The fourth book of rationality

This is part 4 of 6 in my series of summaries. See this post for an introduction.



Part IV

Mere Reality



M
oving on from evolutionary and cognitive models, this part explores the nature of mind and the character of physical law. What kind of world do we live in, and what is our place in that world? We will look at past scientific mysteries, parsimony, and the role of science in individual rationality.

Previous sections have explained patterns in human reasoning and behavior through the lenses of mathematics, physics and biology, but haven’t said much about humanity’s place in nature or the natural world in its own right. Humans are not only goal-oriented systems, but also physical systems. We are built out of inhuman parts, like atoms. We can relate the human world to the world revealed by physics through reductionism: in particular, by applying it to present-day controversies in science and philosophy, like the debates on consciousness and quantum physics.

The philosopher Thomas Nagel famously asked whether anyone can ever know what it’s like to be a bat – what it would subjectively feel like. Even if we could perfectly model bat neurology and predict bat behavior, how could we be certain that the bat isn’t just an unconscious automaton? David Chalmers argues that third-person cognitive models can never fully capture first-person consciousness, and that consciousness is a “further fact” not fully explainable by the physical facts. But can this argument stand up to a technical understanding of how explanation and belief work?

The other topic, quantum mechanics, is our best mathematical model of the universe to date. The Schrödinger equation deterministically captures everything there is to know about the dynamics of physical systems, including “superpositions”… as long as we aren’t looking. Whenever we make observations, the superpositions seem to vanish without a trace, and we need to use Born’s probabilistic rule to make predictions. What all of this even means has produced many views on the nature of quantum mechanics. Yudkowsky uses this scientific controversy as a proving ground.



15

Lawful Truth

This chapter introduces the basic links between physics and human cognition.

In the late 18th century, Antoine-Laurent de Lavoisier discovered that people, like fire, consume fuel and oxygen and produce heat and carbon dioxide. Today, phosphorous (part of ATP, crucial for metabolism) is found in matches. We may use different surface-level rules for different phenomena, but the underlying laws that govern nature are not so divided. Due to this, you can’t change just one thing in the world and expect the rest to continue working as before. For example, if matches didn’t light, we couldn’t breathe. Reality is laced together a lot more tightly than we might like to believe.

Newton’s unification of falling apples with the course of planets gave rise to the idea of universal laws with no exceptions. Even though our models of fundamental laws may not last, reality itself is (and was) always constant and rigid. To think the universe itself is whimsical is to mix up the map and the territory. In our everyday lives we are accustomed to rules with exceptions, but apparent violations of the basic laws of the universe exist only in our models, not in reality. Remember: since the beginning, not one unusual thing has ever happened.

Just as physics came before the physicist, mathematics came before the mathematician in a structured universe. The beauty of math is discovering new properties that you never built into the mathematical objects you created (like building a toaster and realizing that your invention also, for some unexplained reason, acts as a rocket jetpack and MP3 player). Don’t prematurely end the search for mathematical beauty by trying to impose order; sometimes you have to dig a little deeper. Finding hidden beauty isn’t certain, but it has happened frequently enough throughout history to be better than unfounded faith. Consider the sequence {1, 4, 9, 16, 25, …}. Can you predict the next item in the sequence? You could take the first differences to get {4-1, 9-4, 16-9, …} = {3, 5, 7, 9,…} and then second differences to get {5-3, 7-5, 9-7, …} = {2, 2, 2, …}.

If you predict the next second difference is also 2, then the next first difference must be 11, hence the next item in the original sequence must be 36. But perhaps the “messy real world” lacks the order of these abstract mathematical objects? It may seem messy for three reasons: first, we may not actually know the rules due to empirical uncertainty; second, even if we do know all of the math, we may not have enough computing power to do the full calculation due to logical uncertainty; and third, even if we could compute it, we still don’t know where in the mathematical universe we are living due to indexical uncertainty. We are not omniscient. However, uncertainty exists in the map, not in the territory. Our best guess is that the “real” world is perfectly regular math. In many cases we’ve already found the underlying simple and stable level – which we name “physics”.

Bayesian probability theory is attractive because it comprises unique and coherent laws, as opposed to the ad-hoc tools of frequentist statisticians. Bayesians expect probability theory and rationality to be self-consistent, neat, and even beautiful, which is why they think Cox’s theorems are so important. Rationality is fundamentally math, and using a useful approximation to a law (in the map) doesn’t change the law (in the territory). Any approximation to the optimal answer will be explainable in terms of Bayesian probability theory, and just because you may not know the explanation does not mean no explanation exists. Instead of thinking in terms of tricks to throw at particular problems, we should think in terms of the extent to which we approximate the optimal Bayesian theorems.

There is a proverb that “outside the laboratory, scientists are no wiser than anyone else”. If there’s any truth to this, we should be alarmed. Rationality should apply in everyday life, not just in the laboratory. We should be disturbed when scientists believe wacky ideas (e.g. religion), because it means they probably don’t truly understand why the scientific rules work on a gut level, but blindly practice scientific rituals like a social convention. Those who understand the map-territory distinction and see that reality is a single unified process will integrate their knowledge.

The Second Law of Thermodynamics says that the phase space volume of a closed system is conserved over time. The link between heat and probability makes this Bayesian. Any rational mind does “work” in the thermodynamic sense, not just the sense of mental effort. Like a car engine or refrigerator, your brain (an engine of cognition) must interact thermodynamically with the physical world to form accurate beliefs about something. In other words, you really do have to observe it. True knowledge of the unseen would violate the laws of physics.

People tend to think that teachers tell them things that are certain and that this is like an authoritative order that must be obeyed, whereas a probabilistic belief is like a mere suggestion. Probabilities are not logical certainties, but the governing laws of probability are harder than steel. It is still mandatory to expect a smashed egg not to spontaneously reform. It is still mandatory to expect a glass of boiling-hot water to burn your hand rather than cool it, even if you don’t know that with certainty. So beware believing without evidence and saying “no one can prove me wrong!” Otherwise you’ll end up building a vast edifice of justification and confuse yourself just enough to conceal the magical step, similar to people designing “perpetual motion machines”.

Philosophers have spilled ink over the nature of words and various cognitive phenomena. But it was Bayes all along! Mutual information between a mind and its environment is physical negentropy, which is Bayesian evidence, implying that any useful cognitive process must to some extent be in harmony with Bayes-structure. For a mind that arrives at true beliefs or better-than-random beliefs, there must be at least one process with a sort-of Bayesian structure somewhere, or it couldn’t possibly work. The quest for the hidden Bayes can be exciting, because Bayes-structure can be buried under all kinds of disguises.


