Back Original

Highlights From The Comments On The Substack Religion Debate

[Original post: Why I’m Staying Out Of The Substack Religion Debate]

Ben Smith writes:

This post sure reads like participating in the object level debate, not just "writing a post about how I won't participate".

I did joke about this at the end, but I also want to defend calling this a post about how I won’t participate.

This essay presented a kind of hand-wavey heuristic for why I’m not spending lots of time examining the details of every individual argument. This is fine, everyone has to do this or be overwhelmed, but I think it’s important to keep hand-wavey heuristics for not participating in arguments, vs. actual participation in arguments, distinct. Little annoys me more than when I spend weeks carefully gathering citations and making a substantive contribution in a domain, and somebody responds with “well that’s obviously all bullshit because it vaguely reminds me of something I have negative affect towards”. It’s not necessarily wrong to deprioritize a topic because it vaguely reminds you of something that you have negative affect toward, but I would prefer these people admit they’re dismissing/ignoring it rather than claim to be engaging with it. I’m working on a longer post about this topic.

Chaotic Neutral writes:

I think Scott presents a great summary of the default skeptical position here.

But I would like to remind him that in his first Fatima article he accused me of having a trapped prior for following the exact same logic in explaining my casual dismissal of the miracle discourse. Seems a little inconsistent to me.

In any case, I think Scott and anyone who enjoyed his article would also enjoy Onid excellent and relevant essay (using Kolmogorov complexity!) to attack the idea of a “simple” infinity god.

I can see why Chaotic Neutral would think I’m being hypocritical or at least inconsistent, but let me offer a few defenses.

First, I used “trapped prior” in a specific way there, referring to Chaotic not feeling surprised by Fatima. My point was that the miracle Fatima, if we can’t explain it, provides some evidence for Catholicism (let’s say it should update us 10% towards it). We may say, after updating, that Catholicism still seems unlikely on net. If that’s true, then Fatima must be fake somehow. However, even though that should be our bottom line, we can’t go back and say “therefore, Fatima isn’t surprising at all, and shouldn’t even update us 10% of the way to Catholicism”, because that’s double-counting. The correct action imho is to continue thinking of Fatima as a weird unexplained creepy anomaly in our worldview, which we have some other reasons to think will get solved eventually. I thought there was a missing mood in some of the Fatima posts where people didn’t keep it open as a weird unexplained creepy anomaly.

(by analogy, suppose that you time travel back to meet Napoleon, and you notice that there’s a UFO hovering above his camp headquarters. You might, in the end, assume that, given the implausibility of aliens, you must have been hallucinating, or that it must have been a very weird cloud, or something like that. But you wouldn’t think “ah, nothing to see here, this doesn’t worry me in any way”. The fact that it was worrying and you genuinely didn’t know how to explain it would coexist with the fact that it didn’t fully move you towards the alien-Napoleon conspiracy theory and so you assume it must have some natural explanation.)

This doesn’t apply to this post, because I’m not using my atheism to dismiss or explain anything else, I’m just asserting it.

Second, as I said above, I was careful to claim this was a dismissal of the debate, and not a participation in the debate. I think it would be fine to say “I’m not wading into the miracles debate, because miracles are nonsense”. I think it’s worse to claim to have participated in the debate by saying that. I still get a trickle of DMs, maybe one or two per month, by people who are terrified of Fatima and want me to reassure them that it doesn’t prove they’re going to Hell (I’ve tried to help some of these people, but others are too compulsive about it, or strike me as not likely to be productive to talk to). I think that considering the people in this bin makes it easier to distinguish between a dismissal of the debate and a solution to the debate.

Finally, the charge of inconsistency would be stronger if I criticized CN but didn’t criticize myself. But I did criticize myself. In my post, I said things like:

I hope this was enough self-criticism to free me of the charge of hypocrisy.

Kveldred (blog) writes:

I enjoyed this, and broadly concur—but I’m a bit suspicious of the airplane argument.

On one hand, I can’t quite refute the “adding a limit is like adding an incapacity” point: it does seem equivalent, and equivalently illegitimate. On the other hand, I wonder if—in re at least the airplane bit, specifically—we might not be conflating two things: viz., (1) how many bits it takes to specify something, vs. (2) how much effort and time etc. it takes to build something.

