[SEQ RERUN] When Science Can’t Help
Today’s post, When Science Can’t Help was originally published on 15 May 2008. A summary (taken from the LW wiki):
If you have an idea, Science tells you to test it experimentally. If you spend 10 years testing the idea and the result comes out negative, Science slaps you on the back and says, “Better luck next time.” If you want to spend 10 years testing a hypothesis that will actually turn out to be right, you’ll have to try to do the thing that Science doesn’t trust you to do: think rationally, and figure out the answer before you get clubbed over the head with it.
Discuss the post here (rather than in the comments to the original post).
This post is part of the Rerunning the Sequences series, where we’ll be going through Eliezer Yudkowsky’s old posts in order so that people who are interested can (re-)read and discuss them. The previous post was Science Doesn’t Trust Your Rationality, and you can use the sequence_reruns tag or rss feed to follow the rest of the series.
Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day’s sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series.
OK… Now I’m off reading Rathmann and Hutter on Solomonoff Induction. I don’t think I can continue discussing the current posts without knowing what exactly EY is talking about on more than the superficial level given on the german wikipedia or LukeProgs unfinished intro. If anyone has suggestions where to find the same information in a more condensed format (70 pages isn’t an hour-read through for me, exactly), please tell me.
You’re looking for an intro to Solomonoff induction? This might be helpful. It’s an unfinished tutorial, but what is finished might it might save you a few hours of reading.
That’s a perfectly valid reason to use Bayesian logic. Unfortunately, the MWI is designed to be untestable experimentally (hence the term interpretation), so this reason does not apply to the EY’s favorite interpretation.
This is debatable, but I’m in the camp that many worlds is testable against collapse interpretations, if not other interpretations. Seeing typical quantum observations for a macroscopic body would settle it pretty easily. We just need to be able to accelerate baseballs in baseball accelerators.
Not quite baseballs, but a couple of years ago, quantum superposition was observed in a 30 micron object, which is large enough to be seen with the naked eye.
It’s designed to be just quantum mechanics with no frills added. That you can’t distinguish it from QM is a bizarre accusation to make.
Any kind of ontological model behind the math of QM, be it collapse, many worlds or decision theory, is by definition frills (non-essential features).
A) being unable to distinguish it from QM is still a bizarre accusation to make, and
B) MWI is the minimal interpretation of QM. Simply, “This equation describes reality. No other processes occur.”
B isn’t true at all. If you look at something like Ballentine’s ensemble interpretation, its pretty clear thats minimal. Remember the stylized fact- quantum mechanics is a response to:
we can’t successfully predict outcomes to every experiment
we can always predict distributions to classes of experiments.
Ballentine tells us- no worries, wavefunctions are tools used to predict probability distributions. There is no wave function collapse, you’ve just sampled once from the predicted distribution.
Many worlds would prefer the wavefunction to represent single systems (which might be confusing map/territory) and so denies 1 (every result possible does happen.)
Sampling mechanism, please.
If it isn’t “They’re all real, and we just feel one of them”, then it’s adding something big—a whole new dynamical principle.
If it IS “They’re all real, and we just feel one of them”, then hey, it’s just MWI with another name.
The idea of Ballentine’s interpretation is that the wavefunction is a tool to compute a probability distribution. Probability distributions are real. Ballentine would accuse many worlders of confusing the map (wavefunctions) for the territory.
You never predict “this is what happens when I send a photon through the slit”- instead you predict “this is what happens when i send lots of photons through a slit.” Its very much the minimal interpretation of quantum mechanics. I recommend skimming through Ballentine’s text (its a very good for many reasons- deriving the Schroedinger equation from galilean relativity, for instance).
I have no idea what you mean by “bizarre” and “accusation”. I prefer less emotionally charged terminology, that’s why I ignored it the first time you said that. This time I downvoted your comment for straying from a reasonably rational discussion into something looking awfully close to personal attacks.
Only according to its proponents, who tend to be rather biased.
