Recently, Yudkowsky has been recently talking about third countries stealing
Should probably be
Recently, Yudkowsky has been talking about third countries stealing
Recently, Yudkowsky has been recently talking about third countries stealing
Should probably be
Recently, Yudkowsky has been talking about third countries stealing
A few typos.
How’s intuitively going to be better --> Who’s intuitively going to be better
igure out” that they need to to help and how. --> igure out” that they need to help and how.
very single job as though I’d be a life-and-death decision whether or not to apply --> very single job as though it’d be a life-and-death decision whether or not to apply
I found this post interesting but I think there is something wrong with it, even though I estimate that its central point has value. My remarks focus on “this post as an advice” rather than the “phenomenon explanation” part of the post.
Perhaps the first thing I should say is that I agree some people have a tendency to think too much in certain situations. To delay the first try for too long and to waste a lot of time optimizing entire lines of though that will be revealed to be worthless after five minutes of concrete work. I am one of those people and I have spent time thinking (eh) about when and how this must be improved and corrected.
But I do not think a general cheer for the policy of “acting before thinking” is a good thing. I can cite two examples from my personal life, in the last 10 days, where a dozen hours or more were wasted because someone did not spend half an hour thinking when it was “obviously” sensible to do so. The first one involves writing code without thinking, creating bad code that had to be managed afterward (technical debt), the second involves personal relationships and is private (sorry).
I estimate that this post presents danger on the same scale as the gain it might deliver. In fact I expect this post to be a net negative in terms of direct advice. Ideally, the wisdom found in this post would have its place in a more complete framework on decision processes. As it stands, I am tempted to compare this advice to “to walk, move your right leg forward”, dangerously misleading and incomplete without the complementary advice “also move your left leg”.
Perhaps a better immediate advice would be.
When applicable and when your actions will not have non-trivial lasting consequences think just a little then act, make an attempt. Only after your have tried enough can you think again, if you believe it to be useful.
If we apply this advice to different timescales we get both your second and third example. We however do not get the first, insofar as there can be lasting consequences to a botched application.
at first she had qualms literally called “Effective Evil”
I think this sentence is missing “working for an organization”.
I don’t think this is a good illustration of point 6. The video shows a string of manipulative leading questions, falling short of the ” in a way that they would endorse” criteria.
When people understand that a string of questions is designed to strong arm them into a given position they rarely endorse it. It seems to me that point 6 is more about benevolent and honest uses of leading questions.
Admittedly, I am making the assumption that ” in a way that they would endorse” means “such that if people understood the intent that went into writing the string of questions in that way they would approve of the process”.
Thank you for the post. A few typos:
To understand the current conflicts, it’s vital what the Russian discourse means when it talks about Nazis --> Probably “it’s vital to understand”.
One argument made, about why Russia’s claims of far-right influence in Ukraine are overblown, is that far-right parties don’t have much influence is that they have relatively poor electoral results --> Two different versions of the same end of sentence.
Holocaust to be national heroes who are shall not be criticized feels deeply wrong. --> just “who shall not be”.
[MENTOR]
I am a 25-year-old computer science PhD student and I currently work on natural language processing with neurosymbolic approaches (using a mix of logic and neural networks to do reasoning on a normal text input). I have a (very) high degree of education in math and theoretical computer science. I also have some very limited skills and knowledge for more practical aspects of computer science. Also, I have quite a lot of reflection on the topic of the nature of thought, valid inference, and rational behavior. These are condensed in a currently unpublished book length report. The result of a year-long break I took for personal reflection. I expect to be able to help quite a lot with fundamental reflection that relates to epistemology or rationality.
If you intend to learn about some aspect of mathematics or theoretical computer science I can probably point you toward resources or help you understand technical aspects. I am also willing to serve as a background help for someone going through a math education. Alternatively, I might be willing to help an in depth reflection to which you think I can be relevant.
I can offer some asynchronous email/messages exchange and semiregular conversation, perhaps averaging one or two per month. Ideally, we would schedule conversations quickly when you need them. If you have a very different mode of interaction in mind I am willing to adapt.
[APPRENTICE]
Presentation copy/pasted from the mentor section.
