2. On Bayesian Probability

← Index

Veils Cast Aside; Examining Her Breasts

(Bayes in Theory)

What Rationalists emphasize perhaps above all as an axiom is the concept of “Bayesian reasoning” as a formula for thought. They print Bayes’ formula on t-shirts, they call themselves “Bayesians”, they describe a “conspiracy of Bayes”. If there is a single theme to Yudkowsky’s writing, beyond the threat of an unaligned superintelligence, it is the wonders a person can possibly achieve if he has a deeply felt understanding of how to apply Bayes’ theorem to his day-to-day life.

What does this mean, however? How can some theorem of probability be so important? Bayes’ formula is: the likelihood of Y being true given X occurring is equivalent to the likelihood of X occurring given Y being true, multiplied by the likelihood of Y being true and divided by the likelihood of X occurring.

The implications of this are likely not obvious to the average reader, hence why Yudkowsky over the years has taken a few shots at writing Bayes explainers for the general audience which require a few hours to digest yet are meant to make the implications of Bayes’ formula intuitive. We certainly invite you to read Yudkowsky’s writings on LessWrong and Arbital if you have the interest in understanding Bayes in depth, otherwise we will do our best to go forward and make the importance of this idea understood without such a primer.

We can make the explanation more simple for our purposes, and we will avoid perplexing the reader with mathematical formulas. Bayes’ formula is a straightforward derivation from the fundamental axiom of conditional probability. As such, it should be thought of simply as a way we can rearrange the basic axioms to find the likelihood of Y being true given X occurring, if we know the likelihood of X, Y, and X occurring given Y being true.

What is crucial to understand here to illustrate the theorem’s profundity — something which many explanations gloss over — is that X and Y are not of the same ontological register. X is an event which may or may not occur, and Y is a truth about the world.

For a long time, the application of Bayes theorem was described as a field called “inverse probability”. Inverse probability does not predict, but instead sees an event and uses it to discern truth: this is its radical nature.

The basic question of standard probability is: how likely is X to happen? We are able to answer this easily in some toy setup if, for example, we have some distribution of balls bouncing around in predictable ways which can cause X under certain circumstances. You can picture if you will as the most basic physical model of a probabilistic system: a hand-cranked lottery machine, which includes an opaque chamber within which wooden balls bounce as the operator turns a crank, and which spits out a single ball with some number inscribed on it when the lottery is complete.

The basic question of inverse probability is: we saw X happen. What does this imply about the reality which caused it? We go not from the distribution of balls bouncing around and project forward to the prediction, but rather we reason backwards from the observation to describe a small configuration of bouncing balls, which we can now imagine a little better as experience continuously reveals its output.

In other words, probability looks forward to predict the future, but inverse probability attempts to go backwards from an observation to see what factors in the past caused it.

The classic demonstration of the use of Bayesian reasoning — of inverse probability — is in medical diagnosis. We find a small lump in a woman’s breast. What are the odds she has breast cancer?

More tests will be needed to uncover the reality of what is happening in her breast, but we are able to do is assign a probability of what is beneath the symptom if we know 1. the likelihood that an average woman will develop breast cancer, 2. the likelihood a woman with no cancer will develop a small lump like the one we have found, and 3. the likelihood that a woman with breast cancer will develop a similar lump. We go backwards from the event, the discovery of the symptom, to reason about the likelihood of the truth of various inner biological conditions and developing processes which may have caused it.

Inverse probability is a tender question. It is a hermeneutic, an interpretive method. It attempts to cover what is concealed within being. It is the quest to penetrate from beyond the veil of expression to find reality’s second hidden face. I hear my lover’s sweet nothings escape her lips and I wonder if she really loves me like she says she does. Perhaps this is a deterministic question of which neurotransmitters have fired: an inquiry upon a system which is impossible to make, for I will never be able to split open her silly head and peer inside the pulsing operating system that waves her fickle tongue. Somehow inverse probability feels so much more crucial than prediction, does it not? We are seemingly always so much less concerned with predicting than uncovering. I will die if she does not love me like she says she does, a thousand palaces of emeralds laid out in my future cannot convince me to live on.

Bayes’ formula, as the formula of inverse probability, encourages us to gradually discover the world — our ground of being — as a probabilistic process which generates experience, or has the possibility to generate various experiences.

Thus, the concept of a Bayesian reasoner can be described as: the man who creates the ground from which we are able to use probability theory to establish truth. He is the assigner of probabilities to things, without whom predicting is impossible. It is via this process, the process of Bayes assigning a ground for prediction, that God-AI is able to create probabilistic estimations of the ground of reality upon which it can make its optimal decisions, in the sketch of the ideal Bayesian reasoner given by Bostrom.

