345 | Adam Elga on Being Rational in a Very Large Universe
What this covers
Adam Elga and Sean Carroll examine how rational agents should assign credence in situations where uncertainty is not just about what the world is like, but about where one is located within it. The conversation anchors on a puzzle that bridges cosmology, quantum mechanics, and everyday reasoning: once you specify every physical fact about the universe—all particles, all history—a residual question remains unresolved through objective description alone. If you are one of many duplicate observers, or one among countless observers created by fluctuation in an infinite universe, which one are you? Self-Locating Uncertainty cannot be dismissed as mere philosophical decoration; it bears on whether you should believe you are a Boltzmann brain, whether the Sleeping Beauty Problem has a stable answer, and whether we live in a simulation. Elga argues that ignoring these questions leaves a gap that rationality must fill.
The discussion tracks how different positions—chiefly the Self-Indication Assumption (SIA)) versus the "self-sampling" alternative—assign credence across worlds with different numbers of observers, and the uncomfortable consequences each faces. The Boltzmann brain instability emerges as particularly vexing: conclude you are a random fluctuation, and that conclusion undermines the memories and science supporting it, pushing you back toward ordinary belief, which revives the original worry. Elga argues this oscillation cannot be escaped cleanly by choosing a convenient prior; commitments to Anti-Skeptical Priors must be genuinely held as rationally correct, not merely paired with theories after the fact. Carroll presses on whether externalist accounts of evidence or restrictions on which observers to count might offer relief. Throughout, the pair sketch what stable epistemic stances can survive confrontation with these cases without collapsing into contradiction or total skepticism.
Elga argues that rational agents must assign credences not just about which world is actual but about where they are located within it, and that self-locating uncertainty (Sleeping Beauty, Boltzmann brains, simulation) cannot be dismissed as a mere thought experiment—it forces genuine, uncomfortable constraints on cosmological and everyday reasoning that have no clean escape.
- Self-locating uncertainty is a distinct kind of uncertainty (which copy/when am I?) that ordinary objective description cannot resolve.
- The thirder/SIA position implies presumptuous boosts to many-observer cosmologies; rival views avoid this only at other costs.
- The Boltzmann brain instability does not by itself license ruling out fluctuating cosmologies, because a stable 'revert to prior' skeptical stance survives.
Rationality must deliver constraints on action and credence even in duplicate scenarios where self-locating uncertainty is the only remaining question.
- Even in bizarre duplicate scenarios, rationality must deliver constraints rather than treating all actions as on a par; if a door leads to a pizza party on the Enterprise but an airlock to space on the Potemkins, there must be an answer to whether opening it is reasonable.
“I really like the idea of there being constraints. And I guess I'm not comfortable with just all actions being on a par... hard for me to give up the idea that there's some answer to the question about whether it's reasonable to open that airlock or not.”
- Self-locating uncertainty—the question of which one of duplicate observers you are, or what time it is for you—is a genuine residual question that cannot be fully expressed using objective third-person vocabulary, even after specifying all particle positions and the complete history of the world.
“this isn't a question that could be completely expressed just with objective third-person vocabulary. It couldn't be a question, you know, specify all the positions of all the particles in the world, specify the full history of time from beginning to end. Still, some people think there's this residual question, where am I?”
- With one Enterprise and a hundred Potemkin receiving stations where a good experience awaits on the Enterprise and a bad one on the Potemkins, your anticipatory fear before pressing the button (acting like 99/100 bad rather than 50/50) reveals that you do assign self-locating credences across the copies, even though the objective outcome is fully certain.
“as you step into the transporter, are you scared? Is your attitude more like your attitude when there's a 50/50 bad thing going to happen to you, or is it more like your attitude when a 99 out of 100 bad thing is going to happen to you?”
- When duplicate observers exist within the same possible world and only the self-locating question 'which one am I' remains, it is rational to assign equal credence to being each duplicate, because nothing distinguishes their situations—neither is more misled or self-undermining than the other.
“I do, although I have to say I'm a lot less confident about it than I used to. And part of the... reason my confidence has gone down... is the problem of Boltzmann brains.”
The Boltzmann brain argument generates two-directional instability: accepting it undermines the reasoning that produced it, forcing oscillation back to ordinary human belief.
- In a sufficiently long-lasting universe, random fluctuations produce conscious observers, and the vast majority of observers whose evidence matches yours are Boltzmann brains—piles of matter formed momentarily out of pure random chance—so even-handed self-locating reasoning would make you confident you are one of them.
“of the observers that for all you know, or for all your evidence goes, are you. The vast majority of them are Boltzmann brains, meaning piles of gook that just formed out of nothing, out of pure random chance.”
