What this covers

Sean Carroll hosts physicist and philosopher David Deutsch in a wide-ranging conversation about how science produces knowledge and what makes an explanation good. Deutsch, a pioneer in quantum computation and architect of Constructor Theory, challenges the dominant Bayesian framework for understanding scientific progress. Against the view that science advances by adjusting credence numbers attached to propositions, Deutsch argues that scientists seek good explanations—theories that are constrained and hard to vary—and that explanatory knowledge constitutes a unique force in the universe. The exchange touches his foundational work on computability, quantum mechanics, and epistemology, anchored in Karl Popper's philosophy of science.

The conversation ranges across several fault lines. Deutsch lays out how human brains possess both hardware universality (Turing completeness) and explanatory universality, a second all-or-nothing property no other Earth organism shares. He invokes the Popper-Miller theorem to argue that evidence supporting a theory does not generally increase credence in its interesting consequences, thereby undermining Bayesian Epistemology from within its own mathematical framework. Good explanations cannot be formalized and do not obey the probability calculus or ordinary logic—a claim he supports by showing that the negation of an explanation explains nothing. The discussion includes his vision of Constructor Theory as a reformulation of physics in terms of which transformations are achievable and which forbidden, his advocacy for Everettian Quantum Theory, and his conviction that quantum mechanics and general relativity are both false. He also defends the Principle of Optimism—that anything not forbidden by physics is achievable with sufficient knowledge—as a framework for thinking about AGI and the future.

Sharpest takeaway

Deutsch argues that explanatory knowledge created by 'people' (humans, aliens, or future AGIs) constitutes a unique, universal, and unbounded force in the universe, and that science advances by seeking good, hard-to-vary explanations rather than by adjusting Bayesian credences.

  • Human brains have both hardware universality (Turing completeness) and explanatory (software) universality, which is all-or-nothing and the only kind there is
  • The Popper-Miller theorem shows that evidence supporting a theory does not generally increase the credence of its interesting consequences, undermining Bayesian epistemology
  • Good explanations cannot be formalized and do not obey the probability calculus or even ordinary logic, so knowledge growth is about explanatory power, not credence

The argument · threads9 threads · 37 claims
0.70

Explanation is non-formalizable and cannot obey probability calculus or classical logic.

6 pointscentrality 5/5
  • Even though we have no rigorous mathematical definition of explanatory universality, we have a rigorous mathematical understanding that we never will have one, because explanation is non-formalizable in the same way mathematics is.

    “is it then fair to say that even if we don't have a rigorous mathematical definition of explanatory universality, we have a rigorous mathematical understanding that we never will have a rigorous mathematical definition?”

  • Good explanations do not even obey ordinary logic: the negation of an explanation is never itself an explanation. Saying gravity is not caused by spacetime curvature explains nothing about quantum gravity, which can be proven because we have no working theory of quantum gravity—so if even logic fails to model explanation, the probability calculus certainly cannot.

    “the negation of an explanation is never an explanation. So if you say that gravity is caused by the curvature of spacetime, that's a theory... To say gravity is not caused by the curvature of spacetime doesn't explain anything”

  • Deutsch rejects Bayesian epistemology specifically—the theory that knowledge consists of propositions each accompanied by a credence number obeying the probability calculus, and that improving a theory means increasing credence in true theories and decreasing it in false ones. He argues these credences either do not exist or, if they do, cannot obey the probability calculus.

    “I specifically object to Bayesian epistemology, which is the theory of knowledge, that knowledge consists of propositions in a rational mind, each of which is accompanied by a number”

  • The concept of an explanation cannot be formalized because one can always invent new modes of explanation as conjectures; this non-formalizability is exactly parallel to the non-formalizability of mathematics, where no formalism can capture all mathematical truths or what counts as a valid proof.

    “is it then fair to say that even if we don't have a rigorous mathematical definition of explanatory universality, we have a rigorous mathematical understanding that we never will have a rigorous mathematical definition?”

  • The Popper-Miller theorem shows that evidence can increase the credence of a general theory while decreasing the credences of its consequences; the only consequences whose credence is increased substantially are those that merely restate or nearly restate the evidence, so interesting consequences are not supported—and this result is independent of the prior credence distribution.

    “The key is that a piece of evidence can increase the credence of a general theory while decreasing the credences of its consequences”

  • There is an infinite amount of mathematics to be discovered, though some of it can never be discovered; Deutsch conjectures that all the interesting mathematics, which is also infinite, can be discovered.