It is mathematically — that is: necessarily and tautologically — true that the Bayesian reasoner is the ideal reasoner, as long as we assume that applying the axioms of probability to predict our experience is possible and desirable. Or in other words, for Bayes’ theorem to be useful there must in fact be some field of reality which is predictable. In the stock market, they often say: “Past performance does not predict future results”. If this is true, then Bayesian reasoning unfortunately cannot work.

Under the frequentist conception of probability, which Bayesian thought is often contrasted with, probabilities are assigned via repetition of events. We can only meaningfully assign a probability to something which has happened repeatedly. If I have known nine lovers, and five of them were unfaithful, I can say that there is a five out of nine chance that my new lover will betray me. But if I am loving for the first time, I am blind, I cannot predict anything at all. And for you, my love, every time I am touched by you, it is always the first time.

The Bayesian, the wielder of inverse probability, instead always steadies himself in advance with a probabilistic ground, constructing a set of expectations which anticipate each possible new event. As the smith of the ground of probabilistic predictions, he must always have his “prior probability” set. He establishes a tentative truth from which experience is predicted, which he then adjusts with experience to update his ever-developing ground.

The question of how to ground one’s expectations in an unknown domain is not clear, and a matter of debate among Bayesians. If I have never loved before, how am I to know how likely my heart is to be broken? “Just start out by calling it as fifty-fifty”, is one semi-solution. “The probability that she loves me may as well be the same as the probability that there are an odd number of petals on this daisy”, I tell myself as I tear the petals off one by one, whispering my prayers.

The obvious problem one sees is that there are an infinite number of possible truths that one has to have anticipated by assigning probabilities to first for Bayesian reasoning to be possible. This problem is either eased or deepened by saying: one assigns probabilities not to truth-claims, but to entire possible worlds from which experience arises.

At least, this is the case according to the formal notion of a Bayesian reasoner described by Solmonoff induction, the method preferred by Rationalism: one has a probability distribution over a set of algorithms which generate experience. This is a formalism on top of a formalism. Ray Solmonoff derived his epistemological theory around 1960 in order to apply Bayesian reasoning in a computational context. Solmonoff was an early pioneer in artificial intelligence: he had recently been one of the invitees to the Dartmouth Summer Research Conference on Artificial Intelligence, the symposium in which artificial intelligence was given its name as a field. Solmonoff was attempting to articulate a process through which a hypothetical computer intelligence would be able to understand the world around it, and discovered Bayes’ formula as the only available tool that would let him do what he wanted. But in Solmonoff’s formulation of Bayes, he replaces the dominant metaphor of mechanical lottery-ball systems and establishes a new paradigm in which we are attempting to parse sequences of letters generated by computers.

Solmonoff imagines that a computer reasoner will have as its input a string of characters, and then it will attempt to unveil, using inverse probability, the conditions for the generation of the characters before it, which is also given computer program. Two computers talking to each other, trying to read each other’s algorithms. In the formalization by Solmonoff, the reasoner must be able to compute all of these possible conditions itself. So for instance, if the reasoner receives a string “ABABABABAB”, it may reason to itself: “a very simple computer program could have generated this, one which says output A, then output B, then do the same again”. But then, if after receiving one more character, the string reads “ABABABABABC” — the aforementioned simple computer program the reasoner had established as the privileged hypothesis is viciously penalized in the calculation of inverse probability, for it could not have possibly have inserted that extra C. Now maybe there are a few more contenders for what generated this text in front of me: a computer program that alternates A-B five times and then inserts a C, or a computer that alternates A-B and then occasionally inserts a C pseudo-randomistically, etc.

Now, Solmonoff’s method is extraordinarily computationally intractable — for every hypothesis the reasoner has about what computer program might have generated the string, it must re-execute every time when it gets a new character, so that it can test to see if it is generating good predictions or not. It goes without saying that this becomes overwhelmingly resource-intensive as soon as we get outside of toy examples such as strings of A, B, and C — how would one for instance be able to hold in one’s mind simulations of the millions of possible authors behind this text in order to penalize and boost their rankings based around whether they would have said the next word? However, Solmonoff’s formalism has attracted a lot of traction in artificial intelligence circles for its purity and formal completeness.

But then it gets even worse, because how do we apply this method when we are attempting to interpret phenomenon in the real world to discover answers to questions of being, e.g. whether or not a woman has breast cancer? We go not from a string of characters generated by a Turing machine, but the entire gestalt of one’s experience as produced by the entirety of being, something seemingly intractable. The solution, for those such as Yudkowsky who endorse Solmonoff induction as a general frame for discovering truth about the world, is to describe reality as computational — or to conceive of one's experience as a string of data points generated by some algorithm. A generative algorithm describes a world, a world predicts experience.

