Adam Elga on Rationality, Probability, and Reasoning in a Large Universe

Philosopher Adam Elga works through how a rational agent should apportion belief when the evidence runs out — among disagreeing experts, among indistinguishable copies of oneself, and across a universe large enough to contain many observers who feel exactly as you do.

Key ideas

1. Belief comes in degrees, and rationality governs how those degrees move. A credence is a degree of belief — not ‘yes’ or ‘no’ but a number between them, like a 30% confidence that it will rain. Bayesian updating is the rule that carries a prior credence, plus the likelihood of the data, into a posterior credence. Elga’s career-long project is to work out what this discipline requires in the hard cases where intuition gives out, and he holds that we must have some credence in every live possibility, because to act at all is to bet on one.

2. When a peer disagrees, defer to what your past self would have expected. Elga defends a version of the equal-weight view: on meeting an equally-informed person who reached the opposite conclusion, ask what your earlier self — before the dispute — would have said about who was likelier to be right. That test does not collapse into limp averaging, because most of us privately answer it in our own favour, which is itself the thing to notice. The failure mode to rule out is thinking ‘50/50’ in advance yet ‘I’m right’ when the disagreement actually lands.

3. Self-locating uncertainty is a genuine kind of ignorance about who and when you are. If two identical copies of you wake in identical rooms, no fact about the arrangement of every particle answers the question ‘which one am I?’. Elga argues that such questions are real and admit rational answers — typically equal credence across indistinguishable copies in the same world — and that this is not exotic: it is the ordinary puzzle of waking at night unsure whether it is 3am or 4am.

4. The Sleeping Beauty problem exposes whether extra observers deserve a probabilistic ‘boost’. Beauty is woken once if a fair coin lands heads and twice (memory wiped between) if tails; on waking, what credence should she give heads? Elga is a ‘thirder’ — 1/3 heads, 2/3 tails — because the two-waking world carries more instances of her exact state of mind. Whether such worlds deserve that boost is the hinge on which anthropic cosmology and the many-worlds of quantum mechanics also turn.

5. Boltzmann brains threaten to make rational belief eat itself. In an eternal, fluctuating universe, random chance produces vastly more short-lived ‘brains’ with your exact experience than genuine observers — so typicality reasoning says you are probably one. But a Boltzmann brain has no reason to trust its own memories or physics, so the conclusion undermines the reasoning that produced it. Elga maps the stable and unstable ways out and concludes the problem largely survives: we can exclude such cosmologies, but not cleanly or comfortably.

Content

Credence and Bayesian updating

Elga describes himself as ‘addicted to rationality’ — a childhood nickname, Mr Rational, that turned into a philosophical vocation centred on probability, justification and doing the optimal thing. The framework running underneath the whole conversation is Bayesian: you hold prior credences over rival hypotheses, you measure how likely the observed data would be under each, and you update to posterior credences accordingly. Carroll’s standing example is dark matter — a WIMP versus an axion — two live candidates a scientist rationally weighs unequally even before decisive evidence arrives.

Elga’s insistence, against a common dodge, is that a rational agent cannot simply decline to have a credence. Pressed, some people say they have no opinion, or that their state is better captured by an interval — ‘not 0.3 but somewhere from 0.2 to 0.7’. Elga has argued against treating suspension of judgement as a special third attitude: the test is always what such a stance tells you to do, and in practice we act on implicit credences about all manner of uncertainties whether we avow them or not. He and Carroll agree, though, that having credences does not mean spreading them evenly: complicated or ad hoc theories deserve lower priors, and scientific practice smuggles in fine-grained standards — a quantity suspiciously close to but not equal to zero counts against a theory — that no neutral, vanilla probability rule delivers.

The Sleeping Beauty problem and self-locating belief

Self-locating uncertainty is uncertainty not about how the world is but about where — or who, or when — you are within it. Elga introduces it through a teleporter that reconstructs you on two ships whose receiving bays look alike: no third-person description of the cosmos settles which copy you are, yet the question feels real. Scale it up — one Enterprise, a hundred Pumpkins — and the question becomes practical: stepping in, should you sweat as at even odds or as at 99-to-1? Elga’s inclination is equal credence across indistinguishable copies in one world, the same instinct as guessing 50/50 on waking unsure whether it is 3am or 4am.

The Sleeping Beauty problem sharpens this. Beauty sleeps on Sunday; a fair coin decides whether she is woken only on Monday (heads) or on both Monday and Tuesday (tails), with the first waking erased so every waking feels identical. On waking, credence in heads? Elga is a ‘thirder’, reasoning backwards from a variant where Beauty first learns it is Monday and, facing a yet-to-be-tossed fair coin, must surely say 50/50; holding the ratio between the heads-Monday and tails-Monday scenarios fixed as she loses that information forces 1/3 heads, 2/3 tails. He is candid that this is really the claim that a world with more awakenings — more instances of her exact state — gets a boost, and flags it as the crux for Boltzmann brains later. He also lays out the rivals: ‘halfers’, and the lesser-known compartmentalised-conditionalising view that firewalls probability inside each world so it cannot cross between them. Throughout, Elga stresses he has no taste for bizarre conclusions — he resists them, and would happily be shown a common-sense exit — but feels the strange verdicts are forced, not chosen.