    “we know that there's an infinite amount of mathematics to be discovered, even though there... there's a lot of it that we can't discover... Although I have a conjecture that we can discover all the interesting things, which are also infinitely”

0.65

Constructor theory reformulates all physics as which transformations are possible versus impossible.

4 pointscentrality 5/5
  • The basic idea of constructor theory is that all laws of physics can be expressed in terms of which physical transformations can be brought about and which cannot, analogous to computable versus non-computable functions; it functions as an overarching meta-law (like conservation of energy) that constrains and explains subsidiary theories rather than replacing dynamical laws.

    “The basic idea of constructor theory is that all the laws of physics can be expressed in such terms, in terms of a distinction between what can be brought about and what can't be brought about”

  • Popper taught that the content of a scientific theory lies in what it rules out; taking this seriously and asking what a theory rules out versus what it permits leads directly to constructor theory, a step Popper himself never took.

    “Popper taught us that the content of a theory, of a scientific theory, is in what it rules out. And if you just took that seriously as the basis of your worldview, you'd immediately come to constructor theory”

  • Deutsch conjectures, by analogy with the principle of optimism, that there is a principle stating the physical transformations that can be brought about are precisely those a universal constructor could bring about—paralleling Turing's universal computer for the domain of physical construction.

    “is there a principle that says the transformations that can be brought about are precisely the ones that a universal constructor could bring about?”

  • A hoped-for advantage of constructor theory is that it crosses levels, letting biology, chemistry, and physics be discussed under one umbrella; thermodynamics is a prime case, where one wants principles (like the impossibility of converting all heat to work) that transcend the specific physics of the materials involved.

    “in thermodynamics we really don't want to know what the specific physics of the stuff we are dealing with... We want principles that transcend that”

0.63

Quantum theory and general relativity contradict each other and are both definitely false.

3 pointscentrality 5/5
  • Quantum theory is false (not merely approximate) because it conflicts with general relativity: quantum mechanics says the sun is in a superposition of many positions, some far from where we see it (amplified by chaotic interactions), yet neither theory tells us how the sun's gravitational effect differs across these branches—so we lack the right equations, and both theories are at best excellent approximations.

    “quantum theory tells us, with equal confidence, that the sun isn't just in one place, the sun is in a superposition or more generally in a mixed state, where it has many different positions simultaneously”

  • Deutsch holds high credence in both quantum theory and general relativity even though he believes both are definitely false and that they contradict each other—proving his beliefs about theories cannot obey the probability calculus, since otherwise credence in one would equal one minus credence in the other.

    “I think that quantum theory is definitely false. I think that general relativity is definitely false. And I also think they contradict each other. And therefore my credences... they definitely don't obey the probability calculus, because if they did, my credence for one would be one minus my credence for the other. And yet I have a very high credence for both of them”

  • Deutsch proposes a variant quantum field theory that drops the standard axiom that space-like separated fields commute (which forces fields to be discontinuous everywhere and lack a conceptual model), instead requiring fields at space-like separated points to be allowed to fail to commute while being continuous; this changes the meaning of causality, measurement, and subsystem separation, and yields only finitely many possible second-order equations of motion.

    “I have an idea for a variant of quantum field theory where we don't have that axiom, where fields at space-like separated points are allowed to fail to commute and where the thing that they have to do is be continuous”

0.49

Scientists seek best explanations by choosing most constrained viable theories, not highest probability.

3 pointscentrality 4/5
  • When seeking general theories, scientists do not seek high probability—they seek the most constrained, lowest-probability theory that is still viable, because it is the best explanation. Galileo preferred circular orbits over ellipses, and ellipses over epicycles, even though epicycles were the most probable, because the more constrained theory leaves less unexplained.

    “The theory that the planets move on epicycles has got a far higher probability than that they move on ellipses, because an ellipse is a kind of epicycle. So by the Linda argument, Galileo should have preferred the epicycle theory”

  • We cannot use classical decision theory to decide which experiments to fund by maximizing expected value, because the best outcome of an experiment is one you did not foresee—and if you didn't foresee it, you couldn't have assigned it a probability; instead we judge by explanatory prospects for increasing knowledge.

    “The best thing that can happen in an experiment is that you get an outcome that you didn't foresee. But if you didn't foresee it, you also didn't foresee its probability”

  • The principle of optimism holds that everything not forbidden by the laws of physics is achievable with enough knowledge; for if something were permitted by physics yet unattainable by any computable program we could run, then it actually would not be possible, and we could formulate and test a refutable scientific theory saying the supposed laws of physics were not the true laws.