That is, the one does not (have to) bear upon the other at all; we cannot necessarily reason from “the more-limited biplane is easier to construct than is the less-limited starship” to “the biplane *must be* simpler”. E.g.: “a sphere” is undoubtedly simpler than “a small sphere”, conceptually/semantically speaking—but that surely doesn’t require that constructing a sphere of unlimited (= unspecified?)² size will be easier than building a small one!

Or, put another way: is it a valid general principle that “an object with a simpler description ought be easier to construct than will be a like object with a more complex description”, such that “it’s harder to go ‘starship → biplane’ than the converse” is evidence for latter’s simplicity?

Well, it might be some evidence—i.e., simpler things probably are mostly easier to make—but seems to me that so many factors beyond complexity can go into “ease of construction” that it can be, at best, only suggestive of simplicity.

I agree I didn’t do a great job with this part. I was trying to avoid having to describe Kolmogorov complexity too, uh, complexly, and that made the explanation too basic to hold up to full scrutiny. Let me try again.

A starship might be simpler to describe in English than a biplane. The starship is “aircraft without limits”, and the biplane is “aircraft that can only go fly at X miles and hour, go to Y altitude, will tear apart if subjected to Z stress, etc”.

But a starship is harder to draw schematics for than a biplane. The biplane has fewer components and requires less precision; the blueprint could fit in a smaller file. Popular misinterpretations of simplicity rely on English-language description; Kolmogorov complexity is closer to schematics: it asks how long a working program that describes or simulates the object would be.

I think this passes Kveldred’s counterexample. A 3D graphics file describing a sphere of indeterminate size could be shorter than one describing a sphere of size n, because it doesn’t need to include a size parameter, and this holds even if size n is very small.

But hopefully it illuminates why I find “mind without limits” to be fake simplicity: it’s not a schematic; it’s not a file capable of simulating such a mind.

There’s a file capable of simulating a nematode mind; nematodes have only a few hundred neurons, and the file (depending on how you measure) is only a few hundred kilobytes. When you run it, the simulated nematode brain on your computer will try to do nematode things like wiggle and . . . whatever else nematodes do. Some sci-fi envisions a future where we can run emulated human brains on computers. Since human brains have billions of neurons, this would require a much larger file - gigabytes or terabytes - but if you had such a thing, you could do anything a human could do - solve math problems, write a novel - by running the program.

In the real world, the closest equivalent is AI. GPT-1, a very simple AI with few capabilities, is about 100 MB; GPT-5, a modern AI with many more capabilities, is about 10 TB. Again, these are the literal sizes of the weights; once you have this 10 TB file (plus the program to read it), you can run the AI on your computer and ask it questions.

All of this suggests that it takes kilobytes to megabytes to describe very simple minds, and gigabytes to terabytes to describe more complex minds with human-level capabilities.

But IIUC, Adelstein is claiming it takes only 19 bytes - the English language string “mind without limits” - to describe God. Or maybe even 1 byte (the number one, representing maximal everything) or 0 bytes (pure unconditioned existence). On the Kolmogorov complexity interpretation, this is false: you can’t type “mind without limits” into a Python terminal, run it, and get a perfect simulation of God who can answer all your questions for you. It’s also common-sensically false; if small weak minds are 100 MB, and powerful human-level minds are 10 TB, it’s unlikely that the most powerful and perfect mind of all is 0 - 19 bytes.

So my claim is that if you use “mind without limits” as an entity in your explanation, the penalty is however many bytes it would take to fully schematicise a mind without limits - for example, the size of the weights file of a superintelligent AI that has infinite capabilities and knows everything.

This is a known and common argument for atheism; the theist counterargument is that Kolmogorov complexity is meaningless without describing the language/interpreter of the schematic/program, and if it naturally privileges minds, then the description of mind could be very short. As an analogy, if you were using a programming language with a library for God, then the code for a simulated God that knows everything God knows could just be the one-line program “import God”. Or, as per Adelstein’s model:

> import Mind;
> limits = 0;

…and the reason this works for Mind but not aircraft is that there isn’t a library called “aircraft” which the interpreter can import.

This is why I then have Part III where I argue against computational dualism. If you could do this, you would expect our brains to try the “import Mind” trick, but they definitely don’t - they build it up laboriously, neuron by neuron. If “Mind” is an existing function in your library, this is terrible coding technique!