The notion, often entertained by futurists, that reality is a computer simulation, or that there is a second hidden face to reality described by code, can be read as a metaphysical presupposition in Rationalism’s description of how a Bayesian reasoner works. I, a Bayesian reasoner, am able to have expectations of reality because I simulate the laws of physics, as well as other social laws and so on, within my mind. In Yudkowsky’s fears of how AI might arise and eat everyone, the AI does a lot of simulation of physics, of human psychology, of chemistry, etc. Yudkowsky knows in advance it is possible for God-AI to be simulating these things, because as an ideal Bayesian reasoner, simulation is what it does, what it must do.

As a Bayesian, I must simulate my lover to know the truth concealed within her surreptitious words. Within the metaverse of my mind, there are infinite simultaneously unfolding lotteries of love. Infinite virgins and infinite whores swallow rose petals and tea crackers and spit out fortune cookies which reveal their blasphemous secrets. With each word whispered from my lover, some of my whores are killed, and some of them breed. Eventually I hope the quantum superposition of silhouettes I project on my wall resolves itself into a single shimmering woman vibrating quietly before me; but then again, don’t we all.

For every word of the letter I read from her, I must run it by the faithful Penelope simulated in my mind, as well as the lying Jezebel, demanding each of them give me the next letter of the text, then, breathing deeply. checking it against what the next word says... Of course there is an infinite spectrum of women in between these two poles, and with each word, some shrink in size, while others loom terrifyingly large in my mind. Each one in turn must dance, one or another, and eventually the letter ends. By the time I have read her signature, only four dancers remain, all of whom performed perfectly, writing this letter exactly in its entirely, one sincere, one ironic, one sarcastic, and one tragic. I unfocus my eyes and attempt to blur them on top of one another into a single shape. If I don’t know my lover yet, I shall in due time.

Obviously, performing this infinite computation for one’s predictions is intractable. Rationalists say: yes, but it describes an ideal that an actual reasoner may gradually approximate. That it describes an ideal is, again, inarguably and tautologically true, given certain metaphysical and mathematical axioms. We may ask though: is the ideal useful to apply in practice?

Let’s Agree to Disagree

(Bayes in Practice)

First, let’s consider if it is useful for machines. State-of-the-art neural networks — let’s say for instance, GPT-4 can be described as such: GPT-4 is a complex matrix algebra formula which predicts the next word in a text via inputting a matrix representing the existing text thus far and outputting a number representing the next word in the text.

The form of the GPT-4’s equation is defined by the engineer in advance, but its constant factors are not, and must be learned in the training process. To help the reader understand this: the process of training GPT4 is like a more complicated version of some statistical problem where we believe an equation like y = Ax3 + Bx2 + Cx will predict y for a given x, but we must determine the best A, B, and C to discover this equation. In the case of GPT-4, there are over a trillion such A B and Cs. Finding the values for these is what is called learning.

How do we learn A, B, and C? In its training phase, GPT-4 repeatedly tries to guess the next word, and initially it gets it wrong every time. But each time it fails, we can run a calculation to say: which A, B and C etc would have potentially gotten it right? This is called gradient descent. We then push our estimation of A, B and C etc slightly towards the direction of the values that would have been correct in this context (this is called backpropagation), and try again.

The better this predictive system performs, the more it approaches the ideal of a Bayesian reasoner, but this is tautological. Is GPT-4, in its design, modeled after an ideal Bayesian predictor? Not especially. There are explicitly designed Bayesian Neural Networks, but these are for more special purposes, because, as described above, the explicit updating over all possible worlds a Bayesian reasoner must implement is not computationally feasible for anything other than very succinctly defined domains.

GPT-4 updates its truth-notion not with the formal, precise accuracy of Bayes’ formula, but in fits and bursts using gradient descent. GPT-4 takes leaps forwards and backwards. As for the structure of its truth, do GPT-4’s A, B, C, and trillion other parameters describe algorithms for possible worlds? The researcher Janus believes that they do and has argued for this in their post Simulators and elsewhere, but we at Harmless are not entirely convinced... What these numbers encode, what GPT understands as its truth, seems like a profound question to wrestle with, which we might touch upon later.

So it is seen that in the realm of machines, we must make speculative jumps rather than explicitly use Bayes’ formula to update our predictions. What about with humans?

Among Rationalists, amongst the Bayesian community, they will occasionally recommend crunching Bayes’ formula to make some prediction, whether about one’s personal life or some global event. But it is said that what is more important is internalizing the felt sense of Bayes’ formula so that one can reason while conceiving of it as an ideal. Bayes’ formula is, at the end of the day, a vibe you pick up on.