That’s the “shut up and calculate” version, no MWI necessary.
Anyway, my mental model of you is that you think that there is some fixed objective reality behind what we model and measure, not a useful concept when dealing with the quantum world.
The notion of accusation should be clear from context—“designed to be untestable” is tantamount to ‘intellectually dishonest’ in science. It’s not a personal attack for me to describe this as an accusation.
Then what happens to the components of the wavefunction that we do not observe? If you just stick with QM, then nothing special happens to them. They go on being themselves. If this is real—if you take it seriously—then they’re still out there doing their respective things.
This is Many Worlds. It follows directly from just doing “QM, no exceptions”. Calling it MWI is calling it by one particular consequence, but the real heart of it is to take Quantum Mechanics seriously on its own terms. I’d call it Quantum Realism.
Quantum Realism—now quite explicitly called out on its own as ‘some fixed objective reality behind what we model and measure’ is not a useful concept? ? ?
WHAT?
I’ve done QM. I nearly ended up doing a quantum computation post-doc. If I tried to burden my mind with anything more complicated than realism while thinking about what was going on in that system, I would have gone mad.
~~~~
I get the feeling I’m saying, “This is the ground state, we’re all out of energy.” and you’re saying “It has h-bar omega over two energy left!”
I’m not sure its fair to pretend any scientific theory can make statements about things other than what we observe. Keep in mind, what you are saying is that encountering an experiment that seems to be fundamentally probabilistic (I can develop theories that predict probabilities, but NOT theories that predict exactly what will happen) is proof that a. the world is deterministic, b. this determinism is hidden from us in such a way as to make it ‘seem’ like things are probabilistic.
If we are properly Bayesian, shouldn’t we instead update by saying that the existence of experiments where we can only predict probabilities should push us in the direction of the existence of fundamental probability?
That’s cute, and if there weren’t such a thing as interference, you’d be totally right. But there is. It clearly indicates that this isn’t simply probability we’re talking about, here.
I’m not sure ‘cute’ is a useful word in facilitating a rational conversation. I certainly had an emotional reaction to it (feeling like I was being patronized) that I had to fight with in order to think clearly.
Anyway, maybe its best to see where we might be disagreeing. Do you believe its true that: 1. there exist experiments where we cannot predict what will happen
In those experiments we can predict what will happen if we run large ensembles of experiments in the same class
If you agree with that- do you think it MIGHT be problematic to use these facts to update in the direction of a fully deterministic theory?
Looks like we are rehashing our positions without getting through to each other… I guess I’ll stop here. Thank you for trying. If you like, you can often find me on freenode in either lesswrong or physics channels, some immediacy in the discussion might help clear up the hidden differences that seem to prevent us from reaching any common ground.
My understanding of the “shut up and calculate” version (which I ascribe to) is that we have these equations that reflect some kind of process. We don’t know why the equations work, but we know that they do, and we’re rather curious about that underlying process. We’d like to know why they work, but until they can give us some sort of experimental prediction, they are strictly interpretations. The MWI is at the moment, just an interpretation. I suspect it is testable, and I would like to see it tested, once we have the means to do so.
I must confess to harboring the suspicion that there is indeed some objective reality. Why do you believe that this is not a useful concept?
Have you noticed a contradiction in what you said?
The map/territory meme works better in some cases, worse in others. Note that it is only a model, not necessarily “the way things are”. A simpler and just as powerful approach is that repeatably testable models are all there is. It is tempting to assign repeatability of testing to the invisible unachievable “objective reality” behind the veil, something science strives to uncover, but this notion is unnecessary. If this makes you uncomfortable, do the standard exercise and reformulate the question such that it has an answer: “Why do I think that there is objective reality?”. (The answer turns out to be similar to that of “Why do I think I have free will?”)
Again, the minimal approach to the scientific method is: build models and test them. Adopt those with more experimentally confirmed predictive power. And by all means, use Bayes as an aid in constructing less-stupid models.
