Idan Arye

Karma: 224

Idan Arye Sep 26, 2020, 7:26 PM
1 point
in reply to: quanticle’s comment on: Honoring Petrov Day on LessWrong, in 2019
Petrov’s choice was not about dismissing warnings, it’s about picking on which side to err. Wrongfully alerting his superiors could cause a nuclear war, and wrongfully not alerting them would disadvantage his country in the nuclear war that just started. I’m not saying he did all the numbers, used Bayes’s law to figure the probability there is an actual nuclear attack going on, assigned utilities to all four cases and performed the final decision theory calculations—but his reasoning did take into account the possibility of error both ways. Though… it does seem like his intuition gave utility much more weight than to probabilities.
So, if we take that rule for deciding what to do with a AGI, it won’t be just “ignore everything the instruments are saying” but “weight the dangers of UFAI against the missed opportunities from not releasing it”.
Which means the UFAI only needs to convince such a gatekeeper that releasing it is the only way to prevent a catastrophe, without having to convince the gatekeeper that the probabilities of the catastrophe are high or that the probabilities of the AI being unfriently are low.

Idan Arye Sep 26, 2020, 7:12 PM
1 point
in reply to: TAG’s comment on: A Priori
That isn’t what you need to show. You need to show that the semantics have no ontological implications, that they say nothing about the territory .
Actually, what I need to show is that the semantics say nothing extra about the territory that is meaningful. My argument is that the predictions are canonical representation of the belief, so it’s fine if the semantics say things about the territory that the predictions can’t say, as long as everything it says that does not affect the predictions is meaningless. At least, meaningless in the territory.
The semantics of gravity theory says that the force that pulls objects together over long range based on their mass is called “gravity”. If you call that force “travigy” instead, it will cause no difference in the predictions. This is because the name of the force if a property of the map, not the territory—if it was meaningful in the territory it should have had impact on the predictions.
And I claim that the “center of the universe” is similar—it has no meaning in the territory. The universe has no “center”—you can think of “center of mass” or “center of bounding volume” of a group of objects, but there is no single point you can naturally call “the center”. There can be good or bad choices for the center, but not right or wrong choices—the center is a property of the map, not the territory.
If it had any effect at all on the territory, it should have somehow affected the predictions.

Idan Arye Sep 25, 2020, 2:49 PM
1 point
in reply to: TAG’s comment on: A Priori
If you take a heliocentric theory, and substitute “geocentric” for “heliocentric”, you get a theory that doens’t work in the sense of making correct predictions. You know this, because in previous comments you have already recognised the need for almost everything else to be changed in a geocentric theory in order to make it empirically equivalent to a heliocentric theory.
I only change the title, I don’t change anything they theory says. So its predictions are still the same as the heliocentric model.
But you are arguing against realism, in that you are arguing that theories have no content beyond their empirical content, ie their predictive power. You are denying that they are have any semantic (non empirical content), and, as an implication of that, that they “mean” or “say” nothing about the territory. So why would you care that one theory in more complex than another, so long as its predictions are accurate?
The semantics are still very important as a compact representation of predictions. The predictions are infinite—the belief will have to give a prediction for every possible scenario, and scenariospace is infinite. Even if the belief is only relevant for a finite subset of scenarios, it’d still have to say “I don’t care about this scenario” an infinite number of times.
Actually, it would make more sense to talk about belief systems than individual beliefs, where the belief system is simply the probability function P. But we can still talk about single beliefs if we remember that they need to be connected to a belief system in order to give predictions, and that when we compare two competing beliefs we are actually comparing two belief systems where the only difference is that one has belief A and the other has belief B.
Human minds, being finite, cannot contain infinite representations—we need finite representations for our beliefs. And that’s where the semantics come in—they are compact rules that can be used to generate predictions for every given scenario. And they are also important because the amount of predictions we can test is also finite. So even if we could comprehend the infinite prediction field over scenariospace, we wouldn’t be able to confirm a belief based on a finite number of experiments.
Also, with that kind of representation, we can’t even come up with the full representation of the belief. Consider a limited scenario space with just three scenarios X, Y and Z. We know what happened in X and Y, and write a belief based on it. But what would that belief say about Z? If the belief is represented as just its predictions, without connections between distinct predictions, how can we fill up the predictions table?
The semantics help us with that because they have less degrees of freedom. With N degrees of freedom we can match any $K \leq N$ number of observations, so we need $M > N$ observations to even start counting them as evidence. I not sure how to come up with a formula for the number of degrees of freedom a semantic representation of a belief has—this depends not only on the numerical constants but also on the semantics—but some properties of it are obvious:
1. The prediction table representation has infinite degrees of freedom, since it can give a prediction for each scenario independently from the predictions given to the other scenarios.
2. A semantic representation that’s strictly more simple than another semantic representation—that is, you can go from the simple one to the complex one just by adding rules—then the simpler one has less degrees of freedom than the complicated one. This is because the complicated one has all the degrees of freedom the simpler one had, plus more degrees of freedom from the new rules (just adding the rules is some degrees of freedom, even if the rule itself does not contain anything that can be tweaked)
So the simplicity of the semantic representation is meaningful because it means less degrees of freedom and thus requires less evidence, but it does not make the belief “truer”—only the infinite prediction table determines how true the belief is.