16

Reductionism 101

This chapter deals with the project of scientifically explaining phenomena.

A philosophical argument that free will does or does not exist is different from the question, “what cognitive algorithm, as felt from the inside, drives the intuitions about the debate?” Dissolve the question by explaining how the confusion arises. Good philosophy should not respond to a question like “if a tree falls in a forest but no one hears it, does it make a sound?” by picking and defending a position (“Yes!” or “No!”), but by deconstructing the human algorithm to the point where there is no feeling of a question left. What goes on inside the head of a human who thinks they have free will?

Confusion exists in the map, not in the territory. Questions that seem unanswerable, where you cannot imagine what an answer would look like, mark places where your mind runs skew to reality. Where the mind cuts against reality’s grain, it generates questions like “do we have free will?” or “why does anything exist at all?” Bad things happen when people try to answer them: they inevitably generate a Mysterious Answer. These wrong questions must be dissolved by understanding the cognitive algorithms causing the feeling of a question.

When facing a wrong question, don’t ask “why is X the case?” but instead ask “why do I think X is the case?” and trace back the causal history of your belief. You should be able to explain the steps in terms of smaller, non-confusing elements. For example, rather than asking “why do I have free will?” try asking “why do I think I have free will?” This latter question is guaranteed to have a real answer whether or not there is any such thing as free will, because you can explain it in terms of psychological events.

Science fiction artists seem to think that sexiness is an inherent property of a woman, such that an alien invader would find her attractive despite having a different mind and different evolutionary history. Ancient magazine covers depicted “bug-eyed monsters” carrying off girls in torn dresses.

This is a case of the Mind Projection Fallacy, coined by E.T. Jaynes. It is a general error to project your own mind’s properties into the external world. Other examples include Kant’s declaration that space by its very nature is flat, and Hume’s definition of a priori ideas as those “discoverable by the mere operation of thought, without dependence on what is anywhere existent in the universe”.

Probabilities express states of partial information; they are not inherent properties of things. It is only agents who can be uncertain. Ignorance is in the mind, and a blank map does not correspond to a blank territory. And via Bayes’s Theorem, learning different items of evidence can lead you to different states of partial knowledge (unsurprisingly). There is no “real probability” that a flipped coin will come up heads; the outcome-weighting you assign to it depends on the information that you have about the coin. The coin itself has no mind and doesn’t assign a probability to anything.

The morning star and evening star are both Venus, but the quotation “the morning star” is not substitutable for “the evening star”. You have to distinguish beliefs/concepts from things, and remember that truth compares beliefs to reality, but reality is real regardless. (Reality itself does not need to be compared to any beliefs in order to be real.) If you don’t make a clear enough distinction between your beliefs about the world, and the world itself, it is very easy to derive wrong conclusions. As Alfred Tarski said: “snow is white” is true if and only if snow is white.

Confusing belief with reality is easier when using qualitative reasoning, which leads to mistakes like thinking “different societies have different truths” (no, they have different beliefs). Instead of a qualitative binary belief or disbelief, you should use quantitative probability distributions and degrees of accuracy (measured in log base 2 bits). For example, if you assign a 70% probability to the sentence “snow is white” being true, and if snow is white, then your probability assignment is more accurate than it would have been if you had assigned a 60% chance; in fact, it will score log2(0.7) = -0.51 bits. For meta-beliefs (i.e. beliefs about what you believe), you may assign credence close to 1, since you may be less uncertain about your uncertainty than you are about the territory. This way you can avoid mixing up beliefs, accuracy, and reality – which are all different things.

Reality itself is not “weird” or “surprising”. If your intuitions are shocked by the facts, then those are poor intuitions (i.e. models) and they should be updated or discarded. People think that quantum physics is weird, yet they have the bizarre idea that reality ought to consist of little billiard balls bopping around, when in fact reality is a perfectly normal cloud of complex amplitude in configuration space. If you find this “weird”, that’s your problem, not reality’s problem, and you’re the weird one who needs to change. Reality has been around since long before you showed up. There are no surprising facts, only models that are surprised by facts.

Yudkowsky gives the example of how he used to browse random websites when he couldn’t work, and he thought that he couldn’t predict when he would become able to work again. But if you see your hour-by-hour work cycle as chaotic or unpredictable, your productivity may not actually be unpredictable; you may just be committing the Mind Projection Fallacy – i.e. the problem being your own stupidity with respect to predicting it. Inverted stupidity looks like chaos, and we often fail to think of ourselves; we just see a chaotic feature of the environment. Hence we miss opportunities to improve.

Reductionism is disbelief in the mind projection fallacy that the higher levels of simplified multilevel models exist in the territory. Only the fundamental laws of physics are “real”, but for convenience we use different representations at different levels or scales of reality. For example, we use different computer models for the aerodynamics of a 747 and collisions in the Relativistic Heavy Ion Collider (RHIC), but they obey the same fundamental Special Relativity, quantum mechanics and chromodynamics. We build models of the universe that have many different levels of description, but so far as anyone has been able to determine, the universe itself has only the single level of fundamental physics (reality doesn’t explicitly compute protons, only quarks). The scale of a map is not a fact about the territory; it’s a fact about the map.

Poets like John Keats have lamented that “the mere touch of cold philosophy” has destroyed haunts in the air, gnomes in the mine, and rainbows. But one of these things is not like the others. There is a difference between explaining something (e.g. rainbows) and explaining it away (e.g. gnomes and haunts). The former, the rainbow, is still there. The latter has been shown to be an error in the map, since there never were gnomes in the mine! When you don’t distinguish between the multi-level map and the mono-level territory, then when someone tries to explain to you that the rainbow is not a fundamental thing in physics, acceptance of this will feel like erasing rainbows from your multi-level map, which feels like erasing rainbows from the world. But when physicists say “there are no fundamental rainbows”, it does not mean “there are no rainbows”.

It is fake reductionism to profess that something has been explained by Science, without seeing how it is reducible. Imagine a dour-faced philosopher, who isn’t able to see where the rainbow comes from, telling you that “there’s nothing special about the rainbow, scientists have explained it away, just something to do with raindrops or whatever, nothing to be excited about.” The anti-reductionists experience “reduction” in terms of being told that the password is “Science”, with the effect of moving rainbows to a different literary genre (one they’ve been taught to regard as boring). Genuine knowledge and understanding lets you do things like play around with prisms and make your own rainbows with water sprays. Scientific reductionism does not have to lead to existential emptiness.

Poets write about Jupiter the god, but not Jupiter the spinning sphere of ammonia and methane. Equations of physics aren’t about strong raw emotions. Classic great stories (like those told about Jupiter the god) touch our emotions, but why should Jupiter be human when we are humans? It’s not necessary for Jupiter to think and feel in order for us to tell stories, because we can always write stories with humans as their protagonists. We don’t have to keep telling stories about Jupiter. That being said, we could do with more diverse poetry and original stories. The Great Stories are old!


