200 | Solo: The Philosophy of the Multiverse
What this covers
This solo episode tackles a knot of problems that arise when physicists try to reason about probabilities in universes containing many (or infinitely many) observers. Sean Carroll defends multiverse scenarios—whether from inflation, string theory, or quantum mechanics—as legitimate science because theories proposed to explain observable data happen to predict them; the real difficulty is not whether multiverses exist but how to think rationally about your location within one. The episode ranges across three main puzzles: why standard principle of typicality reasoning grants unwarranted leverage to armchair philosophy, what goes wrong with both major statistical approaches to multiverse probability, and why Boltzmann brain scenarios should be rejected on grounds of cognitive instability rather than observation.
Carroll argues that the consensus principle of mediocrity—the assumption that you are a typical observer—sounds humble but is actually presumptuous, since it lets you infer facts about the distant future or the existence of distant aliens from the mere fact of your own experience. Both the world-first Self-Sampling Assumption (which generates the Doomsday argument) and the observer-first Self-Indication Assumption suffer this flaw. He advocates instead for Radford Neal's Fully Non-Indexical Conditioning: condition on everything you know about yourself except your location in spacetime, and assume you are typical only among observers macroscopically identical to you. This preserves anthropic predictions about the cosmological constant—as Steven Weinberg demonstrated in 1988—without the presumptuous moves. Carroll also confronts the Boltzmann problem directly: in eternally fluctuating universes, there exist vastly more random fluctuations of your entire macroscopic past and present than thermodynamically normal observers, so even excluding bare brains, you should expect tomorrow's telescope to show no cosmic microwave background. The solution is not observation but a principled refusal: belief that you are a random fluctuation is cognitively unstable, undermining the reasoning that led to it.
Carroll argues that multiverse scenarios are legitimate science because they affect how we explain observable data, but that the standard 'principle of typicality' used to reason about probabilities in a multiverse is secretly presumptuous; instead we should reason that we are typical only within the set of observers macroscopically identical to ourselves (fully non-indexical conditioning), supplemented by ruling out cognitively unstable Boltzmann brain cosmologies.
- Each multiverse type (cosmological, many-worlds, eternally fluctuating) is a prediction of theories proposed to explain data, not invented for fun.
- Both world-first and observer-first typicality approaches grant unwarranted armchair leverage over the universe (Doomsday/Jovian and presumptuous philosopher problems).
- Boltzmann brain scenarios should be rejected a priori because believing you are one is cognitively unstable, not because you can observe you are not one.
Multiverse theories are scientifically justified because they are predictions of inflation and string theory, not invented for aesthetics, and make definite claims despite practical unobservability.
- The popular objection that the multiverse is unfalsifiable misreads Popper, who worried about theories like Marxist history or Freudian psychoanalysis that could explain anything; multiverse theories do say something definite (they predict some things and forbid others), the problem is only that what they predict is unobservable in practice, which does not make them unscientific.
“all these different multiverse scenarios are absolutely un-disputably saying something than saying non other things. The problem is that what they're saying happens in the universe are things that we can't see... So it's saying something definite, but you don't know”
- The cosmological multiverse was not invented because it sounds cool; it is a prediction and consequence of theories (inflation, string theory) proposed to explain data, so it must be judged by judging those theories rather than dismissed on its own.
“the cosmological multiverse is not a theory. It is a prediction of various theories, and it is a consequence of those theories, but you should judge the theories that it's a consequence of not the prediction all by itself”
- Weinberg argued ten years before the cosmological constant was observed that if its value is set not by a dynamical mechanism forcing it to zero but by anthropic selection across regions with different vacuum energies, then it should be observably small but nonzero, because observable values compatible with our existence outnumber unobservably tiny ones; this prediction was vindicated by the 1998 discovery of cosmic acceleration.
“Steven Weinberg pointed this out 10 years before we discovered the cosmological constant. He pointed out specifically that if the explanation for the vacuum energy is not some dynamical mechanism that sets it equal to zero, but rather some anthropic selection... then you should predict that the cosmological constant should be observable”
- Whether or not you take a multiverse seriously affects your actual scientific practice: if you reject the cosmological multiverse you must find a dynamical mechanism explaining the vacuum energy's value, whereas if you accept it that value can be an environmental selection effect; similarly your stance on many-worlds determines whether you must posit hidden variables or objective collapse, so the multiverse is science because it plays an explanatory role in observable physics.