I am a 25-year-old computer science PhD student and I currently work on natural language processing with neurosymbolic approaches (using a mix of logic and neural networks to do reasoning on a normal text input).
I have a (very) high degree of education in math and theoretical computer science. I also have some very limited skills and knowledge for more practical aspects of computer science.
Also, I have quite a lot of reflection on the topic of the nature of thought, valid inference, and rational behavior. These are condensed in a currently unpublished book length report. The result of a year-long break I took for personal reflection.
I am looking for one of the followings:
Resources and help on the topic of the nature of thought and characterization of valid inference. This is a long standing topic for me. A string of reflections and readings that has been and continues to be an important side project.
Like 1 but with a focus on AI or in general automation of thought.
Help better managing my work and motivation. I need to reduce my work related stress. The help of someone who does a high amount of intellectual work without stress could be precious.
By default, I would imagine a few discussions near the beginning and then much less frequent conversations as needed (once a month?). But that’s just off the top of my head.
[NORMAL]
Whether as a mentor, as an apprentice, or because you think we could work together in some other capacity, feel free to DM me through lesswrong. Give me a few days to get back to you. If I do not answer it probably means your message did not go through. In that case just comment below.
I get the idea but I am not sure how to move to a richer domain. The only obvious idea I see is to go to continous time, but that not the usual paradigm for games.
We could go the opposite direction and try to get a result for a more restrictive class of games. I listed some in the post; the only case I thought of for which I do not know if the result holds is bounded games.
Alternatively, it is also possible to take another hypothesis than the strategy not being dominated. The result has shape “if a strategy is then it is a utility maximisation”. Maybe we can find some better .
Is there some other way to change the conjecture that I missed?
I see. If we were to make this formal it would depend on the notion of “complexity” we use.
Notably it seems intuitive that there be counterexample games that pump complexity out of thin air by adding rules and restriction that do not really matter. So “just” adding a simple complexity threshold would certainly not work, for most notions of complexity.
Maybe it is true that “the higher the complexity the larger the portion of nondominated strategies that are utility maximisation”. But
The set of strategies is often infinite, so the notion of “portion” depends on a measure function.
That kind of result is already much weaker than the “coherence result” I have come to expect by reading various sources.
Interesting idea anyway, seems to require quite a bit more work.
I have the suspicion that you read “more complexity” as meaning “more restrictions”, while I meant the contrary (I do realize I didn’t express myself clearly). Is that the case?
My intuition for the idea of complexity is something like “the minimal number of character it takes to implement this game in python”. The flaw is that this assume computable games, which is not in line with the general formulation of the conjecture I used. So that definition does not worK. But that’s roughly what I think of when I read “complexity”. Is that compatible with your meaning?
Note that this is for the notion of complexity of a given game. If you mean the complexity of a class of games then I am less certain how to define it. However if we are to change the category of games we are talking about then the only clear ways to do so I see involve weakening the conjecture by restricting it to speak of strictly fewer games.
In EJT’s post, I remember that the main concrete point was that being stubborn, in this sense:
if I previously turned down some option X, I will not choose any option that I strictly disprefer to X
To my understanding that was a good counter to the idea that anything that is not a utility maximisation is vulnerable to money pumps in a specific kind of games. But that is restricted to “decision tree” games in which in every turn but the first you have an “active outcome” which you know you can keep until the end if you wish. Every turn you can decide to change that active outcome or to keep it. These games are interesting to discuss dutch book vulnerability but they are still quite specific. Most games are not like that.
On a related note:
a non-dominated strategy for a preference tree compact enough compared to the world it applies to will be approximately a utility maximizer
I think I didn’t understand what you mean by “preference tree” here. Is it just a partial order relation (preference) on outcomes? If you mean “for a case in which the complexity of the preference ordering is small compared to that of the rest of the game” , then I disagree. The counterexample could certainly scale to high complexity of the rules without any change to the (very simple) preference ordering.
The closest I could come to your statement in my vocabulary above is:
For some value , if the ratio “complexity of the outcome preference” / “complexity of the total game” is inferior to then any nondominated strategy is (approximately) a utility maximisation.
Is this faithful enough?
Weird. I didn’t expect this to be wrong and I did not expect the other one to be right. Glad I asked.
