superrationality to work in humans, you’d probably have to broadcast it directly into the minds of everyone on the planet, and even then it’s uncertain how many defectors would remain. You almost certainly could not possibly get TDT or UDT to work in humans because the majority of them cannot even understand them. I certainly had trouble, and I am not exactly one of the dumbest members of the species, and frankly I’m not even sure I understand them now.
The original question remains. It is not rhetorical. Do people think TDT/UDT/superrationality are supposed to be useable by humans?
(I am aware of this; it is no surprise that a very smart and motivated person can use TDT to cooperate with himself, but I doubt they can really be used in practice to get people to cooperate with other people, especially those not of the same tribe.)
Some ways humans act resemble TDT much more than they resemble CDT: some behaviours such as voting in an election with a negligible probability of being decided by one vote, or refusing small offers in the Ultimatum game, make no sense unless you take in account the fact that similar people thinking about similar issues in similar ways will reach similar conclusions. Also, the one-sentence summary of TDT strongly reminds me of both the Golden Rule and the categorical imperative. (I’ve heard that Good and Real by Gary Drescher discusses this kind of stuff in detail, though I haven’t read the book itself.)
(Of course, TDT itself, as described now, can’t be applied to anything because of problems with counterfactuals over logically impossible worlds such as the five-and-ten problem; but it’s the general idea behind it that I’m talking about.)
It’s perhaps more useful to see these as (frameworks for) normative theories, describing which decisions are better than their alternatives in certain situations, analogously to how laws of physics say which events are going to happen given certain conditions. It’s impossible in practice to calculate the actions of a person based on physical laws, even though said actions follow from physical laws, because we lack both the data and the computational capabilities necessary to perform the computation. Similarly, it’s impossible in practice to find recommendations for actions of a person based on fundamental decision theory, because we lack both the problem statement (detailed descriptions of the person, the environment, and the goals) and computational capabilities (even if these theories were sufficiently developed to be usable). In both cases, the problem is not that these theories are “impossible to implement in humans”; and certain approximations of their conclusions can be found.
Some people think so; they are wrong. (Examples: 1, 2, 3, 4, 5, 6, 7. Most of these take overly broad vague definitions of a person’s “platonic algorithm”; #5 is forgetting that natural selection acts on the level of genes, not people.)
Eliezer: “This is primarily a theory for AIs dealing with other AIs.” Unfortunately, it’s difficult to write papers or fiction publicizing TDT that solely address AI’s- especially when the description of TDT needs to be in a piece of Harry Potter fanfiction.
On a slightly more interesting side note, if TDT were applicable in real life, people would likely be computation hazards, since a simulation of another person accurate enough to count as implementing the same, simulated platonic algorithm as the one they actually use would also be quite possibly be complex enough to be a person.
TDT is not even that good at cooperating with yourself, if you’re not in the right mindset. The notion that “If you fail at this you will fail at this forever” is very dangerous to depressed people, and TDT doesn’t say anything useful (or at least nothing useful has been said to me on the topic) about entities that change over time, ie Humans. I can’t timelessly decide to benchpress 200 pounds whenever I go to the gym, if I’m physically incapable of it.
A failure or so, in itself, would not matter, if it did not incur a loss of self-esteem and of self-confidence. But just as nothing succeeds like success, so nothing fails like failure. Most people who are ruined are ruined by attempting too much. Therefore, in setting out on the immense enterprise of living fully and comfortably within the narrow limits of twenty-four hours a day, let us avoid at any cost the risk of an early failure. I will not agree that, in this business at any rate, a glorious failure is better than a petty success. I am all for the petty success. A glorious failure leads to nothing; a petty success may lead to a success that is not petty.
Do people think superrationality, TDT, and UDT are supposed to be useable by humans?
I had always assumed that these things were created as sort of abstract ideals, things you could program an AI to use (I find it no coincidence that all three of these concepts come from AI researchers/theorists to some degree) or something you could compare humans to, but not something that humans can actually use in real life.