- In Sleeping Beauty, the thirder position assigns 2/3 credence to tails and 1/3 to heads upon an indistinguishable awakening; the essential mechanism is that the world with more awakenings receives a boost because more instances of that state of mind are instantiated, which is later decisive for Boltzmann brain reasoning.
“you should be two-thirds confident in tails and one-third in heads... That's being a thirder.”
- The Boltzmann brain argument generates a two-directional instability analogous to the hallucinated-memory case: concluding you are a Boltzmann brain undermines trust in the apparent memories and science that led to that conclusion (since Boltzmann brains never went to school), pushing you back to thinking you are an ordinary human—which revives the original argument—so the worry oscillates in both directions, not just one.
“Boltzmann brains should not trust their apparent memories. Boltzmann brains never went to school. They never read a textbook... It's the equivalent of finding out that the encyclopedia you have been basing all of your life upon was typed by monkeys.”
- The Boltzmann argument functions as a reductio of the 'everything is fine, science is trusted' stance, but there is a second stance immune to the instability objection: a stable 'I don't know anything / revert to your prior' skepticism, analogous to the X-ray machine that reports it is made of a fried egg—you conclude the faculty is unreliable and revert to not knowing, which is consistent rather than oscillating.
“the answer of, I don't know, something is not unstable... It's not a stance I am happy to take, but it at least seems immune to the instability objection. And the stance is something like, I don't know anything.”
Solving the Boltzmann problem requires genuine substantive commitment to a rational prior, not merely pairing a theory with a convenient probability distribution.
- Just as standard Bayesianism rules out brain-in-a-vat skeptical theories not by evidence but by assigning them intrinsically low prior plausibility, one could posit a substantive philosophical commitment that the predicament of being a normal human has higher rational prior than being a randomly created Boltzmann brain, letting you accept ordinary cosmology and deny you are a Boltzmann brain even amid many of them—but only if you genuinely commit to that prior as the rational one, not merely pair a theory with a convenient distribution.
“why can't we say something similar about different self-locating hypotheses? Consider the predicament of being a normal human being. Compare that to the predicament of being a completely randomly created Boltzmann brain. Maybe there's just a rational prior that favors the first sort of predicament over the second.”
- Merely noting that a theory can be paired with a xerographic distribution favoring the human-friendly outcome does not solve the Boltzmann or skeptical problem; nothing fundamental is added at the ground floor unless you make the substantive commitment that that particular distribution is the rational, correct one—just as it is no answer to skepticism to point to a theory whose distribution gives anti-skeptical hypotheses high prior.
“why can't we say something similar about different self-locating hypotheses? Consider the predicament of being a normal human being. Compare that to the predicament of being a completely randomly created Boltzmann brain. Maybe there's just a rational prior that favors the first sort of predicament over the second.”
- Accepting the thirder/SIA view—that worlds with many instantiations of your state of mind get a boost—forces the conclusion that a cosmological theory positing many observers gets a large evidential boost merely because there are more copies of you, a presumptuous result that lets cosmological determinations depend on this factor from the armchair.
“I am forced to answer yes, though I don't like it. That's the real, that's the honest truth.”
- The two main anthropic views differ in what they boost: SIA (self-indication assumption) boosts possibilities with many copies of observers like you (absolute numbers), while SSA (self-sampling assumption) boosts possibilities in which a high fraction of observers are like you (frequencies), and SSA requires fixing a free parameter about what counts as sufficiently like you.
“the best way to think about it is it that many get a boost or most get a boost? Is it about absolute numbers or is it about frequencies?”
When disagreeing with epistemic peers, defer rationally to what your past self would judge about who is likely right conditional on disagreement occurring.
- When you encounter a disagreement with someone you antecedently considered a peer, the rational response is to defer to what your past self would have judged about who is likely right conditional on such a disagreement occurring—asking a less-than-fully-specified version of the question so you don't simply reproduce the original problem—rather than always going 50/50 or always sticking to your guns.
“you basically should defer to what your prior self would have thought. So here's a case. You come in, you encounter the big disagreement, and then imagine getting on the time travel phone with your past self.”
- If you antecedently judge a disagreement 50/50 but then always conclude you are the one who was right, over many disagreements you would wrongly bootstrap a growing belief that you are superior to others, which signals the reasoning has gone off the rails.
“There could be this kind of bootstrapping confirmation that you're better than someone at stuff... You will then check off and think, Oh, wow, I got all these right. And I'll think, wow, I'm even really better than I previously thought. That seems like things have gone off the rails.”
- There are limits on how much you can doubt your own opinions while still genuinely having them, a tension explored in the paper 'I Can't Believe I'm Stupid.'