Idan Arye Sep 24, 2020, 11:56 PM
3 points
on: Decoherence is Falsifiable and Testable
We require new predictions not because the theory is newer than some other theory it could share predictions, but because the predictions must come before the experimental results. If we allow theories theories to rely on the results of already known experiments, we run into two problems:
1. Overfitting. If the theory only needs to match existing results it can be constructed a way that matches all these results—instead of trying to match the underlying rules that generated these results.
2. We may argue ourselves into believing our theory made predictions that match our results, regardless of whether the theory naturally makes these predictions.
Now, if the new theory is a strictly simpler versions of an old one—as in “we don’t even need X” simpler—then these two problems are nonissue:
1. If the more complicated theory did not overfit, the simpler version is as good as guaranteed to not overfit either.
2. We don’t need to guess what our new theory would predict if we didn’t know the results—the old theory already did those predictions before knowing the results, and it should be straightforward to show it wasn’t using the part we removed.
So… I will allow it.

Idan Arye Sep 23, 2020, 12:44 PM
1 point
in reply to: Idan Arye’s comment on: A Priori
OK, I continued reading, and in Decoherence is Simple Eliezer makes a good case for Occam’s Razor as more than just a useful tool.
In my own words (:= how I understand it) more complicated explanations have a higher burden of proof and therefore require more bits of evidence. If they give the same predictions as the simpler explanations, then each bit of evidence counts for both the complicated and the simple beliefs—but the simpler beliefs had higher prior probabilities, so after adding the evidence their posterior probabilities should keep being higher.
So, if a simple belief A started with −10 decibels and a complicated belief B started with −20 decibels, and we get 15 decibels of evidence supporting both, the posterior credibility of the beliefs are 5 and −5 - so we should favor A. Even if we get another 10 decibels of evidence and the credibility of B becomes 5, the credibility of A is now 15 so we should still favor it. The only way we can favor B is if we get enough evidence that support B but not A.
Of course—this doesn’t mean that A is true and B is false, only that we assign a higher probability to A.
So, if we go back to astronomy—our neogeocentric model has a higher burden of proof than the modern model, because it contains additional mysterious forces. We prove gravity and relativity and the work out how centrifugal forces work and that’s (more or less) enough for the modern model, and the exact same evidences also support the neogeocentric model—but they are not enough for it because we also need evidence for the new forces we came up with.
Do note, though, that the claim that “there is no mysterious force” is simpler than “there is a mysterious force” is taken for granted here...

Idan Arye Sep 20, 2020, 4:55 PM
1 point
in reply to: TAG’s comment on: A Priori
They assert different things because they mean different things, because the dictionary meanings are different.
The Quotation is not the Referent. Just because the text describing them is different doesn’t mean the assertions themselves are different.
Eliezer identified evolution with the blind idiot god Azathoth. Does this make evolution a religious Lovecraftian concept?
Scott Alexander identified the Canaanite god Moloch with the principle that forces you to sacrifice your values for the competition. Does this make that principle an actual god? Should we pray to it?
I’d argue not. Even though Eliezer and Scott brought the gods in for the theatrical and rhetorical impact, evolution is the same old evolution and competition is the same old competition. Describing the idea differently does not automatically make it a different idea—just like describing $f (x) = (x + 1)^{2}$ as $g (x) = x^{2} + 2 x + 1$ does not make it a different function.
In case of mathematic functions we have a simple equivalence law: $f \equiv g ⟺ \forall_{x} f (x) = g (x)$ . I’d argue we can have a similar equivalence law for beliefs - $A \equiv B ⟺ \forall_{X} P (X ∣ A) = P (X ∣ B)$ where A and B are beliefs and X is an observation.
This condition is obviously necessary because if $A \equiv B$ even though $\exists_{Y} P (Y ∣ A) \neq P (Y ∣ B)$ and we find that $P (Y) = P (Y ∣ A)$ , that would support A and therefore also B (because they are equivalent) which means an observation that does not match the belief’s predictions supports it.
Is it sufficient? My argument for its sufficiency is not as analytical as the one for its necessity, so this may be the weak point of my claim, but here it goes: If $A \equiv̸ B$ , even though they give the same predictions, then something other than the state and laws of the universe is deciding whether a belief is true or false (actually—how much accurate is it). This undermines the core idea of both science and Bayesianism that beliefs should be judged by empirical evidences. Now, maybe this concept is wrong—but if it is, Occam’s Razor itself becomes meaningless because if the explanation does not need to match the evidences, then the simplest explanation can always be “Magic!”.