17

Joy in the Merely Real

This chapter touches on the emotional, personal significance of the scientific world-view.

Keats wrote in his poem Lamia that rainbows go in “the dull catalogue of common things”. Anything real is, in principle, scientifically explicable. Nothing that actually exists is inherently mysterious. But being unmagical and knowable doesn’t make something not worth caring about! We have to take joy in the “merely” real and mundane, or else our lives will always be empty. Don’t worry if quantum physics turns out to be normal; if you can’t take joy in things that turn out to be explicable, you’re going to set yourself up for eternal disappointment – a life of irresolvable existential ennui.

Solving a mystery can feel euphoric – especially if you discover the answer to a problem that nobody else has answered. But if you personally don’t know, why should it matter if someone else has the answer? If you don’t understand a puzzle, there’s a mystery, and you should still take joy in discovering the solution. We shouldn’t base our joy on the fact that nobody else has done it before. Stop worrying about what other people know, and think of how many things you don’t know! Rationalists shouldn’t have less fun.

Why talk about joy in the merely real when discussing reductionism? One reason is to leave a line of retreat; another is to improve your own abilities as a rationalist by learning to accomplish things in the real world rather than in a fantasy. Getting close to the truth requires binding yourself to reality. Don’t invest your emotional energy in magic (or lotteries), but redirect it into the universe. This should not create existential anguish, because that which the truth nourishes should thrive. Emotions that cannot be destroyed by the truth are not irrational. Understanding the rainbow does not subtract its beauty; it adds the beauty of physics.

If people don’t have a scientific attitude in this universe, why would they become powerful sorcerers in a fantasy world if magic were “merely real”? Magic, like UFOs, gets much of its charm from the fact that it doesn’t actually exist. If dragons were real, they wouldn’t be exciting, because people don’t take joy in the merely real. If we ever create dragons or find aliens, people would treat them like zebras – most people wouldn’t bother to pay attention, while some scientists would get oddly excited about them. If you’re going to achieve greatness anywhere, you may as well do it in reality.

Rationalists should bind themselves emotionally to a lawful reductionistic universe, and direct their hopes and care into merely real possibilities. So why not make a fun list of abilities that would be amazingly cool if they were magic, or if only a few chosen people had them? Imagine if, instead of the ordinary one eye, you possessed a magical second eye which enabled you to see into the third dimension and use legendary distance-weapons – an ability we’d call Mystic Eyes of Depth Perception. Speech could be “vibratory telepathy”. Writing could be “psychometric tracery”. Etc. We shouldn’t think less of them for commonality.

Science doesn’t have to be recent or controversial to be interesting, beautiful, or worth learning. Indeed, cutting-edge news is often wrong or misleading, because it’s based on the thinnest of evidence (and often conveys fake explanations). By the time anything is solid science, it is no longer a “newsworthy” headline. Scientific controversies are topics of such incredible difficulty that even people in the field aren’t sure what’s true. So it is better to read well-written elementary textbooks rather than press releases. Textbooks will offer you careful explanations, examples, test problems, and likely true information. Study the settled science before trying to understand the outer fringes.

Breaking news in science is often controversial and hard to understand (and have a high chance of not replicating), so why don’t newspapers report more often on understandable explanations of old science? Sites like Reddit and Digg do this sometimes already. Perhaps journalists should make April 1st a new holiday called “Amazing Breakthrough Day”, in which journalists report on great scientific discoveries of the past as if they had just happened and were still shocking (under the protective cover of April Fool’s Day). For example, “BOATS EXPLAINED: Centuries-Old Problem Solved By Bathtub Nudist” (Archimedes).

Trying to replace religion with atheism, humanism or transhumanism doesn’t work, because it would just be trying to imitate something that we don’t really need to imitate. Atheistic hymns that try to imitate religion usually suck. But in a world in which religion never existed, people would still seek the feeling of transcendence; and this isn’t something we should always avoid. A sense of awe is not exclusive to religion. Unlike theism, space travel is a lawful dream; and humanism is not a substitute for religion because it directs one’s emotional energies into the real universe. It’s not just a choice of drugs: humanity actually exists.

17.9 Scarcity
We value the same objects more if we believe they are in short supply. Scarcity makes the unobtainable more desirable, and “forbidden” information seems more important or trustworthy. This is shown by experimental evidence (see Robert Cialdini’s ‘Influence: The Psychology of Persuasion’). Psychologically, we seek to preserve our options (and leaping on disappearing options may have been adaptive in a hunter-gatherer society). However, desiring the unattainable is likely to cause frustration. And as soon as you actually get it, it stops being unattainable. Tim Ferriss recommends that, instead of asking yourself which possessions or status-changes would make you happy, ask which ongoing experiences would make you happy.

Watching the birth of a child or a space shuttle launch can inspire feelings of “sacredness”, but religion corrupts this; religion makes the experience mysterious, faith-based, private, and separated from the merely real – but it doesn’t have to be! Religion twists the experience of sacredness to shield itself from criticism. Some folks try to salvage “spirituality” from religion. But the many bad habits of thought that have developed to defend religious and spiritual experience aren’t worth saving. Let’s just admit we were entirely wrong, and enjoy the universe that’s actually here.

Too few people study science, because they think it’s freely accessible, so it doesn’t fit their need for deep secret esoteric hidden truth. (In fact, you have to study a lot before you actually understand science, so it’s not public knowledge.) Because it is perceived that way, people ignore science in favor of cults that conceal their secrets, even if those secrets are wrong. So if we want to spread science, perhaps we should hide it in vaults guarded by mystic gurus wearing robes, and require fearsome initiation rituals.

This parody sketch about “the Bayesian Conspiracy” depicts how Brennan, a character in the Beisutsukai series, is inducted into the Conspiracy. After climbing sixteen times sixteen steps, Brennan passes through a glass gate into a room lined with figures robed and hooded in light-absorbing cloth. They chant the names of Jakob Bernoulli, Abraham de Moivre, Pierre-Simon Laplace, and Edwin Thompson Jaynes, who are dead but not forgotten. Brennan is asked to perform a Bayesian calculation, and is given a ring when he finally gives the correct answer.


18

Physicalism 201

This chapter is on the hard problem of consciousness.

As we’ve seen, the reductionist thesis is that we use multi-level models for computational reasons, but physical reality has only a single level. You can see how your hand (the higher level) reduces to your fingers and palm (the lower level). These are the same, but from different points of view. While it is conceptually possible to separate them, that doesn’t make it logically or physically possible. It is silly to think that your fingers could be in one place and your hands somewhere else. And just because something can be reduced to smaller parts doesn’t mean the original thing doesn’t exist.