Imagine if I said that the universe was created by a universe-creating factory using skilled German union labor. A natural objection would be “how did that exist before the universe?” The counterobjection “I’ve decided that universe-creating factories and Germans might be infinitely simple, and you can import them with a one-line command” would sound hollow, because all the factories we know are very complex, and there’s a whole long geologic/biologic/evolutionary/civilizational history for why factories, unions, and Germans exist. This is my objection to “mind” as a one-line primitive too.

Reginald K writes:

Stephen Barr says (I think in Modern Physics, Ancient Faith) that there’s little anthropological evidence that gods were originally posited as explanations for natural phenomena that couldn’t otherwise be explained, and that in the last couple of millennia, where we have better evidence, it shows the opposite: Religious natural philosophers/scientists were typically inclined to first reach for natural reasons for unexplained phenomena instead of immediately claiming God did it. (Newton and mutual planetary perturbations was a notable exception.) Anyone know how much truth there is to this vs. the standard “God of the gaps” argument?

I don’t find this convincing. Reginald seems most interested in arguing that classical philosophers and medieval/Renaissance Christians didn’t do this, which fine, whatever. But the two examples I gave (the Storm God, the Sun God) were both pagan.

Trying to generalize from modern primitive people to the original condition of humankind is always fraught, but modern (and historical) primitive people seem to do this, to the point where individual animals and terrain features are regularly mythologized. Snakes often lose their legs because a god punished a primordial snake committed some infraction (in Chinese myth, the Jade Emperor punished a snake for hunting humans; in African myth, a generic Creator punished a snake for boasting about its beautiful legs too much). Australian aborigine tribes have myths about how various Dreamtime spirits created each watering hole and mountain in their ancestral territory. Just the myths about thunder and lightning could fill an entire book: in Norse myth, thunder is Thor throwing his hammer Mjolnir at giants; in Shinto myth, it’s the god Raijin playing his drums; in Finnish myth, it’s the creator Ukko making love to his wife.

Even the more “sophisticated” traditions like Greek and Jewish religion have plenty of these. The story about Persephone and the pomegranate doesn’t make sense outside a context of trying to explain the existence of seasons. Judaism’s “how the snake lost its legs” story is that God punished it for tricking Eve into eating the apple (pain during childbirth is also a punishment for Eve’s transgression). Babel explains why people have different languages. Even the story of Lot’s wife looking back at Sodom is plausibly an attempt to explain a particular pillar of salt that stands on that spot.

Given all this, I don’t find it relevant that Leibniz or whoever was good at preferring natural explanations to divine ones.

archsine writes:

What would you say is the fundamental difference between a theism with a god, and the chance that we are living in a simulation? What criteria for this argument against God existing would not also apply to the simulators existing?

Other people asked similar questions about superintelligence. Sorry, I don’t like to make fun of commenters, but to me this sounds kind of like “You don’t believe in ghosts? But ghosts are just pale white people who are dead and dress in sheets, so what’s the difference between that and believing in ancient Romans?” Yes, if you squint the results look similar, but they get there in different ways that overturn different amounts of our previous worldview!

To me the two biggest differences are:

First, neither simulation theory nor AI make any claim about dualism, mind as independent from matter, or the simplicity of mind or “perfect mind”. They posit completely normal minds, made of matter like ours, created because of some specific history. That’s totally different from claiming that Mind is a primitive of the universe and “without limits” is a predicate you can cheaply apply to it, and I think Matthew Adelstein and nearly every philosopher of religion would agree on this point.

Second, neither simulation theory nor AI are conspiracy-theory-like attempts to explain features of the world. “The superintelligent AI build the Pyramids” or “the simulators orchestrated the rise of Napoleon” would be powerful-hidden-entity explanations for specific otherwise-inexplicable anomalies. But the evidence for superintelligent AI (existing in the future) or simulators (existing now) doesn’t go through that kind of abductive reasoning. In the AI case, it goes through reasoning that if AI continues gaining capabilities at the current rate, we would expect it to eventually become very capable. In the simulator case, it goes through reasoning about how if each simulation-capable species simulates many worlds, there should be more simulations than natural worlds.

…but I have trouble listing all the differences because they’re totally different theories. How do you explain believing in Romans but not ghosts, even though they’re both white dead people in sheets? You just have to consider everything you know about ghosts, everything you know about Romans, and conclude that they’re different things that only resemble each other on the most superficial level. Likewise, I believe in prediction markets but not Nostradamus, in Toyota Camries but not the divine chariot-throne merkavah, and in the three branches of the federal government but not the Holy Trinity.