“Minimal number of character it takes to implement this game in python” would be small because the “game code” part is the laws and the reward.
Not so sure about that. The game has to describe and model “everything” about the situation. So if you want to describe interaction with details of a “real world” then you also need a complete simulation of said real world. While everything is “contained” in the reward function, it is not like the reward function can be computed independently of “what happens” in the game. It is however true that you only need to compute the most minimal version of the game relevant to the outcome. So if your game contains a lot of “pointless” rules that do nothing then they can be safely ignored when computing the complexity of the game. I think that’s normal.
In the case of the real world, even restricting it to a bounded precision, the program would need to be very long. It is not just a description of the sentence “you win if there are a lot of diamonds in the world” (or whatever the goal is). It is also a complete “simulation” of the world.
Btw, the notion I was alluding to is Kolmogorov complexity.
and it would be difficult to write a non-dominated strategy which tries to be non-dominated just by not accruing that much energy overall and instead moving a lot of energy at some time step. Yet it’s probably possible [...]
Depending on the exact parameter, an intuitive strategy that is not dominated but not optimal long terms either could be “never invest anything”, in which you value the present so much that you never move because that would cost energy. Remark that this strategy is still “a” utility maximisation (just value each step more than the next to a high amount). But it is very bad for the utility you described.
But this kind of trick becomes more difficult if I restrict the number of branches you can make in the preferences tree; it’s possible to be non-dominated “just” because you have many non-comparable branches and it suffices to do OK on just one branch. As I restrict the number of branches, you’ll have to do better overall.
I still get the feeling that your notion of preference tree is not equivalent to my own concept of a partial order on the set of outcomes. Could you clarify?
Sorry for only answering now, I was quite busy in the last few days.
I think simulating the real world requires a lot of memory and computations, not a large program. (Not that I know the program.) Komogorov complexity does not put restrictions on the computations. Think about Conway’s game of life.
You also need a specification of the initial state, which dramatically increases the size of the program! Because the formalism only requires turing machines (or any equivalent computation formalism), there is no distinction between the representation of the “rules” and the rest of the necessary data. So even if the rules of physics themselves are very simple (like in the game of life), the program that simulates the world is very big. It probably requires something like “position of every atom at step ”.
Sorry, my choice of expression is confusing. I was thinking about a directed acyclic graph representing the order in my mind, and called that “tree”, but indeed the standard definition of tree is acyclic without orientation, so the skeleton of a DAG does not qualify in general. A minimal representation total order would be a chain, while a partial order has to contain “parallel branches”.
Ok thank you. I will keep reading “order relation” for those.
You seem to interpret that in my example the energy “in the battery of the agent” counts, such that not moving can’t be dominated. I said “energy accumulated in some region of the universe” to avoid this kind of thing. Anyway, the point of the example is not showing a completely general property, but to point at things which have the property, so I expect you to fix yourself counter-specifications that make the example fail, unless of course you thought the example was very broken.
Sure. I agree counterexamples that rely on a small specification flaw are not relevant to your point.
I don’t know if that class of examples works. My intuition is somewhat that there will be nondominated strategies that are not utility maximization “by default” on that sort of games. At least if we only look at utilities that are weighted sums of the energy at various points in time.
On the whole and in general, it is still not intuitive to me whether utility maximization become ubiquitous when the “complexity” ratio you defines goes down.
I agree it might be too ambitious to look at all nondominated strategy. I went for “nondominated” as a condition because it was, in my eyes, the best formal translation of the initial intuitive claim I was trying to test. Besides, that’s what is used in the complete class theorem.
There might be interesting variations of the conjecture with stricter requirements on the strategy. But I also think it would be very hard to give a non-tautological result that uses this notion of “no matter the odds”. The very notion that there are odds to discuss is what we are trying to prove.
Let’s put aside the distiction between initial conditions and rules, I think it is just a distraction at this point.
In general I would even expect a complete simulation of the univers to be non-computable. Ie I expect that the univers contains an infinite amount of information. If we bound the problem to some finite part of time and space then I expect, just as an intuition, that a complete simulation would require a lot of information. Ie, the minimal turing machine / python code that consistently outputs the same result as the simulation for each input is very long.