How can you reduce anger to atoms, when atoms themselves are emotionless? Instead of professing passwords, try to understand how causal chains in your mind compute the consequences of options, and how this echoes the environment. But this is challenging, and the ideas of neurons, information processing, computing etc. give us the benefit of hindsight. Without them it is hard to understand how little bouncing billiard balls could combine in such a way as to make something angry. But yes, even something like anger can be reduced to atoms.

For a very long time, people had a detailed understanding of kinetics (with concepts like momentum and elastic rebounds) and an understanding of heat (with concepts like temperature and pressure). In hindsight it is obvious that heat is motion, but in 1824 it was conceivable that the two were separate. It took an extraordinary amount of work to understand things deeply enough to make us realize that heat and motion were really the same thing. To cross the gap, you’d have to conceive it possible for heat and motion to be the same, and then see how the former reduces, which is hard.

As an example of “Amazing Breakthrough Day”, one could write an article talking about how the brain has recently been discovered by a multinational team of scientists to be made from a complicated network of cells called neurons, which use electrochemical activities to perform thought. This discovery indicates that mind and body are of one substance, contrary to Descartes. The brain, as the seat of reason, could according to Darwin be the product of a history of non-intelligent processes, and therefore mental entities are probably not ontologically fundamental.

In the hunter-gatherer era, animism (i.e. the belief that things like rocks and streams had spirits or minds) wasn’t obviously stupid; if the idea were obviously stupid, no one would have believed it. But now we know, thanks to microscopes, neuroscience, cognition as computation, and Darwinian natural selection, that trees and rivers don’t think or have intentions. Anthropomorphism only became obviously wrong when we realized that the tangled neurons inside the brain were performing complex information processing, and that this complexity arose as a result of evolution.

18.6 A priori
Your brain is an engine that works by processing entangled evidence; this includes thoughts themselves, because thoughts are existent in the universe (in the form of neural patterns, i.e. the operation of brains). The facts that philosophers call “a priori” arrived in your brain by a physical process. You might even observe them within some outside brain. There is no “a priori truth factory” that works without a reason. The reason why simple algorithms are more efficient, and Occam’s Razor works, is because our simple low-entropy universe has short explanations to be found.

The map is multi-level but reality is single-level, and truth involves a comparison between belief and reality. The higher levels of your map have referents in the single level of physics. Concepts note only that a cluster exists, and do not define it exactly. We can’t practically specify everything in terms of quarks, so our beliefs are “promissory notes” that refer to (and implicitly exist in) empirical clusters in what we think is the single, fundamental level of physics. Implicit existence is not the same as nonexistence. Virtually every belief you have is not about elementary particle fields, but this doesn’t mean that those beliefs aren’t true. For example, “snow is white” does not mention quarks anywhere, and yet snow nevertheless is white – it may be a computational shortcut, but it’s still true. This view of how we can form accurate beliefs way above the underlying reality is not circular, but self-consistent upon reflection.

Your philosophical zombie is putatively a being that is exactly identical to you, except that it’s not conscious. Some argue that if zombies are “possible” then we can deduce a priori that consciousness is extra-physical. Epiphenomenal property dualists (who think zombies are possible) like David Chalmers believe in consciousness, but also that is has no real-world effects. (They are not the same as substance dualists like Descartes who believe that mind-substance is causally active.) But you can be aware of your awareness and write about it, which requires physical causality. Whatever makes you say “I think therefore I am” causes your lips to move, and is thus within the chains of cause-and-effect that produce our observed universe. Philosophers writing papers about consciousness would seem to be at least one effect of consciousness upon the world. Zombie-ists would need an entirely separate reason within physics of why people would talk about subjective sensations. So epiphenomenalism is pointlessly complicated! It postulates a mysterious, separate, extra-physical, inherently mental property of consciousness, and then further postulates that it doesn’t do anything. That is deranged. Don’t try to put your consciousness or your personal identity outside physics.

A few more points on p-zombies: Your internal narrative can cause your lips to say things, and consciousness is probably that which causes you to be aware of your awareness. The reductionist position is that that-which-we-name “consciousness” happens within physics, even if we don’t fully understand it yet. It should be logically impossible to eliminate consciousness without moving any atoms. It seems deranged to say that something other than consciousness causes my internal narrative to say “I think therefore I am” but that consciousness does exist epiphenomenally.

The argument against zombies can be extended into a more general principle, albeit with some difficulty. The Generalized Anti-Zombie Principle (GAZP) says that something that can’t change your internal narrative (which is caused by the true referent of “consciousness”) can’t stop you being conscious. Any force smaller than thermal noise won’t “switch off” your consciousness, because it doesn’t significantly affect the true cause of your talking about being conscious. This implies that you could, in principle, transfer your consciousness to artificial silicon neurons.

A Giant Lookup Table (GLUT) could in principle replace a human brain and thus seem conscious, but a randomly generated GLUT is extremely unlikely to have the same input-output relations as the human brain. The GAZP says that the source of a GLUT that mirrors the brain is probably a conscious designer. When you follow back discourse about “consciousness”, you generally find consciousness, because it is responsible for the improbability of the conversation you’re having. This kind of GLUT would have to be precomputed by a human using a computational specification of a human brain.

That it is impossible to observe something is not enough to conclude that it doesn’t exist. If a spaceship goes over the cosmological horizon relative to us, so that it can no longer communicate with us, should we believe that the spaceship instantly ceases to exist? We cannot interact with a photon outside our light cone, but it continues to exist as a logical implication of the general laws of physics, which themselves are testable. Believing this doesn’t violate Occam’s Razor, because Solomonoff induction applies to laws or models, not individual quarks. The photon is an implied invisible, not an additional invisible. Under the Minimum Message Length formalism of Occam’s Razor (which is nearly equivalent to Solomonoff Induction), if you have to tell someone how your model of the universe works, you just have to write down some equations to simulate your model – not specify individually the location of each quark in each star in each galaxy – and the amount of time it takes to write down the equation doesn’t depend on the amount of “stuff” that obeys it.

This is a satirical script for a short scene in a zombie movie –but not about the folkloric lurching and drooling kind of zombie; the philosophical kind. Imagine a world where normal people are infected with an “epiphenomenal virus”, which cannot be experimentally detected. Things would go on just as before. Or would they? The victims sure look conscious… but don’t let the fact that they look like humans, talk like humans, claim to have qualia, and are identical to humans on the atomic level, fool you. These clever scumbags are zombies!