Matthew Adelstein wrote a response, cleverly titled Why I’m Staying Out Of Responding To Scott Alexander’s Piece. Some of his points are covered above, others are minor, but I respond here to what I consider the most important.

It would be one thing if God were just being trotted out to explain one or two random anomalies. Instead, God provides a good explanation of about 30 features of the world. Now, when you pad your list with 30 arguments, they won’t all be winners. But there are a bunch of good arguments on the list—at least the first nine strike me as very forceful, even considered on their own.

These arguments don’t require positing weird eccentric motivations of God—they posit more general facts like “a perfect being would prefer life to random floating clouds of hydrogen gas,” “if God’s going to create at all, it’s not terribly unlikely he’d do a lot of creating,” and “God prefers agents with minds to random mental static.” None of these require that we guess God’s favorite toothpaste flavor!

The aliens-helped-Napoleon theory can also explain thirty features of Napoleon’s rise. For example, it explains how he became leader of France, how he gathered such a big army, how he won at Jena, how he won at Austerlitz, how he escaped Elba, what happened to him in the Great Pyramid, why he was predicted by Nostradamus, etc, etc.

There are two tricks here. First, it treats every surprising thing as an anomaly to be explained, then gives one big overfit solution which solves every anomaly by virtue of its bigness and overfitness. Second, it gets to fit itself to the reason we’re talking about this at all (in this case, that Napoleon was very successful and won a lot) and count each sub-example as a different thing being explained.

The God example is similar. We have to exist and reason to be having this discussion, therefore whatever universe we find ourselves in will be optimized for human existence and reason, therefore insofar as we haven’t explained the universe perfect enough for there to be no anomalies, there will be lots of anomalies in the direction of the universe being more optimized for human existence and reason than expected. You can resolve this by positing a God who’s optimizing for human existence and reason, even though this is not at all obvious even given the definition of God (maybe God should be a negative utilitarian and avoid creating things, or maybe He should just create pure hedonium without associated computation).

Suppose I come back to my home and see that many of my valuables have been stolen. However, surprisingly, a good number of valuables weren’t taken. Positing that a burglar broke in involves thinking:

“Why did this anomalous thing happen? Because there was a powerful entity trying to make that specific thing happen. No, you can’t see it, it hid all evidence of its own existence. Yeah, the outcome doesn’t match any sane entity’s plausible goals, but I can sort of tell a story for why they kind of sort of vaguely resemble those goals.”

This is in fact a famous failure mode of schizophrenics; if one of their objects is in the wrong place, they assume someone was rooting through their things for sinister reasons, when more often eg they just forgot where they put it.

Of course, sometimes the answer really is a burglar, but this should have a high prior given that we know burglars exist. The entire tribe of burglars is not hidden, only certain actions by individual members.

Likewise, it’s not a “conspiracy theory” to claim that Mossad sabotaged Iran’s nuclear program, even though it awkwardly involves the Jews secretly conspiring to destroy stuff. At least part of the reason is that we know Mossad exists, even if we don’t know all of their specific agents and actions, and we know more or less what they want (the sabotage of Iran).

If we knew that gods existed and performed miracles, the application to any specific miracle would of course be on much firmer ground!

Suppose you’re considering between two theories of physics. One might just be totally horribly broken and incoherent. But if it works, it’s ridiculously simple. The other is definitely very complicated. Overall, you should prefer theory 1 to theory 2. Because theories that are a bit simpler have a much higher prior, even with only a low prior on theism being coherent, it will still gobble up a lot more of the probability space than naturalism. I talk about this more here, if you’re interested […]

Third, there are a lot of other potential ways of God being simple. [List of ways]. One of these accounts succeeding doesn’t seem outrageously unlikely. It seems that you should think there’s at least a 1% chance that one of these accounts succeeds on which God has at least a 1% intrinsic probability. Combined, that gets you at least an intrinsic probability of 1/10,000.

This seems like the epistemic version of Pascal’s Wager, with equally dire failure modes. For example, real quantum gravity seems quite difficult to figure out and may be complex. But gremlin-based quantum gravity (the differences between quantum mechanics and relativity are smoothed over by magical gremlins) would be very simple, as long as we posit that the gremlins are perfectly simple entities that require no further explanation. Since we don’t do that, I think Adelstein must be applying the complexity penalty in the wrong place, or failing to sufficiently penalize the hypothesis that something is perfectly simple. I admit that I have no good solution to any form of Pascal’s Wager, including epistemic forms.