But having read the original superrationality essays, I realize that Hofstadter makes no mention of using this in an AI framework and instead thinks about humans using it. And in HPMoR, Eliezer has two eleven-year old humans using a bare-bones version of TDT to cooperate (I forget the chapter this occurs in), and in the TDT paper, Eliezer still makes no mention of AIs but instead talks about “causal decision theorists” and “evidential decision theorists” as though they were just people walking around with opinions about decision theory, not the platonic formalized abstraction of decision theories. (I don’t think he uses the phrase “timeless decision theorists”.)
I think part of the rejection people have to these decision theories might be from how impossible they are to actually implement in humans. To get superrationality to work in humans, you’d probably have to broadcast it directly into the minds of everyone on the planet, and even then it’s uncertain how many defectors would remain. You almost certainly could not possibly get TDT or UDT to work in humans because the majority of them cannot even understand them. I certainly had trouble, and I am not exactly one of the dumbest members of the species, and frankly I’m not even sure I understand them now.
The original question remains. It is not rhetorical. Do people think TDT/UDT/superrationality are supposed to be useable by humans?
(I am aware of this; it is no surprise that a very smart and motivated person can use TDT to cooperate with himself, but I doubt they can really be used in practice to get people to cooperate with other people, especially those not of the same tribe.)
Some ways humans act resemble TDT much more than they resemble CDT: some behaviours such as voting in an election with a negligible probability of being decided by one vote, or refusing small offers in the Ultimatum game, make no sense unless you take in account the fact that similar people thinking about similar issues in similar ways will reach similar conclusions. Also, the one-sentence summary of TDT strongly reminds me of both the Golden Rule and the categorical imperative. (I’ve heard that Good and Real by Gary Drescher discusses this kind of stuff in detail, though I haven’t read the book itself.)
(Of course, TDT itself, as described now, can’t be applied to anything because of problems with counterfactuals over logically impossible worlds such as the five-and-ten problem; but it’s the general idea behind it that I’m talking about.)
I have. It does. Strongly recommended.
adds Good and Real at the end of the queue of books I’m going to read
It’s perhaps more useful to see these as (frameworks for) normative theories, describing which decisions are better than their alternatives in certain situations, analogously to how laws of physics say which events are going to happen given certain conditions. It’s impossible in practice to calculate the actions of a person based on physical laws, even though said actions follow from physical laws, because we lack both the data and the computational capabilities necessary to perform the computation. Similarly, it’s impossible in practice to find recommendations for actions of a person based on fundamental decision theory, because we lack both the problem statement (detailed descriptions of the person, the environment, and the goals) and computational capabilities (even if these theories were sufficiently developed to be usable). In both cases, the problem is not that these theories are “impossible to implement in humans”; and certain approximations of their conclusions can be found.
Some people think so; they are wrong. (Examples: 1, 2, 3, 4, 5, 6, 7. Most of these take overly broad vague definitions of a person’s “platonic algorithm”; #5 is forgetting that natural selection acts on the level of genes, not people.)
Eliezer: “This is primarily a theory for AIs dealing with other AIs.” Unfortunately, it’s difficult to write papers or fiction publicizing TDT that solely address AI’s- especially when the description of TDT needs to be in a piece of Harry Potter fanfiction.
On a slightly more interesting side note, if TDT were applicable in real life, people would likely be computation hazards, since a simulation of another person accurate enough to count as implementing the same, simulated platonic algorithm as the one they actually use would also be quite possibly be complex enough to be a person.
Why do you think we would need to get everyone to use UDT for it to be useful to you? It’s not like UDT can’t deal with non-UDT agents.
TDT is not even that good at cooperating with yourself, if you’re not in the right mindset. The notion that “If you fail at this you will fail at this forever” is very dangerous to depressed people, and TDT doesn’t say anything useful (or at least nothing useful has been said to me on the topic) about entities that change over time, ie Humans. I can’t timelessly decide to benchpress 200 pounds whenever I go to the gym, if I’m physically incapable of it.
-Arnold Bennett, How to Live on 24 Hours a Day
A dangerous truth is still true. Let’s not recommend people try at things if a failure will cause a failure cascade!
The notion of “change over time” is deeply irrelevant to TDT, hence its name.