“If you don't think that the multiverse makes sense or is there, then it is beholden on you to come up with some dynamical mechanism that explains why the cosmological constant has the value we observe. If you do think that the multi-verse is there, then arguably, you don't need to do that, it's just an environmental selection effect”
Typicality reasoning in multiverse contexts is poorly defined and generates absurd conclusions like the Doomsday argument regardless of which approach you adopt.
- There are two main ways to apply typicality: the world-first approach (self-sampling assumption) assigns a prior to each cosmological scenario and then assumes you are typical within it, so a single observer in a small universe is more probably you than any one observer in a big universe; the observer-first approach (self-indication assumption) assumes you are typical across all observers in all scenarios, boosting scenarios that contain more observers.
“world-first says, 'Give your prior probabilities to all the different scenarios and then assume you're typical within them.' observer first says, 'Assume that you're somehow typical in the space of all observers in all possible worlds and therefore favor those scenarios with more observers in them.'”
- The world-first approach generates the Doomsday argument: if we assume we are typical humans, then finding ourselves within the first ~100 billion humans is likely in a scenario where only 200 billion humans ever live but extremely unlikely in one where 200 trillion live, so we can conclude from the armchair that humanity will end soon, an unwarranted leverage over the future.
“I predict just on the basis that I am a typical person, that humanity won't last for much longer. That's why it's called the doomsday argument”
- The principle of typicality (also called the principle of mediocrity or Copernican principle), holding that we should reason as if we are typical observers in some reference class, is more or less consensus among multiverse cosmologists, but Carroll argues it is not well-defined and not a good idea, partly because it is obviously false that he is a typical observer (most people lack podcasts) and partly because the reference class is undefined.
“some version of this idea is more or less universally accepted among modern cosmologists... And number two, I don't think it's very well-defined or a very good idea at all, I think we need to do better”
- Claiming you are a typical observer sounds humble but is actually a strong, presumptuous claim that typical observers are like you (in your era, on Earth not Jupiter), granting enormous armchair leverage over the rest of the universe; both world-first (Doomsday, Jovians) and observer-first (presumptuous philosopher concluding the universe is infinite) approaches inherit this same flaw.
“the proposal that you are a typical observer in the universe is really proposing the typical observers in the universe are like you... So you're giving yourself enormous leverage over the rest of the universe by assuming that there aren't many, many more observers that are not like you”
Boltzmann brain scenarios are ruled out not by observation but by cognitive instability: trusting your reasoning to conclude you are a fluctuation undermines the reasoning itself.
- The real Boltzmann problem is not minimal brains but 'Boltzmann yous': in an eternally fluctuating universe there are vastly more random fluctuations matching your entire macroscopic state (including your whole past light cone) than thermodynamically sensible observers, so even granting Hartle and Srednicki their fiat exclusion of brains, you should still expect to be a fluctuation whose telescope will show no microwave background tomorrow, which falsifies their scenario.
“the problem is not Boltzmann brains, the problem is Boltzmann yous. The problem is, whoever you are, whatever you think about your current macroscopic state, you could have fluctuated randomly into existence”
- In the many-worlds view with an infinite-dimensional Hilbert space of quantum gravity, de Sitter space can settle into a static quantum state where the thermal spectrum is only what you would measure if you interacted with it, not a record of dynamical fluctuations coming and going; so when nothing disturbs the state, nothing actually happens, and Boltzmann brains need not arise.
“when you're not measuring it... when there's nothing disturbing the state there's also nothing happening. There's no dynamical fluctuations. The thermal-ness of this quantum state is a statement of what you would observe, but not a statement about the dynamical things coming and going, like brains coming into existence”
- If our universe's future is an eternal de Sitter thermal state, then like a box of gas at fixed temperature lasting forever, rare thermal fluctuations will occasionally decrease entropy; smaller fluctuations are far more likely than large ones, so the minimum fluctuation that counts as an observer (a brain that appears, observes equilibrium, then dissolves) dominates, making typical observers Boltzmann brains rather than post-Big-Bang observers like us.
“Most observers... you get who knows how many billions or trillions of observers like you and me after the Big Bang, but then you get an infinite number of observers that are random fluctuations in the future. So who cares about us living right after the Big Bang. Most observers in this situation are going to be random fluctuations”
- We should give essentially zero credence to Boltzmann-brain-dominated cosmologies not because we observe we are not Boltzmann brains (which is what a Boltzmann brain would say) but because the belief is cognitively unstable: if you use logic and reason to conclude you are probably a random fluctuation, you have no right to trust the very reasoning steps you used, since they too would be randomly fluctuated into your brain.