Anthropic reasoning and Boltzmann brains

The philosophers’ puzzle maps directly onto physics. If more observers earn a boost, then a cosmology predicting many observers is confirmed simply because you exist in it — donning, in Elga’s phrase, ‘the robes of the presumptuous philosopher’ and settling a question about the whole universe from an armchair. He accepts the parallel to the anthropic multiverse and to Everett’s many worlds, and admits that consistency forces him to say yes, more-observer theories get more credence, ‘although I don’t like it’. He tidies the jargon: the boost-for-many-copies view is the self-indication assumption (SIA); the boost-for-high-frequency view is the self-sampling assumption (SSA); a third is compartmentalised conditionalising (CC). Carroll counters with Radford Neal’s fully non-indexical conditioning — judge a theory only by the chance it produces at least one observer exactly like you, which rewards big stochastic universes without runaway boosts as duplicates multiply.

Boltzmann brains are where confidence collapses. Given enough time, random fluctuation yields conscious observers — most of them fleeting, disembodied, formed for just long enough to have your present experience before decaying, ‘a sad short life’. If typicality reasoning holds, the sheer number of them should make you think you are probably one. Elga’s contribution is the self-undermining structure of that thought: a Boltzmann brain ‘never went to school’, has no reliable link to any past, so believing you are one destroys your grounds for the cosmology that said so — ‘the equivalent of finding out that the encyclopedia you have been basing all of your life upon was typed by monkeys’. He and Carroll agree the fluctuation-dominated cosmologies should be excluded, but Elga resists the tidy escape. Using analogies — a memory-doctor’s warning that you hallucinate memories, an X-ray machine that reports it is made of a fried egg — he shows the instability admits a stable response (‘revert to your prior; I don’t know anything’), so the problem is not dissolved by the instability alone. His own recent paper offers, by his own account, no way out, only a sharper map of the permissible wiggle room; the honest residue is that the puzzle ‘fully or almost fully survives’.

Dutch books and money pumps

Alongside self-locating belief, Elga names Dutch book arguments and money pumps — worked out with co-authors — as a strand of his rationality project, though the episode only gestures at them. Both are coherence arguments: a Dutch book is a set of bets that guarantees you a loss, and a money pump a cycle of trades that returns you to where you started poorer. The point of each is diagnostic — if your degrees of belief or your preferences are incoherent, a clever bookmaker can in principle turn you into a money pump, extracting a sure loss purely from the inconsistency. That threat is the pressure behind the whole conversation: it is why a rational agent’s credences must hang together as probabilities rather than float free.

Rational disagreement

Elga’s best-known applied work asks what to do when someone you already judged smart and well-informed reaches the opposite conclusion on the same evidence. One camp says the party who is in fact right should stick to their guns; the other, the equal-weight camp, says that since you had no prior reason to expect you would be the right one, learning of the disagreement is no evidence in your favour, so you should move substantially towards them. Following David Christensen, Elga favours a version that defers to your prior self: imagine phoning your past self and asking, without the full disputed details, how likely it would be that you rather than this person would be the one in error. Christensen’s split-the-check case — diners totting up a bill and getting different answers — makes the deferral intuitive, and it even accommodates asymmetry, as when you know you are bad at arithmetic and should weight the other person at 90/10.

The trap the exercise exposes is bootstrapping. If you privately favour yourself every time, then across a run of disagreements with no independent check you will tally up a string of apparent wins and conclude you are even better than you thought — ‘things have gone off the rails’. Elga’s target is the agent who says 50/50 in advance but ‘I’m right’ when the moment comes; the legitimate criticism of such a person is not that they hold a current opinion but that they will never admit they could be wrong (he cites his paper with Andy Egan, ‘I can’t believe I’m stupid’, on the limits of doubting your own views while still holding them). He also puts a harder option on the table, the level-splitting view: stay confident you are right about tomorrow’s weather while granting that the rational credence is 50/50 — a first-order belief paired with a second-order doubt about its own rationality, which he flags as a possible escape hatch for the Boltzmann case. The conversation closes on an AI twist: given how readily an AI could be reset to any state, perhaps it should be as cautious about trusting us as we are about trusting it.

See also

• Adam Elga — guest; philosopher of rationality, probability and decision at Princeton
• Sean Carroll — host
• Rationality — the discipline the whole conversation elaborates
• Many-Worlds Interpretation — the quantum multiverse Elga treats as parallel to the anthropic one
• Knightian Uncertainty — the ‘no probability, just an interval’ stance Elga argues against
• Annie Duke on Poker, Probabilities, and How We Make Decisions — a companion take on credences and decisions under uncertainty
• Iain McGilchrist on Quantum Mechanics, Teleology, and the Anthropic Principle — another Mindscape angle on anthropic reasoning
• Joel David Hamkins on Infinity, Gödel Incompleteness and the Mathematical Multiverse — infinity and multiplicity of worlds from the mathematician’s side