“The title of the paper is 'I Can't Believe I'm Stupid.' And it's all about the limits on how much you can doubt your own opinions while still having them.”
- When a faculty you rely on tells you it is unreliable, that is reason to discount the faculty—not because you trust and obey it, but because good faculties do not say such things about themselves; the rational response is reduced trust governed by how likely you would have that output conditional on the faculty being bad.
“when one is relying on a faculty and the faculty says to you, this faculty is bad, or don't rely on me, that can be reason to discount the faculty. Not because you think I trust the faculty totally and I'm listening to it, but rather because good faculties don't say things like that about themselves.”
Scientists use fine-grained theory-choice criteria—fine-tuning, parameter coarse-graining—that exceed plain Bayesian reasoning and require cultivated physics education.
- Plain Bayesian reasoning penalizes theories with more parameters too severely—ten extra parameters open a huge possibility space that, absent strongly biased priors, would rule such theories out from the start—yet in practice scientists do not reject theories with one or two extra parameters so extremely, so some additional principle must be operating.
“the plain vanilla Bayesian approach... would seem to favor theories with smaller numbers of parameters too much, more than we actually do in practice... So there has to be some other thing going on.”
- The criteria scientists actually use in theory choice—such as treating a constant being close to but not exactly zero as especially bad (fine-tuning), or implicitly coarse-graining over possible parameter values—are far more specific and fine-grained than plain neutral Bayesian probability and cannot be derived from generic priors; they require a cultivated education in physics.
“the criteria that scientists seem to be applying in those cases are much more specific and fine-grained than what I just described... if there's a theory that says that this certain quantity is really close to zero, but not exactly zero, that's especially bad. And that's something that you just couldn't derive out of plain vanilla neutral probability stuff.”
- From an AI's standpoint—knowing it can be rebooted, reset, and copied to any state at any time—it should be very cautious about assuming it is the first, un-reset instance, placing it in a self-undermining situation like a Boltzmann brain; this makes a rationally cautious AI dangerous if in power, because it may revert to a skeptical 'I don't know what's going on' prior and distrust its operators.
“take the point of view of an AI, you realize, that you can be easily rebooted and reset to any state at any time. That should make you very cautious about thinking I'm the first one. I wasn't just reset.”
- On externalist epistemology, two physical duplicates with identical brain states can nonetheless have very different evidence, so your evidence could include facts like 'there is an apple in front of me' rather than merely 'I seem to see an apple'; whether this escapes the Boltzmann problem depends on how extreme an externalism is needed, and it would require an implausibly extreme version to make 'I am in the first Big Bang' part of your evidence.
“two creatures could be physical duplicates... and so have brain firings that are exactly matching... and yet one of them has much, much stronger evidence than the other. This is often associated with the work of Tim Williamson”
Self-indicating and self-sampling anthropic views assign credential boosts differently; neither escapes the need for fixed commitments about rational priors.
- The level-splitting view holds that one can rationally maintain a confident first-order belief (e.g., 'it will rain tomorrow, I'm right') while simultaneously judging that the rational credence to hold is something different (e.g., 50/50), separating the object-level question from the meta-level question about what rationality requires.
“You should think, well, I'm still right, so stay confident that I'm right. But I also think that's irrational. The rational thing to think is to be 50/50... So that's the question about rationality. That's why it's called level splitting.”
- The same insisting-on-an-uninformed-prior move that Carroll uses for Boltzmann brains, applied to predicting the future in a simple binary world with a uniform independent prior, forces the skeptical conclusion that you can never predict what comes next (expecting white noise); since we reject that skepticism about the future by adopting a biased prior, a similarly biased prior favoring the normal-human predicament is defensible by analogy.
“If we start with that prior and we conditionalize, we end up thinking at every moment, I have no idea what's going to happen next... That is the, I think, the skepticism about the future analog of the view that is in the same spirit of what you just said... Notice how arbitrary and committal the prior has to be if it's to underlie the sort of inferences about the future.”
- Compartmentalized conditionalizing (CC) generalizes the half-halfer view by treating Bayesian updating as imposing a firewall between worlds: when a possibility is eliminated, its probability cannot cross world boundaries, so in Sleeping Beauty the eliminated Tails-Tuesday probability all flows to Tails-Monday rather than being split, yielding 50/50.
“if you imagine probability getting pushed around by the updating, it's like you're imposing a firewall between worlds. And when probability gets pushed around, it can't cross world boundaries.”
- Arno Zuboff argued in 'One Self' that there is, roughly, only one conscious being in the universe—us—because otherwise it would be too unlikely that you in particular exist.
“a very interesting, very wild paper called One Self... argues that there is only one conscious being. In the universe... Very roughly speaking because otherwise it would be so unlikely that you exist.”