Is science only about “natural” things? People say this when invoking the supernatural, which appeals to ontologically basic mental entities like souls. But firstly, this is a case of non-reductionism, which is a confusion because it is incoherent. It is not clear what an irreducibly mental and fundamentally complicated universe is even supposed to look like. Secondly, if you test supernatural explanations the way you would test any other hypothesis (by converting them into a reducible and natural formulation), you will probably still find out that they aren’t true.

There is a prediction of supernatural models: that information can be transferred “telepathically” between brains in the absence of any known material connection between them. If psychic powers were discovered they would be strong Bayesian evidence that non-reductionism is correct and that beliefs are ontologically fundamental entities. But more likely, this will not be discovered, and the reason we are tempted by non-reductionist worldviews is that we just lack self-knowledge of our own brains’ quirky internal architecture. If naturalism is correct, then the attempt to count “belief” or the “relation between belief and reality” as a single basic entity is simply misguided anthropomorphism.


19

Quantum Physics and Many Worlds

This chapter is on the measurement problem in physics. Yudkowsky discusses many-worlds interpretations as a response to the Copenhagen interpretations. Since he is not a physicist, you are free to consult outside sources to vet his arguments or learn more about the physics examples. Note that the Many-Worlds Interpretation is still controversial.

Quantum mechanics doesn’t deserve its fearsome reputation. Quantum mechanics is considered by many to be weird, confusing or difficult; yet it is perfectly normal reality. There are no surprising facts, only models that are surprised by facts: the issue lies with your intuitions, not with quantum mechanics. A major source of confusion is that people are told that quantum physics is supposed to be mysterious, and they are presented with historical erroneous concepts like “particles” or “waves” rather than a realist perspective on quantum equations from the start.

What is the stuff reality is made of? The universe isn’t made of little billiard balls, nor waves in a pool of aether, but mathematical entities called configurations that describe the position of particles, and amplitude flows between these configurations (measured as square moduli of complex numbers). A configuration can store a single value in the form of a complex number (a + bi) where i is defined as √(-1). These complex numbers are what we call amplitudes, and they are out there in the territory. We cannot measure amplitudes directly, only the ratio of absolute squares of some configurations. To find the amplitude of a configuration, you sum up all the amplitude flows into that configuration.

The figures above depict the classic split-photon experiment with half-silvered mirrors. On the left, a photon is sent from the source toward the half-silvered mirror A; this is a configuration. We can give the configuration “a photon heading toward A” a value of (-1 + 0i). From there, the photon can take alternative pathways – B and C are full mirrors and D is another half-mirror. The half-mirrors multiply by 1 when the photon goes straight and multiply by i when the photon turns at a right angle. The full-mirrors always multiply by i. Note that all these amplitude flows happen, and that they can cancel each other out, which is why no photon is detected at E. In the diagram on the right, the configuration of “a photon going from B to D” has an amplitude value of zero (because it’s blocked), so Detector 1 now goes off half the time.

While the previous experiment deals with one moving particle, real configurations are about multiple particles (in fact, all the particles in the universe). Configurations specify where particles are (e.g. “a photon here, a photon there…”), but they don’t keep track of individual particles!

In the figure above, two photons head toward D at the same time. But whether both photons are deflected or both go straight through, the resulting configuration is the same. Amplitude flows that put the same types of particle in the same places flow into the same configuration, even if the particles came from different places. Unlike probabilities, amplitudes can have opposite signs (positive or negative) and thus cancel each other out (giving a squared modulus of zero for the sum), which is how we can detect which configurations are distinct. So it is an experimentally testable fact that “photon B here, photon C there” is the same configuration as “photon C here, photon B there”. This is why we may see Detector 1 go off twice or Detector 2 go off twice, but not both Detectors go off at the same time.

A configuration is defined by all particles. What makes a configuration distinct is at least one particle in a different state – including particles constituting the experimental equipment! Adding a sensor that tries to “measure” things introduces a new element into the system and thus makes it a distinct configuration. If amplitude flows alter a particle’s state, then they cannot flow into the same configuration as amplitude flows which do not alter it.

In the above experiment, sensitive thingy S is in a different state between “a photon from D to E and S in state no” and “a photon from D to E and S in state yes”. By measuring the amplitude flows, we have stopped them from flowing to the same configurations. In this case the amplitudes don’t cancel out, leading to a different experimental result. We find that the photon has an equal chance of striking Detector 1 and Detector 2. (Remember, in the first diagram in “Configurations and Amplitude”, the photon always struck Detector 2.) This phenomenon confused the living daylights out of early quantum experimenters. But the distinctness of configurations is a physical fact, not a fact about our knowledge, and there’s no need to suppose that the universe cares what we think.

Macroscopic decoherence, also known as “many-worlds”, is the view that the known quantum laws that govern microscopic events simply govern at all levels without alteration. Collapse postulates assume that wavefunction superposition “collapses” at some point before reaching the macroscopic level (some say due to conscious awareness!), leaving only one configuration with non-zero amplitude and discarding other amplitude flows. Many Worlds proposes that configurations where we observe and don’t observe a measurement both exist with non-zero amplitude, but are decoherent since they are too different from each other for their amplitude flows to flow into common configurations. But early physicists used to assume that measurements had single outcomes. They simply didn’t think of the possibility of more than one world, even though it’s the straightforward result of applying the quantum laws at all levels. So they invented an unnecessary part of quantum theory which says that parts of the wavefunction spontaneously and mysteriously disappear when decoherence prevents us from seeing them anymore. Yet collapse theories are still supported by physicists today.

The idea that decoherence fails the test of Occam’s Razor is wrong as probability theory. Occam’s Razor penalizes theories for explicit entities that cannot be summed over; but the Many-Worlds interpretation of quantum mechanics does not violate Occam’s Razor because decoherent worlds follow from the compact laws of quantum mechanics. Measurements obey the same quantum-mechanical rules as all other physical processes. Some physicists just use probability theory in a way that is outright mathematically wrong, on the level of 2+2=3. This is one of the reasons why Yudkowsky, as a non-physicist, dares to talk about physics.

To probability theorists, words like “simple”, “testable” and “falsifiable” have exact mathematical meanings; e.g. we can use Bayes’s Theorem to concentrate the probability mass of a hypothesis into narrow outcomes (falsifiability) and look for evidence that would favor one hypothesis over another (testability). Macroscopic decoherence is falsifiable for the same reasons quantum mechanics is, and compared to the collapse postulate, it is strictly simpler – because decoherence is a deductive consequence of the wavefunction’s evolution. Within the internal logic of decoherence, the many superposed worlds are simply a logical consequence of the general laws that govern the wavefunction, and as such, do not cost us extra probability. Adding collapse is a useless complication.