Idan Arye Sep 20, 2020, 3:38 PM
1 point
in reply to: TAG’s comment on: A Priori
In the thought experiment we are considering , the contents of the box can be er be tested. Nonetheless $10 and $100 mean different things.
I’m not sure you realize how strong a statement “the contents of the box can be never be tested” is. It means even if we crack open the box we won’t be able to read the writing on the bill. It means that even if we somehow tracked all the $20 and all the $100 bills that were ever printed, their current location, and whether or not they were destroyed, we won’t be able to find one which is missing and deduce that it is inside the box. It means that even if we had a powerful atom-level scanner that can accurately map all the atoms in a given volume and put the box inside it, it won’t be able to detect if the atoms are arranged like a $20 bill or like a $100 bill. It means that even if a superinteligent AI capable of time reversal calculations tried to simulate a time reversal it wouldn’t be able to determine the bill’s value.
It means, that the amount printed on that bill has no effect on the universe, and was never affected by the universe.
Can you think of a scenario where that happens, but the value of dollar bill is still meaningful? Because I can easily describe a scenario where it isn’t:
Dollar bills were originally “promises” for gold. They were signed by the Treasurer and the secretary of the Treasury because the Treasury is the one responsible for fulfilling that promise. Even after the gold standard was abandoned, the principle that the Treasury is the one casting the value into the dollar bills remains. This is why the bills are still signed by the Treasury’s representatives.
So, the scenario I have in mind is that the bill inside the box is a special bill—instead of a fixed amount, it says the Treasurer will decide if it is worth 20 or 100 dollars. The bill is still signed by the Treasurer and the secretary of the Treasury, and thus has the same authority as regular bills. And, in order to fulfill the condition that the value of the bill is never known—the Treasurer is committed to never decide the worth of that bill.
Is it still meaningful to ask, in this scenario, if the bill is worth $20 or $100?

Idan Arye Sep 20, 2020, 2:59 PM
1 point
in reply to: TAG’s comment on: A Priori
I’m not sure I follow—what do you mean by “didn’t work”? Shouldn’t it work the same as the heliocentric theory, seeing how every detail in its description is identical to the heliocentric model?

Idan Arye Sep 20, 2020, 12:28 PM
1 point
in reply to: TAG’s comment on: A Priori
So if I copied the encyclopedia definition of the heliocentric model, and changed the title to “geocentric” model, it would be a “bad, wrong , neo-geocentric theory [that] is still a geocentric theory”?

Idan Arye Sep 20, 2020, 12:25 PM
1 point
in reply to: TAG’s comment on: A Priori
If A and B assert different things, we can test for these differences. Maybe not with current technology, but in principle. They yield different predictions and are therefore different beliefs.

Idan Arye Sep 19, 2020, 10:40 PM
1 point
in reply to: TAG’s comment on: A Priori
But this is not the dictionary definition of the geocentric model we are talking about—this we have twisted it to have the exact same predictions as the modern astronomical model. So it no longer asserts the same things about the territory as the original geocentric model—its assertions are now identical to the modern model. So why should it still hold the same meaning as the original geocentric model?

Idan Arye Sep 19, 2020, 10:17 PM
1 point
in reply to: TAG’s comment on: A Priori
If a universe where the statement is true is indistinguishable from a universe where the statement is false, then the statement is meaningless. And if the set of universes where statement A is true is identical to the set of universes where statement B is true, then statement A and statement B have the same meaning whether or not you can “algebraically” convert one to the other.