I do not have a good number to give that translates this intuition of “very long”. Let’s say that simulating the earth during the last 10 days would take multiple millions of terrabits of data? Of course the problem is underspecified. We also need to specify what the legal inputs are.
Anyway, do you agree with this intuition?
Thank you for this post.
Seems kinda weird, right? Well, consider this: Turing showed that there is no computation that this machine can’t perform, that another machine can perform.
I am not sure to which extent you already know what follows, but I thought this might be worth clarifying. The “basic” church turing thesis is that our intuitive notion of “formal computation process” is entirely captured by the turing machine. A consequence is that any “good” formal notion of algorithm / fully described computation process is equivalent or weaker to the Turing machine. So far, this has been true of all proposals and it is taken for granted that this statement is true. Note that the thesis is about being a good formal definition of an intuitive notion. Hence we cannot prove it to be true. We can only prove that various attempts at formalizing computation indeed ended up equivalent to each others.
The “physic church turing thesis” is that no machine can be built that performs computations impossible for a turing machine. For example: there is no way to build a halting oracle in real life. This is less certain.
But if you take a God’s eye view and had the power to get the community to all shift at once to a different model, it sounds to me like there are probably better ones than Turing’s.
People implicitly refer to models much closer to actual programing languages when they prove stuff related to computability. The tedious details of turing machines are rarely discuss in actual proof, at least in my experience. In fact, the precise model at hand is often somewhat ambiguous, which can be slightly problematic. I think the turing machine is a good starting point to understand the notion of computability, just because it is simpler than alternatives (that I am aware of).
One of my favorites: he proved that, given some extremely, extremely reasonable assumptions[13], you should Shut Up and Multiply. Well, he didn’t phrase it that way. He said “maximize expected utility”.
Not sure he did, at least to the extent that it is taken when we say “shut up and multiply”. I guess you refer to the Von Neumann–Morgenstern utility theorem, but that theorem does not provide a full “it is in our best interest to do utility maximization”. There was some discussion on the topic recently: a first claim and my own answer.
I didn’t check the article yet but if I understood your comment correctly then a simpler example would have been “turing machines with a halting oracle”, which is indeed stronger than normal turing machines. (Per my understanding) the church-turing thesis is about having a good formal definition of “the” intuitive notion of algorithm. And an important property of this intuitive notion is that a “perfectly rigorous idiot” could run any algorithm with just paper and a pen. So I would say it is wrong to take something that goes beyond that as a counterexample.
Maybe we should clarify this concept of “the intuitive notion of algorithm”.
PS: We are running dangerously close to just arguing semantics, but insofar as “the church-turing thesis” is a generally consensual notion I do not think the debate is entirely pointless.
Ok, I think I can clarify what people generally mean when they consider that the logic Church-Turing thesis is correct.
There is an intuitive notion of computation that is somewhat consensual. It doesn’t account for limits on time (beyond the fact that everything must be finitely long) and does not account for limits in space / memory. It is also somewhat equivalent to “what a rigorous idiot could be made to do, given immortality and infinite paper/pen”. Many people / most computer scientist share this intuitive idea and at some point people thought they should make more rigorous what it is exactly.
Whenever people tried to come up with formal processes that only allow “obviously acceptable” operations with regard to this intuitive notion, they produced frameworks that are either weaker or equivalent to the turing machine.
The Church-Turing thesis is that the turing machine formalism fully captures this intuitive notion of computation. It seems to be true.
With time, the word “computable” itself has come to be defined on the basis of this idea. So when we read or hear the word in a theoretical computer-science context, it now refers to the turing machine.
Beyond this, it is indeed the case that the word “computation” has also come to be applied to other formalisms that look like the initial notion to some degree. In general with an adjective in front to distinguish them from just “computation”. We can think for example of “quantum computing”, which does not match the initial intuition for “computation” (though it is not necessarily stronger). These other applications of the word “computation” are not what the Church-Turing thesis is about, so they are not counterexamples. Also, all of this is for the “logic” Church-Turing thesis, to which what can be built in real life is irrelevant.
PS: I take it for granted that computer scientists, including those knowledgeable on the notions at hand, usually consider the thesis correct. That’s my experience, but maybe you disagree.
I know this comment is a few months old, but an answer might still be helpful.
I think the following quote from the article answers your question.