I had not, actually. Thank you. Perhaps I should refer to it as the Presently Untestable Many Worlds Hypothesis (PUMWH).
In The Simple Truth, EY argued that we use the word “reality” to mean whatever it is that actually determines our observations, and “models” to mean whatever we use to generate our predictions. That seems like a very useful approach to it. And if there isn’t any “objective reality” that determines my observations, why do my observations seem so frequently to correspond to an ordered, lawful universe?
(upvoted)
So, you and shminux both seem to agree that there exists some X such that X determines observations, and that there exists some Y such that we use Y to constrain our predictions.
You label (X,Y) (“reality”,”model”).
Shminux labels (X,Y) (“repeatably testable model”,”model”).
Is there a situation where one framework makes different predictions than the other?
More generally, is there anything y’all are actually disagreeing about, other than labels?
It sounds like you both agree, for example, that different observers can share an X while having different Ys, as can the same observer at different times, and that different Ys can vary in terms of how closely they approximate X, and that it’s best to adopt the most reliably X-approximating Ys I can.
What don’t you agree on?
I admit I find shminux’s labeling a little bizarre, as it leads me to say things like “the behavior of a Neandrathal hunter who falls off a cliff is constrained by a repeatably testable model”, but pretty much whatever labels I use I will end up saying some bizarre things about how the world really works, so that doesn’t bug me too much.
EDIT: Based on discussions below, I revise my understanding of the Xs and Ys to:
(X,Y)minibearrex (“reality”,”set of models”).
(X,Y)shminux (“set of repeatably testable models”,”set of models”).
That is not quite the model of natural science I have in my mind. Yes, experimental testing is used to check models’ validity, but I never said that the results of this testing are determined by a single coherent whole ‘X’. The model of having such an ‘X’ works well in classical physics, including the unfortunate dude from the Simple Truth. It does not work well in QM. Specifically, it makes one ask questions like “what is the momentum and position of a particle, really?” (incidentally, this one EY has dealt with quite well in his QM sequence). Unfortunately, the mental model of a fixed underlying territory keeps leading one astray with questions like “do many worlds exist?”, “does wave function collapse violate relativity?”. Once you reformulate the questions as “what predictions does the MWI make?” and “How do we detect violation of relativity in the collapse model?”, the questions are immediately dissolved.
If all we have is observations and models and predictions, and the whole enterprise of trying to reason about whatever might underlie those observations is misguided, presumably abstract assertions like “there exists a single X” or “there exists more than one X” are also misguided except insofar as they can be grounded out in differential predictions about future events.
That is, I don’t see how the distinction you’re drawing here between those two superficially distinct accounts is at all meaningful on your ontology.
That is the minimal model I currently prefer, yes. And “the whole enterprise of trying to reason about whatever might underlie those observations” is not misguided, it’s useful for coming up with better models, but that’s all it is useful for. Assigning it any ontology is unnecessary.
I am not drawing any distinction, neither matters in the approach “build models and test them”. You are the one who started discussing some metaphysical “X”, I simply pointed out that your suggestion of a single “X” is not a requirement.
You raised the question of whether X refers to a “single coherent whole” entity or refers to multiple entities (e.g., multiple repeatably testable models).
Which, as I said, puzzled me, since my understanding was that on your account that’s a meaningless question.
And it sounds like we agree that on your account it’s a meaningless question.
I’m still not really sure why you chose to raise it in the first place, but I hope we can agree that the proper thing to do with a meaningless question is to point out why it’s meaningless and otherwise ignore it.
As far as I can tell, you are the one who invented “X”, no? I never said I need it, just pointed out certain flaws with it.
I suppose it is not meaningless in any model that includes X, and the one I favor does not. I hope we agree on at least that much.
Tapping out.
*tries to pronounce PUMWH… *fails miserably. I just wanted to note that a hypothesis that is specifically designed to not make new predictions (remember, it has exactly the same math as any other interpretation) is not a hypothesis, but a cocktail party talk.
Because map/territory is a very useful model in so many cases. Just not all of them.