“if you use logic and reason to conclude that you are probably a Boltzmann fluctuation, then you must also think that you have no right to believe any of the steps you used along the way to do that reasoning”
- While a smooth low-entropy early universe would produce the cosmic microwave background we see, the reverse does not follow: conditioning only on current telescope data without assuming the past hypothesis, it is overwhelmingly more probable that the early universe was wildly inhomogeneous and we are a random fluctuation, because there are vastly more high-entropy pasts that could fluctuate into any medium-entropy present than low-entropy pasts that naturally lead to it.
“It's true that if the early universe was smooth, then we would see what we see in our telescopes. But it's not true that if we see what we see in our telescopes, then the early universe must have been smooth”
Infinite or very large universes necessarily contain infinite copies of you, creating multiverse problems independent of inflation or quantum mechanics.
- Guth's inflation posits that if the early universe's energy density was dominated by a false vacuum energy, it would expand at a hugely accelerated rate that smooths and flattens the universe (like pulling a wrinkly bed sheet), then converts that false vacuum energy into ordinary matter and radiation, explaining the observed smoothness, flatness, and absence of magnetic monopoles.
“Guth used this idea of inflation, which says that if a the universe starts out... its energy density is dominated by what we call a false vacuum energy... it would cause the universe to expand at a hugely accelerated, very fast rate. And that's kind of like pulling the edges of a wrinkly bed sheet”
- If the universe is spatially flat or negatively curved and not topologically twisted, it is infinite in size; since there are only a finite number of ways the ~10^80 particles in an observable-universe-sized patch can arrange themselves, an infinite universe guarantees an infinite number of regions containing exact copies of you, raising multiverse philosophical problems without invoking inflation, string theory, or quantum mechanics.
“So there will be an infinite number of people exactly like you and me, somewhere else very, very, very far away in this infinitely big universe. That right there raises these philosophical problems of how to deal with the multiverse”
- Because the inflaton is a quantum field, fluctuations can occasionally push it back up its potential, where inflation proceeds faster and generates huge volumes of space; for reasonable parameters inflation therefore never globally ends, producing eternal inflation in which inflation ends in some regions while continuing elsewhere.
“for very reasonable values of the parameters, inflation never really ends, it will end in some region of the space, but somewhere else, the inflaton field... it quantum fluctuated up the hill, and even though that's relatively rare when it happens, it generates a huge amount of space”
Solving cosmological problems requires both philosophical rigor to identify logical errors and physical insight to choose the right theory, neither discipline alone sufficing.
- Carroll chose the title 'professor of natural philosophy' at Johns Hopkins to indicate doing both philosophy and physics, because in his view there is no real boundary between thinking hard about the fundamental workings of reality in physics versus philosophy, even though academia has evolved a substantial perceived barrier.
“there's no boundary between these two areas, in both cases, you're thinking hard, trying to understand the fundamental workings of reality... so I don't perceive a barrier, but because of various ways which academia has evolved over the years, there is a quite a substantial barrier that other people perceive”
- Cosmology's biggest questions (the beginning of the universe, why there is something rather than nothing) have no straightforward experimental path, so unlike most physics where data eventually corrects sloppy reasoning, cosmology requires careful philosophical reasoning to avoid error.
“there is no straightforward path to addressing these questions just by doing the appropriate experiment. Nevertheless, they're important questions, and we should think about them, and therefore we should think carefully about them”
- Philosophers excel at analyzing the logical chain of reasoning and pinpointing where an argument fails but care less about reaching the right answer, often investing heavily in theories known to be incomplete; physicists excel at picking the right theory but reason sloppily, getting right answers for wrong reasons because data eventually corrects them, which fails in the multiverse where no immediate experiment guides you, so both disciplines are needed.
“Philosophers are really good at telling you why your theory is wrong... physicists just wanna get the right answer, and over and over again in the history of physics, they get the right answer for the wrong reasons... the data always save you in physics”
Self-locating uncertainty in multiverse contexts requires distinguishing credence about which theory is correct from credence about who you are within that theory's observer populations.
- Adopting fully non-indexical conditioning still recovers standard anthropic predictions: by interpreting the probability of the data as the probability that an observer like you exists (roughly proportional to the number of observers until it saturates), and temporarily pretending you do not already know the cosmological constant, you reproduce exactly what Weinberg did in 1988 using the number of galaxies as a proxy for observers, so physicists get the right answer for the wrong reasons.