    “the principle of optimism is that everything which is not forbidden by laws of physics is possible with enough knowledge”

0.49

Explanatory universality, possessed by humans and future AGIs, makes them all classifiable as people.

3 pointscentrality 4/5
  • Extraterrestrial civilizations and future artificial general intelligences will necessarily possess the same explanatory knowledge property that humans have, since they could not build technology without it; therefore all such entities should be classed as 'people,' and there is nothing beyond this category in terms of capability.

    “Humans are people, extraterrestrials are people, AGIs will be people. And I argue in my book that there's nothing beyond that”

  • We have no clue how to build an AGI; what stands between us and it is an explanatory, largely philosophical theory of what creativity and explanation are, not more compute or data. The eventual program will likely be relatively simple—humans differ from great apes by only a few kilobytes of code that bootstraps a qualitatively different kind of program.

    “I think what's standing between us and making an AGI is an explanatory theory. It'll be largely a philosophical theory rather than a computer science theory”

  • Beyond Turing completeness (a hardware property), human software has explanatory universality—a second kind of universality that no other surviving Earth organism has; like Turing universality it is all-or-nothing with no gradations, and extinct human relatives like Neanderthals and Homo erectus that had campfires also possessed it.

    “The other kind of universality, explanatory universality, is a property of software, which I say we have, our software has that property, and no other surviving organism on Earth has explanatory universality”

0.46

Bayesian reasoning applies only to finite frequencies in narrow cases, not to general theory choice.

2 pointscentrality 4/5
  • Bayes' theorem is legitimately useful for particular (not general) questions—like a doctor estimating dengue fever from travel history—but this is really reasoning about finite frequencies approximated as probabilities, used because of an absence of explanation rather than a genuine credence; such methods cannot decide between general theories because the relevant set of individuals would be infinite.

    “He's looking at frequencies, first of all, not probabilities... He's approximating those frequencies of probabilities, and he's using the approximation that his putative patient is randomly chosen”

  • Support is intransitive: evidence that Linda is a banker supports 'banker and feminist,' which logically implies 'feminist,' yet being a banker is itself no support (even counter-support) for being a feminist. So increasing credence in a theory does not generally increase credence in its consequences—only rarely, in uninteresting cases.

    “her being a banker is no kind of support for her being a feminist. So it, there's a non-transitivity there”

0.44

Major transitions in the universe occurred rapidly after long stasis without current explanation.

12 pointscentrality 2/5
  • Chance encounters (like the lunch with DeWitt) did not truly steer Deutsch toward the Everett interpretation but merely hastened a conviction he would have reached anyway, illustrating a kind of convergent evolution in intellectual development.

    “I wasn't exactly steered, I was just hastened. I think I would have come round to this eventually”

  • Because of explanatory knowledge on Earth, the planet will soon be in a position to repel incoming asteroids and comets rather than attract them, making it potentially the only planet in the universe to do so—a purely physical fact requiring no anthropocentric framing.

    “the planet Earth will soon be in a position where if asteroids or comets head towards it, they will be repelled, rather than attracted. For all we know, every other planet in the universe attracts asteroids and comets, and ours will be the only one to repel”

  • Hugh Everett left research physics not because he was disappointed by the reception of his ideas but because he wanted to do other things and make a fortune, which he did.

    “He left it because he wanted to do other things and to make a fortune. And you know, he did.”

  • The sudden transitions after long monotony are not examples of punctuated equilibrium because in none of these cases was there a genuine equilibrium—all the states were unstable to the change happening—so Gould's punctuated equilibrium is at best a description of what sometimes happens, not a theory of evolution.

    “this punctuated equilibrium, as advocated by Gould, for example, in my view is not a theory of evolution, it's just at best, it's a description of what sometimes happens”

  • Major transitions in the universe—origin of planets, origin of life, multicellular life, explanatory creativity in humans, and the Enlightenment—all occurred rapidly after very long periods of nothing happening, and in no case can we currently explain why they took as long as they did.

    “all those things happened very rapidly after a long period of not happening, and in all cases, you can't really put your finger on why it took so long”

  • Popper and Miller's theorem was written in only three or four pages, accepted by the British Journal for the Philosophy of Science and Nature; some people got very angry while most didn't notice—Deutsch thinks everybody should notice and nobody should get angry.