Adelstein on simplicity:

Suppose we have two theories. Theory 1 says the universe is infinite. Theory 2 says the universe is 10 quintillion meters across exactly. Should we say theory 2 has a higher prior because theory 1 is “subtracting negatives, ie adding?” No! Sometimes theories that posit more total things are simpler. The theory of modal realism—that every possible world exists—is a lot simpler than the theory that every possible world exists except one, even though it has one more world.

When does a theory pay a cost for having a limit? Dustin Crummett has a great post on this, but two contributing factors:

This is why the theory that nature is uniform is much better than the theory that nature breaks down in one random patch somewhere. The simplest way to describe the patch theory is to describe uniform laws and then add a proviso about them not being uniform in some region. The theory is also arbitrary. Why do laws break down in that patch rather than some other one?

So if we can characterize all of God’s properties in terms of a single unlimited thing set to infinity, then theism is a very good theory. I claim we can: God is a mind set to infinity. And minds are fundamental bits of the universe, not reducing to any non-mental thing.

I agree that modal realism seems simpler and more likely than all-worlds-but-one realism.

But with minds, I still think this is incorrectly using this complex concept as a primitive. For example, an unlimited mind should be able to add numbers. And adding numbers requires certain algorithms, which are implemented in different ways across calculators, LLMs, and humans, but which are never zero bytes long.

The “limit” of GPT-5 is just where OpenAI stopped building it. When they stopped building it, it had 10 TB of weights, and it was smart enough to add numbers but not smart enough to cure cancer.

When you “describe” GPT-5 in its native language of weights and parameters, there’s no term for “limits”. It either has enough weights and parameters and algorithms and heuristics to add numbers (or cure cancer), or it doesn’t.

Both “mind” and “limit” are emergent terms that come when we see so many carefully-constructed-neuron-by-neuron artifacts that we need a word to refer to them. But the word itself carries no simplifying power. I can use the term “Taj Mahal” to quickly point to a specific building, and then if I wanted to describe some other lesser building I might have to describe it as “like the Taj Mahal, but smaller, and with fewer minarets”, and then that other building’s description would sound longer and more complex than the Taj Mahal. But that’s just an illusion caused by “Taj Mahal” being a convenient concept handle. At base, you actually have to describe the Taj Mahal and where every brick and window is, and that description will be long.

Likewise, you can’t just say “mind” and “limits” and expect Reality to know what you’re talking about. You have to tell it what a mind is, in its own language. And that’s going to involve a lot of specifying synaptic connections.

As an analogy, consider Scott’s favorite theory: Tegmark’s mathematical realism. On this theory, every single mathematical structure is instantiated. Every world that you can describe with math exists. Across the infinite worlds, there is infinite computation. But that’s not because you’re building in at the ground level infinite computational ability. Rather, you’re deriving it from the fundamental elements of the theory.

I think it’s simple in principle (though impossible in implementation) to specify an (infinite) computer program that simulates Tegmark’s mathematical realism. You would first go through all one-character programs in some language, all two-character programs in that language, and so on. The computation would fall out naturally - you could observe where in the structures each thought was being generated.

(if this is the stupidest thing I’ve ever written and this is provably mathematically impossible in five different ways, sorry)

This is different from Adelstein’s world where you can just write “import God” and all the computation happens under the hood.

But why is a mind “allowed to wax to its full power” omniscient and omnipotent? Why not just infinitely intense conscious experience? I agree this step is the shakiest, but let me say two things about it.

One: just intuitively, a mind that doesn’t know something is limited in some essential respect. Something lies outside the bounds of its awareness. Imagine a mind wishing for something to be done but being unable to bring it about. Definitely that’s limited in some way! Now, there’s a deeper question about whether this kind of unlimitedness gets to be built into our fundamental theory, but it at least doesn’t seem like the answer is obviously negative.

I think this is based on the same argument I criticized - rounding off consciousness to “mind”, and then noting that a “mind” should surely be able to think and so if you didn’t let it then it would be limited. I think if you didn’t start with mind - if you just called it “consciousness” - this problem would go away. What if unlimited consciousness that could do anything wanted to think? Well, it could, by the definition of unlimited, but why would it want that? What if unlimited consciousness that could do anything wanted to paint the Taj Mahal pink? Well, it could do that too - but again, why would it?