Suppose a murder case in a big city leaves no evidence, yet one of the police detectives says, “Well, we have no idea who did it, but let’s consider the hypothesis that it was Mortimer Q. Snodgrass.” This can be called the fallacy of privileging the hypothesis. Before promoting a specific hypothesis to your attention, you need to have some rational evidence already at hand, because it takes more evidence to narrow down the space of all possibilities, than to figure out which of the handful of candidate hypotheses is true. The anti-epistemology is to talk endlessly about how you “can’t disprove” an idea, how the negative evidence is “not conclusive”, how future evidence could confirm it but hasn’t happened yet, and so on. Single-world quantum mechanics (i.e. collapse postulates) has no evidence in favor of it, and there are a billion other possibilities that are no more complicated, so it’s not worth even thinking about it. But due to historical accident, collapse postulates are indeed spoken about.

Some people may be disturbed by the straightforward prediction of quantum mechanics that they are constantly splitting into zillions of other people. Egan’s Law says: “It all adds up to normality”. The many worlds of quantum mechanics have always been there, and they are where you have always lived – not some strange, alien universe into which you have been thrust. But you cannot causally affect other worlds, and decoherence has nothing to do with the act of making decisions. Living in multiple worlds is the same as living in one. Worrying about an extremely pleasant or awful world in your future is like the lottery. So live in your own world. Quantum physics is not for building strange philosophies around many-worlds, but for answering the question of what adds up to normality. If there were something else there instead of quantum mechanics, then the world would look strange and unusual.

Before Hugh Everett III proposed his relative state formulation (aka many-worlds) in 1957, none of the theories were very good and the best quantum physicists could do was to “shut up and calculate”. But that is not the same as claiming that “Shut up!” actually is a theory of physics, and that the equations definitely don’t mean anything. Nevertheless, some jumped to the conclusion that the wavefunction is only a probability. This contributed to quantum non-realism, which is a semantic stopsign. If you can’t say exactly what you mean by calling the quantum-mechanical equations “not real”, then you’re just telling others to stop asking questions. The equations do describe something – the quantum world is really out there in the territory and the classical world exists only implicitly within the quantum one (at least from the realist perspective).

If wavefunction collapse actually happened, it would be the only informally specified (qualitative), non-linear, non-unitary, non-differentiable, discontinuous, non-local in the configuration space, acausal (non-deterministic), and superluminal (faster than light) law in quantum mechanics, and the only fundamental phenomenon to violate CPT symmetry, Special Relativity and Liouville’s Theorem, and be inherently mental. It would be the only fundamental law adopted without precise evidence to nail it down. If early physicists like Niels Bohr had never made the mistake, and thought immediately to apply the quantum laws at all levels to produce macroscopic decoherence, then “collapse postulates” would today seem like a completely crackpot theory. There would be many better hypotheses proposed to explain the mysterious Born probabilities.

Early quantum physicists made the error of forgetting that they themselves were made of atoms, so they concluded that conscious observation had a fundamental effect. They didn’t notice that a quantum theory of distinct configurations already explained the experimental result, without any need to invoke consciousness. In retrospect, could philosophers have told the physicists that this was a big mistake? Philosophical insight would not have helped them, because it’s usually science that settles a confusion. That’s why we don’t usually see philosophers sponsoring major advances in physics. At the frontier of science, it takes intimate involvement with the scientific domain in order to do the effective philosophical thinking.

Some people think that free will and determinism are incompatible. If the laws of physics control everything we do, then how can our choices be meaningful? But “you” and physics are not competing causal nodes; you are within physics! Your desires, plans, decisions and actions cannot determine the future unless we live in a lawful, orderly universe.

Things should not seem like the causal network on the left, but the one on the right. If the future were not determined by reality, it could not be determined by you. Yudkowsky calls this view “Requiredism”: that planning, agency, choice etc. require some lawful determinism. Anything you control is necessarily controlled by physics.

If collapse theories (or any theory of a globally single world) were true, they would violate Special Relativity; and although Special Relativity seems counterintuitive to us humans, what it really says is that human intuitions about space and time are simply wrong. Given the current state of evidence, the “many-worlds interpretation” (i.e. macroscopic decoherence) wins outright. There is no reason to suppose that quantum laws are different on the macroscopic level, so it seems obvious that there are other decoherent Earths. You shouldn’t even ask, “Might there only be one world?” The argument should have been over fifty years ago. New physical evidence could reopen it, but we have no particular reason to expect this. The main problem for Many Worlds is to explain the Born probabilities.


20

Science and Rationality

This chapter relates the ideas of previous ones to scientific practice. So if it was many-worlds all along, and collapse theories are silly, did physicists in the first half of the 20th century really screw up that badly? How did they go wrong, and what lessons can we learn from this whole debacle?

This is another short story set in the same universe as “The Ritual” and “Initiation Ceremony”. Future physics students look back on the cautionary tale of quantum physics. Einstein, Schrödinger, and Von Neumann failed to see Many Worlds, perhaps because of administrative burdens imposed by a system of science that thought it acceptable to take 30+ years to solve a problem, rather than resolving a major confusion faster. Eld Science was based on getting to the truth eventually… but people can think important thoughts in far less than thirty years if they expect speed of themselves.

The failure of physics in the first half of the 20th century was not due to straying from the scientific method, because the physicists who refuse to adopt many-worlds are obeying the rules of Science. Science and rationality (i.e. Bayesianism) aren’t the same thing. The explanation of quantum mechanics was meant to illustrate (among other things) the difference between the scientific method, which demands new testable predictions, and Bayesian probability theory, which suggests that macroscopic decoherence is simpler than collapse. Science says that many-worlds doesn’t make new testable predictions because we can’t see all the other worlds; Bayes says that the simplest quantum equations that cover all known evidence don’t have a special exception for human-sized masses. This is a clear example of when it comes time to break your allegiance to Science.

The social process of Science doesn’t trust the rationality of individual scientists, which is why we give them the motive to make falsifiable experimental predictions. But the rational answer comes from Bayes’s Theorem and Solomonoff Induction. Science doesn’t always agree with the exact, Bayesian, rational answer, and Science wants you to go out and gather overwhelming experimental evidence, because it doesn’t trust you to be rational. It assumes that you’re too stupid and self-deceiving to just use Solomonoff induction.

Science doesn’t care if you waste ten years on testing a stupid theory, as long as you recant and admit your error. But some things (e.g. cryonics) cannot be experimentally tested right now despite huge future consequences; so you have to think rationally to figure out the answer. You should not automatically dismiss such theories. You have to try to do the thing that Science doesn’t trust you to do, and figure out the right answer before you get clubbed over the head with it. Evolutionary psychology is another example of a case where rationality has to take over from science.