    “It didn't, it only took three or four pages. They sent it off to, I think the British Journal for the Philosophy of Science and also to Nature. The papers were accepted. Some people got very angry, and most people didn't notice.”

  • Assigning a probability per unit time to events like the evolution of multicellular life does not explain why they took as long as they did; saying the probability was one in a million years does not explain why it actually took a billion years—something substantive is needed instead.

    “if we were told that the probability of multicellular life evolving per unit time was actually 1 in a million years, we'd be wondering why it's a billion, but saying it's a million doesn't help to explain why it was a billion”

  • The defining vice of a bad explanation is that it can explain anything—like solipsism or 'the patient could be lying'—and such explanations must be excluded because they short-circuit any approach to truth; this criterion has a transcendent compellingness without being an arbitrary axiom, since rejecting it lands one in self-defeating positions like denying that mountains exist.

    “this argument that an explanation that can explain anything is a bad explanation, I think has got a transcendent compulsiveness about it, a compellingness about it. Which doesn't involve any axioms”

  • Human brains have hardware universality—they are Turing complete, meaning they can execute any computable program, with finite memory and lifetime being only trivial restrictions that can be augmented across generations.

    “our brains are Turing complete. That is, we can execute any program that can be executed at all. Now, it might take us more than a lifetime. It might require more memory than we have, but we can augment our memory”

  • We should resist the tendency to assume we know almost everything; in fact we know almost nothing, and during the pandemic many things treated as known were not known and still are not—intellectual humility about the limits of current knowledge is warranted.

    “I don't like this jumping into thinking we know almost everything now. We know almost nothing. During the pandemic, I was tweeting all the time: This isn't known.”

  • The Schrödinger picture is misleading because its state is global, leading Everett and DeWitt to think the whole universe splits whenever the state changes; in the Heisenberg picture only the observables split locally, leaving the distant universe unchanged by local quantum phenomena, which makes the many-worlds view easier to accept.

    “the Schrödinger picture is very misleading, because the Schrödinger state is global and it leads Everett and DeWitt to thinking about the whole universe as splitting every time the state changes”

  • Newton may have known much of what we now call modern Lagrangian or Hamiltonian dynamics but deliberately left it out of his work as irrelevant to his goals, preferring three laws of motion to five; he definitely knew about Galilean invariance.

    “I think Newton did know a lot about what we now call modern theory of dynamics or Lagrangian or Hamiltonian dynamics, but he decided not to include it because it was irrelevant to what he wanted to show”

0.33

Quantum computation involves temporary splitting into vast separate universes that later recombine.

2 pointscentrality 3/5
  • Turing conjectured that his model of computation captured everything computable; given quantum theory, Deutsch was able to prove this conjecture, so if quantum theory is true there is only one kind of universal computation and nothing beyond it—meaning aliens and AGIs could not compute non-Turing-computable functions.

    “once we went to quantum computation, I was able to prove that given quantum theory. So if quantum theory is false, it might still be false, but if quantum theory is true, then Turing's conjecture is now proven”

  • There are two situations where treating quantum systems as split into separate worlds is a good approximation: when decoherence occurs (with no hope of recombining), and during a quantum computation where vast numbers of universes each run autonomous classical computations before recombining; in the intermediate splitting/recombining process it is not useful to treat them as separate universes since they affect each other too much.

    “there are two situations where it is a good approximation to think of quantum physics, of quantum systems as splitting into worlds. One of them is when there's decoherence, but the other one is where there's a quantum computation in progress”

0.33

Understanding local physical systems requires understanding essentially everything in the universe.

2 pointscentrality 3/5
  • To predict the behavior of a champagne bottle kept on ice by a research team awaiting a breakthrough, one must understand essentially everything in the universe—extraterrestrial life, quasars, dark matter, dark energy—because explanatory knowledge connects local human artifacts to the entire cosmos, unlike a genome which only encodes a parochial slice of its environment.

    “you need to understand basically everything before you can understand how champagne bottles behave on the surface of the Earth. And that again is because of the peculiar properties of explanatory knowledge”

  • The human being is a hybrid: the brain is both the controller and creative programmer of a universal constructor (the body), where the constructor itself must be perfectly obedient rather than creative; bodily errors are correctable in principle with sufficient knowledge, so constructor theory ultimately reduces to knowledge.

    “a brain is both the controller of a universal constructor, which is the human body... But it's also the programmer. It's also the entity that creatively invents the programs, which a universal constructor is not allowed to be creative. It has to be perfectly obedient. So obedient is the opposite of creative”