Idan Arye Sep 19, 2020, 12:11 PM
1 point
in reply to: Teerth Aloke’s comment on: A Priori
If the content of the box is unknown forever, that means that it doesn’t matter what’s inside it because we can’t get it out.

Idan Arye Sep 19, 2020, 10:08 AM
1 point
in reply to: TAG’s comment on: A Priori
Of course it is different. Heliocentricism says something different about reality than geocentricism.
Different… how? In what meaningful ways is it different?

Idan Arye Sep 17, 2020, 11:30 PM
1 point
in reply to: Idan Arye’s comment on: A Priori
Actually… if you squint a bit there is a compact way to represent the fitted geocentric mode:
- The Earth is at the center.
- There is a mysterious force, originating from the earth, that pushes all objects away. It’s strength is what you would expect from the Earth’s centrifugal force according to the modern model.
- All the objects in the universe, other than the Earth, are accelerated at the opposite direction and size than what the Earth’s acceleration in the modern model.
With relativity in mind these rules may not be enough, but let’s ignore that for the sake of the argument.
At this point, I’ll ask the neogeocenterists (pun intended), wouldn’t it be simpler and easier to just use the modern model for calculating my predictions?
“But then you’ll get wrong results!”, they’ll say.
How so? The centrifugal force from assuming the Earth rotates mimics your mysterious force that pushes all things away, and the acceleration the Earth mimics the acceleration your model adds to all other celestial bodies. Then, when predictions for the relative position and velocity of each pair of objects should be identical in both models.
“Yea, sure, but you’d still get wrong results—the Earth will not be in the cener.”
So… what? What difference does being in the center make? If it makes a difference, we should test for that difference and support or disprove your model!
“No, this is not a difference you can test for, but it makes us special!”
Special… how?
“There are countless planets in the universe, and infinite positions to put the center. What is the probability that we are the ones in the center? That we are the only planet that doesn’t move? That these mysterious unexplainable forces make sure we are kept in the center of the universe?”
Pretty damn high, I’d say, considering how you picked the origin to be our position, you decided to use our velocity for calculating the relevant velocities of all other objects as if they were absolute velocities, and you are the ones who added these mysterious forces instead of picking a model that does not require them. Sure, I can’t scientifically prove that your model is wrong, but you can’t prove that all the other models that don’t put the Earth in the center are wrong—and therefore you cannot claim that the Earth is special for being in the center of the universe.
--------------------------------------
By defining beliefs as the predictions you can get from them, I don’t need Occam’s Razor to be true—it is enough that it is useful. The neogeocentric model is not different than the model I use—not in any meaningful way, for if it was different in any meaningful way that would be a difference in predictions that we could test. So I don’t need to argue that simpler = truer—I just let them have their complicated representation of the belief, and instead draw the line on trying to get any meaningful insight from it that cannot be obtained from the more compact representation that I use.

Idan Arye Sep 17, 2020, 8:32 PM
1 point
in reply to: TAG’s comment on: A Priori
You’ll need more than just epicycles to make the geocentric model yield accurate predictions. For example, what will happen if we launch a rocket straight up, and observe the Earth from that rocket?
According to the geocentric model, the Earth does not spin—it is the Sun that revolves around it. So if we launch a rocket straight up, it should not observe the Earth rotating. With our modern model, or even with the heliocentric model, we would predict that the rocket see the Earth rotating because the ratio between the perpendicular velocity the rocket started with and its distance from the Earth gets lower and lower as the rocket gets farther away.
So that’s one different prediction.
Now, say that we modify the geocentric model so that the Earth is still in the center, but also rotates. What’s is the angular velocity of its rotation? If we calculate it based on the observations from our rocket, we will come the the conclusion that the Sun’s rotational velocity is extremely low. So low, in fact, that it should not be able to maintain its centrifugal force—and should have been pulled into the Earth a long time ago.
So you’d have to change the rules of gravity too. And then the rules of relativity. And the description will be infinite—because not only you’ll need to match not only the known epicycles, not only the existing scenarios, but any possible setting and formation that can come to mind.
And because it is infinite, it can never be actually used to predict things before they are observed—because calculating these predictions will take infinite time. We can always ever use a finite sub-representation of it, which will not yield accurate predictions for all cases.
But still—if we could, it will not be different than a correct model, just like the sum of an infinite Taylor series is the same as the infinitely differentiable function it is derives from. Even if the Taylor series no longer represents the intuition behind that function.