“adopting this strategy... still gives the same anthropic predictions as the typical street cosmologist would get... I think they are getting the right answer but for the wrong reasons”
- Radford Neal's fully non-indexical conditioning holds that you should condition on everything you know about yourself (memories, age, observations) except your location in the universe, and assume you are typical only within the set of observers macroscopically identical to you; this eliminates the Doomsday and Jovian presumptuousness because you can no longer infer anything about future humans or aliens, while still allowing anthropic predictions.
“if you limit yourself to the set of all observers exactly identical to you macroscopically, then I do think you kinda gotta be typical in that sense, 'cause you have no criterion to distinguish between them”
- Reasoning in a multiverse requires two different kinds of credence: the standard scientific credence over which theory is correct (is there a cosmological multiverse, is many-worlds right), and a distinct credence about who we are within a scenario that contains many observers (self-locating or indexical uncertainty), and the latter is the genuinely novel philosophical problem.
“there's actually two different sets of credences we need to think about. One is the standard scientific question of which theory is correct... But there's also a different credence that comes in... who are we in the universe”
- Carroll tentatively endorses giving a bonus to cosmological scenarios with more observers, but only because larger universes have a higher probability of producing an observer like you, not because of an a priori preference for many observers; this bonus saturates once observers like you become nearly inevitable, distinguishing his view from the pure self-indication assumption which would boost indefinitely with observer count.
“it is okay to favour universes with lots of observers and therefore have a larger probability of predicting me... once the number of observers becomes so large that people like you are almost inevitable, I don't give an extra boost”
- Applying Bayes' rule to cosmological scenarios is stuck on the likelihood function because the only data is 'I exist'; it is unclear whether to ask the probability that any observers exist, observers like me, or exactly me, and whether one's own existence is genuinely new evidence or old evidence already baked into the priors (the problem of old evidence), making the update ambiguous.
“the data we're talking about is, I exist. That's the data... What is the probability that I would exist or you would exist... In fact, this is a well-known problem in Bayesian reasoning called the problem of old evidence”
Many-worlds quantum mechanics and decoherence provide branches that become independent worlds, parallel to how multiverse reasoning divides observers across cosmological regions.
- The Sleeping Beauty problem (waking once on heads, twice with memory wiped on tails, and asked her credence the coin came up heads) parallels multiverse reasoning: the thirder answer (1/3) corresponds to the observer-first approach (more copies of you in the tails world boosts that world), while Lewis's halfer answer (1/2) corresponds to the world-first approach where extra copies don't matter.
“The thirder position... If I know that in the tails universe, there will be two copies of me... and in the heads universe, there is only one copy of me if I'm giving equal probability to all three of those ideas. That's the observer first approach”
- In the many-worlds interpretation, measuring an electron's position means you become entangled with it so the wave function contains a branch where it is here and you saw it here and another where it is there and you saw it there; decoherence (developed in the 1970s-80s) explains why these branches become independent worlds that cannot affect each other, which is what licenses calling them separate worlds.
“the best reasons for talking about these different parts of the wave function as completely independent worlds come down to what we call decoherence... it explains why these different worlds become independent from each other. So that what happens in one world does not affect anything that happens in another world”
- Hartle and Srednicki proposed escaping the Boltzmann brain problem with a xerographic distribution that simply assigns zero probability to being a Boltzmann brain by fiat, declaring 'I am not one of them'; Carroll argues this is illegitimate cheating because there is no principled reason to exclude such observers, and a slightly more careful concept of the problem (Boltzmann yous, not Boltzmann brains) shows the distribution fails.
“they're putting this extra probability distribution that just says by Fiat, I am not one of them. How can they justify doing that? Well, they look around and they know they're not one of them. This might be considered cheating”
- Due to new teaching responsibilities at Johns Hopkins, Carroll plans to fold the monthly Ask Me Anything episode into a regular Monday slot, reducing the schedule from roughly five episodes per month to four.
“instead of roughly five podcast episodes per month will be getting four”
- Wald's cosmic no-hair theorem says a universe with a positive cosmological constant and not too much matter empties out toward de Sitter space; Carroll and Chatwin-Davies proved this is equivalent to entropy increase and equilibration (the Second Law), by defining an entropy that approaches a maximum and showing that condition yields de Sitter space even without Einstein's equations.
“I had long conjectured that this cosmic no-hair theorem was probably equivalent to equilibration. To entropy increase... To the Second Law of Thermodynamics. And finally, with a graduate student, Aidan Chatwin-Davies I was able to prove that”