Your private epistemic standard should not be as lax as the ideal of Science, which asks only that you do the experiment and accept the results. Science lets you believe any stupid idea that hasn’t been refuted by experiment, and it lets people test whatever hypotheses they like, as a social freedom. Science accepts slow, generational progress. But you need Bayes (and Tversky & Kahneman) to tell you which hypotheses to test and precisely how much probability to assign to them. Bayesianism says that there is always an exactly rational degree of belief given your current evidence, and this does not shift a nanometer depending on your whims.

It seems that scientists are not trained in precise rational reasoning on sparse evidence (e.g. the formal definition of Occam’s Razor, the conjunction fallacy, or the concepts of “mysterious answers” and “fake explanations”). These are not standard. Hence why modern world-class scientists, like Sir Roger Penrose, still make mistakes like thinking that consciousness is caused by quantum gravity. They were not warned to be stricter with themselves. Maybe one day it will be part of standard scientific training, but for now it’s not, and the absence is visible.

You can’t think and trust at the same time. To grow as a rationalist, you must lose your emotional trust in the sanity of normal folks. You must lose your trust that following any prescribed pattern will keep you safe. Not even Science or Eliezer can save you from making mistakes. There is no known procedure you can follow that makes your reasoning defensible. Since the social rules of Science are verbal rather than quantitative, it is possible to believe you are following them; but with Bayesianism, it is never possible to do an exact calculation and get the exact rational answer that you know exists. You are visibly less than perfect. So learn to live with uncertainty while still having something to protect and striving to do better.

Yudkowsky is less of a lonely iconoclast than he seems, as many of these ideas are surprisingly conventional and are being floated around by other thinkers (e.g. Max Tegmark and Colin Howson). Perhaps Popper’s falsificationism should be replaced with Bayesianism. But this is not as simple as redefining science. It would require not just teaching probability theory, but also things like cognitive biases and social psychology. We need to form a new coherent Art of Bayescraft before we are actually going to do any better in the real world than modern science.

Whether scientists accept an idea depends not just on epistemic justification, but also a social pack and extra evidence to overcome cognitive noise – hence why Science is inefficient at reaching conclusions. And Science doesn’t say which ideas to test. Yet the bulk of work in progressing knowledge is in elevating the right hypothesis to attention. Hence Bayesianism is faster than science. Ironically, science would be stuck if there weren’t some people who could get it right in the absence of overwhelming experimental proof, because in many answer spaces it’s not possible to find the true hypothesis by accident.

Albert Einstein is an example of an individual scientist who, in the presence of only a small amount of experimental evidence, was unusually good at arriving at the truth faster than the social process of Science (and even more unusually, he admitted it). Deciding which ideas to test can entail high-minded thoughts like noticing a regularity in the data, or inferring a new law from characteristics of known laws (as Einstein did with relativity). Einstein used the data he already had more efficiently – in his armchair. It’s possible to arrive at the right theory this way, but it’s a lot harder.

Einstein used evidence much more efficiently than other physicists, but he was still extremely inefficient in an absolute sense. Imagine a world where the average IQ is 140 and a huge team of physicists and cryptographers was examining an interstellar transmission; going over it bit by bit for thirty years, they could deduce principles on the order of Galilean gravity just from seeing two or three frames of a picture. But a Bayesian superintelligence would make much more efficient use of sensory data, such that it could invent Newtonian mechanics the instant it sees an apple fall.

On the scale of intelligence, the distance between Einstein and “village idiot” is tiny compared to the distance between Einstein and a Bayesian superintelligence. In other words, it looks not like this:

But more like this:

Yudkowsky was disappointed when his childhood hero, Douglas Hofstadter, disagreed. Perhaps this is a “cultural gap” explained by Yudkowsky reading a lot of science fiction from a young age. This was helpful for thinking beyond the human world. That is why he looked up to the ideal of a Bayesian superintelligence, not Einstein. The ideal role model you should aim for should come from within your dreams, because by letting your ideals be composed only of dead humans, you limit yourself to what has already been accomplished, and you will ask too little of yourself.

People talk as if Einstein had magical superpowers or an aura of destiny. There’s an unfortunate tendency to talk as if Einstein, even before he was famous, had a rare inherent disposition to be Einstein. But Einstein chose an important problem, had a new angle of attack, and persisted for years. An intelligent person under the right circumstances can do better than Einstein. (Not everyone, but many have potential.) The way you acquire superpowers is not by being born with them, but by seeing with a sudden shock that they are perfectly normal. What Einstein did isn’t magic; just look at how he actually did it!

In another story from the Beisutsukai series, Brennan and the other students are faced with their midterm exams, and are given one month to develop a theory of quantum gravity. Einstein was too slow to formulate General Relativity (taking ten years), and his era lacked knowledge of cognitive biases and Bayesian methods. It should be possible to do better, if you expect it. An open challenge in science is quantum gravity, which is the question of how to unify General Relativity and quantum mechanics.




Interlude: A Technical Explanation of Technical Explanation

YOU’VE SEEN THE Intuitive Explanation of Bayesian reasoning, but when do the mathematical theorems apply and how do we use the theorems in real-world problems? Is there a controversy? We begin by asking: What is the difference between a technical understanding and a verbal understanding?

One visual metaphor for “probability density” or “probability mass” is a lump of clay that you must distribute over possible outcomes. This lets you visualize how probability is a conserved resource – to assign higher probability to one hypothesis requires stealing some clay from another hypothesis. This matters when you have to bet money, because by not being careful with your bets, you will not maximize your expected payoff.

Imagine there is a little light that flashes red, green, or blue each time you press a button. You have to predict the color of the next flash and you can bet up to one dollar. If the game uses a proper scoring rule, then if the actual frequencies of the lights are 30% blue, 20% green and 50% red, you maximize your average payoff by betting 30 cents on blue, 20 cents on green and 50 cents on red. If you press the button twice in a row, you get the same score regardless of whether you are scored once for your prediction P(green1 and blue2) or twice for P(green1) and P(blue2|green1).

In Bayesian terms this is known as an invariance. Note that P(green1) x P(blue2|green1) = P(green1 and blue2). So it doesn’t matter whether we consider it two predictions or one prediction; we get the same result. Mathematically, the scoring rule would be: Score(P) = log(P); meaning that your score is the logarithm of the probability you assigned to the winner. And your expected score would be:

This is different from the colloquial way of talking about degrees of belief.