Idan Arye Sep 17, 2020, 2:13 PM
0 points
on: A Priori
Modus Ponens can be justified by truth tables. A is $\begin{matrix} _{B}^{A} & T & F T & T & F F & T & F \end{matrix}$ , and $A \to B$ is $\begin{matrix} _{B}^{A} & T & F T & T & T F & F & T \end{matrix}$ . By combining these to $A \land A \to B$ we get the truth table $\begin{matrix} _{B}^{A} & T & T T & T \land T = T & F \land T = F F & T \land F = F & F \land T = T \end{matrix}$ which is only true if both A and B are true.
Of course, one can always reject the notion of truth tables and then we are back to square one...
------
As for Occam’s Razor—I used to think of it in terms of avoiding overfitting. More complex explanations have more degrees of freedom which makes it easier for them to explain all the datapoints by “twisting” instead of by uncovering some underlying rule.
Now that I’ve been exposed to Bayesianism though, and consider beliefs to be defined by the predictions you can get from them, I see Occam’s Razor as a matter of pragmatism:
1. If two explanations yield the exact same predictions, then they are different representations of the same belief. We should choose the simpler one not because it is more true or more accurate—they are identically true and identically accurate—but because it is easier to work with.
2. If the two explanations yield some different predictions, then we don’t even need Occam’s Razor—we need to test these predictions and see which explanation is more accurate.
3. If for whatever reason we can’t test these predictions (too expensive? unethical? cannot be done with current technology?) but we still need to pick one, picking the simpler one is still a good rule of thumb because it is bounded—we can always make the explanation more complex, but there is a limit to how much simpler we can make it before it no longer explains the observations. So we have a stopping condition—if the rule was to pick the more complex explanations we would be stuck forever in a race to make the explanation more and more complicated.

Idan Arye Sep 16, 2020, 11:47 PM
1 point
on: What Does “Signalling” Mean?
But the alarm call is a fact about the signaller—it conveys the fact that it is aware of the predator. For a bird to give the alarm call without being aware of a predator, it would have to do it a lot more often, wasting precious time and energy. Dishonest signal is more expensive than honest signal because a bird that sounded the alarm because it noticed a predator can minimize the alarm time and get back to whatever it was doing as soon as the predator is gone.

Idan Arye Sep 10, 2020, 10:48 AM
1 point
on: Wrong Questions
A key requirement of free will is to be unexplainable. If we can explain free will then it’s no longer “free will”—it’s just a process, deterministic or probabilistic, that can be followed step by step.
Even if current science cannot explain it, the idea that it can be explained already disqualifies it from being free will.
So, the state of affairs where we have free will is to have some component in our decision making process that is complex enough and yet fundamentally unexplainable.

Idan Arye Sep 9, 2020, 6:57 PM
1 point
on: Dissolving the Question
I think we care about whether or not we have free will because we associate it with accountability—both our own and others.
If someone picks me up and throws me on you, you should not blame me for getting slammed—this is not my fault, and I had no say in the matter. If someone points a gun at me and tells me to hit you, you probably won’t blame for complying. But if you had to rank my accountability in these two cases, it’s obvious that I’m more accountable in the latter because I did have a choice—I could not hit you and get shot. This is a very unfavorable choice and you would not expect me to pick it, so in the global scale of accountability it doesn’t really count as a choice—but if we zoom in to just these two cases it’s more choice than the no choice I got at the former case.
Moving on: if I steal food because I’m hungry, should I be held accountable?
This question is controversial in ethical philosophy, and I won’t form an opinion because the point of this exercise is not to solve such controversies—it’s to understand the cognition behind them. If I don’t eat for long enough, I will die. But unlike the case with the gun, I will not die immediately if I don’t steal this bread right now—so I don’t face immediate death, only hunger. I have more choice here—not enough choice to make me consensually accountable, but enough to push it from consensually unaccountable to controversial.
So, accountability is directly linked to will—the more freely I can use my will, the more accountable I should be for my actions. We want to know if people have free will because we want to know if (or—to what degree) they should be held accountable.
Why do we care about accountability? Because we want to punish and/or reward, but we don’t want to do that based on luck. If you punish me for stealing, but I stole because I’m hungry, then you punish me for being unlucky enough to go hungry—which is morally wrong. But it it “really” was my will to steal, then you are punishing me for stealing—which is morally right.
We care about free will because we don’t want to punish/reward people because of their circumstances—only because of their essences. To discuss the actual difference between circumstances and essences is to answer the question of free will—which is outside the scope of this exercise.