When people say “I am 98%” certain…” what they usually mean is “I’m almost but not entirely certain”, which reflects the strength of their emotion rather than the expected payoff of betting 98% of their money on that outcome. But technically, your confidence would be poorly calibrated if you said that you were “98% sure” but you get more than two questions wrong out of a hundred independent questions of equal difficulty. The Bayesian scoring rule rewards accurate calibration. But calibration is just one component; the other component is discrimination. This refers to discriminating between right and wrong answers: the more probability you assign to the right answer, the higher your score. You can be perfectly calibrated by saying “50% probability” for all yes/no questions, but this is merely confessing ignorance, and you can do better.

Now imagine an experiment which produces an integer result between zero and 99. You could predict “a 90% probability of seeing a number in the fifties”, which is a vague theory compared to “a 90% probability of seeing 51”. The precise theory has an advantage because it concentrates its probability mass into a sharper point. So if we actually see the result 51, this is evidence in favor of the precise theory. If the prior odds were 10:1 in favor of the vague theory, seeing a 51 would make the odds go to 1:1 and seeing a 51 again would then bring it to 1:10 in favor of the precise theory.

However, the vague theory would still score better than a hypothesis of zero knowledge (or maximum-entropy, which makes any result equally probable). Even worse than the ignorant theory would be a stupid theory which predicts “a 90% probability of seeing a result between zero and nine”, and thus assigns 0.1% to the actual outcome, 51. By making confident predictions of a wrong answer, it is thus possible to have a model so bad that it is worse than nothing. Ignorance is better than anti-knowledge.

Under the laws of probability theory, it is not possible for both A and not-A to be evidence in favor of B, so a true Bayesian can only test ideas of which they are genuinely uncertain; they cannot try to prove a preferred outcome or to prevent disproof. Unfortunately, human beings are not Bayesians. People don’t distribute conserved probability mass over advance predictions, but try to argue that whatever event did happen “fits” the hypothesis they had in mind beforehand. The consequence of this is that people miss their chance to realize that their models did not predict the phenomenon.

So when a class’s physics students observe a square plate of metal next to a hot radiator, and feel that the side next to the radiator is cool and the distant side is warm, the students may guess that “heat conduction” is responsible. This is a vague and verbal prediction. If they had measured the heat of the plate at different points at different times and applied equations of diffusion and equilibrium, they would soon see a pattern in the numbers, and their sharp predictions might lead them to guess that the teacher turned the plate around before they entered the room.

You now have a technical explanation of the difference between a verbal explanation and a technical explanation, because you can calculate exactly how technical an explanation is. Bayesian probability theory gives you a referent for what it means to “explain” something. Other technical subjects (like physics, computer science, or evolutionary biology) permit this too – which is why it is so important that people study them in school. And as long as you can apply the math, and distinguish truth from falsehood, you are allowed to have fun and be silly.

A useful model is knowledge you can compute in reasonable time to predict real-world events you know how to observe. This is why physicists use different models to predict airplanes and collisions in a particle accelerator, even though the two events take place in the same universe with the same laws of physics. A Boeing 747 obeys Conservation of Momentum in real life, even if some aerodynamic models, which are cheap approximations, violate Conservation of Momentum a little. As long as the underlying fundamental physics supports the aerodynamic model, it can be a good approximation. Even a “vague” theory can be better than nothing, and given enough experiments, vague predictions can build up a huge advantage over alternate hypotheses. Such theories, if they produce not-precisely-detailed but still correct predictions, may be called “semitechnical” theories.

But aren’t precise, quantitative theories still better than vague, semitechnical theories? Well, in the nineteenth century, Darwinian evolutionism was a semitechnical theory (they did not yet have quantitative models) but physics was already precise and mathematical. The physicists said that the Sun could not have been burning for that long, so people like Lord Kelvin challenged natural selection. Of course, evolution turned out to be correct. Nineteenth century physics was a technical discipline, but it was incomplete – they didn’t know about nuclear reactions. The lesson is that every correct theory about reality must be compatible with every other correct theory, and if there seems to be a conflict between two well-confirmed theories, then you are applying one of the two theories incorrectly or applying it outside the domain it predicts well.

The social process of Science requires you to make advance predictions. This custom exists to prevent human beings from making human mistakes. But the math of probability theory does not distinguish between advance predictions and post facto ones, which is why nineteenth century evolutionism still worked. Today, evolutionary theory can make quantitative predictions about DNA and genetics, and as a technical theory it is far better than a semitechnical theory. But controversial theories at the cutting-edge of science are often semitechnical, or even nonsense.

The discipline of rationality is very important for distinguishing a good semitechnical theory (truth) from nonsense (falsehood). But scientific controversies should matter only if you are an expert in the field, or if it affects your life right now. For the rest of us, elementary textbook science shows the comprehensible beauty of settled science. If you do have a reason for following a scientific controversy, then you should pay attention to the warning signs that historically distinguished vague hypotheses that turned out to be gibberish, from those that later became confirmed theories.

A sign of a poor hypothesis is that it expends great effort in avoiding falsification. In terms of Bayesian likelihood ratios, falsification is stronger than confirmation. As Popper emphasized, the virtue of a scientific theory lies not in the outcomes it permits, but the outcomes it prohibits. The same Bayesian scoring rule we saw earlier can be used to accumulate scores across experiments, which we should strive to maximize. The only mortal sin of Bayesianity is to assign probability 1 or zero to an outcome, because this is like accepting a bet with a payoff of negative infinity.

By making predictions in advance, it is easier to notice when someone is using too much probability mass to try to claim every possible outcome as an advance prediction. Imagine waking up one morning to find that your left arm has been replaced by a blue tentacle. How would you explain this? Well, you wouldn’t because it’s not going to happen. There are verbal explanations (like divine intervention, or hallucination) that could “fit” the scenario, but such explanations can “fit” anything. If aliens did it, why would they do that particular thing to you as opposed to the other billion things they might do? What will the aliens do tomorrow? If you guess a model with no internal detail or a model that makes no further predictions, why would you even care? People play games with plausibility, explaining events they expect to never actually encounter.

If you had a “good explanation” for the hypothetical experience, then you would go to sleep worrying that your arm really would transform into a tentacle. Under Bayesian probability theory, probabilities are anticipations: if you assign probability mass to waking up with a blue tentacle, then you are nervous about waking up with a blue tentacle. To explain is to anticipate, and vice versa. If you don’t anticipate waking up with a tentacle, you need not bother crafting excuses you won’t use. Remember: since the beginning, not one unusual thing has ever happened.

Next: The fifth (and penultimate) book of rationality

Comments

  1. Loved it, compelling and throughlly enjoyable

    ReplyDelete

Post a Comment

Popular posts from this blog

Don't accept a Nobel Prize -- and other tips for improving your rationality