What does this mean? Again, the imperative to become a Bayesian reasoner is the imperative to continuously construct the grounds for probabilistic determinations. The Bayesian reasoner must see himself as someone who knows his priors — he possesses his distribution of prior probability. And when challenged on his expectations of the world, he must present it as a probabilistic claim.

This becomes a set of Rationalist community epistemic norms. When among Bayesians, act like a Bayesian reasoner. Rationalists will ask you “what are your priors?” and it is rude not to answer. For any truth claim you output, they will ask you “What is the probability that this is true?” — no truth claim may be served without this. It is polite to put a probability value, the “epistemic status”, at the beginning of any Rationalist essay you might write.

Rationalists describe a postulate derived from the axioms of Bayesian reasoning called Aumann’s Agreement Theorem which says that any two Bayesian reasoners, assuming goodwill, must eventually converge on an identical set of prior probabilities. When disagreeing with a Rationalist, the most important question becomes what aspect of one’s priors led the disagreement to occur in the first place. Any deviation from a potential shared set of priors means that one person must be held in the wrong. The disagreement should be reconciled quickly, otherwise there is a possibility for pollution in the epistemic commons. To be wrong for too long is considered potentially dangerous, as one falsehood begets another through a chain of corrupted priors, and the picture of reality becomes smudged. It is imperative that when disagreement happens, the interlocutors find the precise point of divergence so that they may re-align. For someone to spend time reading a long-winded critique, one which challenges fundamental assumptions and spends time elaborating upon bizarre metaphors, is to deviate from the efficient ideal of Bayesian reasoning, and if the Rationalist reading this is still with us, we thank you for your patience.

Something very remarkable happens once people start acting this way. It is as if the community itself strives to become the artificial Bayesian Neural Network which GPT-4 for example is not; a collective hivemind that forwards predictions to each other to produce a sense of reality, a prior distribution upon which one can make predictions, which the Rationalists for instance do on the prediction aggregator Metaculus. As we have said, it is like as if to figure out how to outmaneuver the emerging superintelligence, the Rationalist community must first become a superintelligence themselves, the only way we know how, by aggregating the power of many high-IQ human brains.

Rationalists have a very strong sense of their own exceptionality as a community; it seems they feel like they are the only ones capable of uncovering truth. If to act collectively within these norms of Bayesian reasoning is the ideal way to uncover truth, then this is true, for they are the only semi-organized group who acts this way, at least that we know of.

It’s interesting to note that goodwill between the nodes in the Bayesian Network is necessary to perform this process. If someone is duplicitious, or dismissive, or excessively disagreeable, they cannot perform the proper function of forwarding information within the hivemind. As such, it must be the case that people within Rationalism share certain goals. It must be a curated space free from foundational conflicts. There is a remarkable essay by famed Rationalist Scott Alexander called “Conflict Theory v. Mistake Theory” in which he contrasts two theories of disagreement, one in which people disagree because of deviating beliefs about reality which they can resolve, and one in which people disagree due to conflicts. After spending so long immersed in the politics-free Rationalist space in which Bayesian reasoners with remarkably little drama work on gradually converging on their shared set of priors so they may coordinate action, Scott realizes with a sort of shock that most people exist in a world where political disagreement arises from inextricable conflicts (such as competing claims on shared resources, national and class antagonisms, etc). This leads to a situation where truly competing wills are not present in Rationalism. One can entertain a lot of bold proposals in a Rationalist space, but if one is committed to the idea that, for example, libertarian capitalism is a bad axiom, or that software-engineering types should have less power in the world rather than more, one is not able to integrate oneself into Rationalism.

As such, the Rationalist community, despite its thriving debate and blogging culture, is not exactly a forum for open, free, unguided inquiry like an Athenian gymnasium or Enlightenment coffeehouse or French intellectual salon. The Rationalist community is an hivemind constructed for the purposes of something — what exactly? Rationality is winning, but winning at what? It depends on who you ask, for some Rationalists it is merely to increasingly cultivate the art of rationality: increasingly honing its own powers of superintelligence, suspending the moment where it gets applied to a particular task. For some Rationalists it is just to make friends. For Yudkowsky, it is to establish a community of people who think like him so that he does not need to solve the AI Alignment problem alone.

How has Rationalism fared at this so far? In its initial days, it seemed as if the Rationalists believed that their methods of reasoning would give them near-superpowers and allow them to take over the world very quickly. Scott Alexander wrote an entertaining post in 2009 titled “Extreme Rationality: It’s Probably Not That Great” urging them against some of their boundless optimism with respect to their methods. But there have since been some attempts at Rationalists to gain serious power — exactly which ones qualify probably depends on finding some difficult boundary of what counts as a true Scotsman. Is Vitalik Buterin a Rationalist? Is Sam Bankman-Fried?

It’s clear that Rationalism failed in its primary task of allowing Yudkowsky to form a collective mind capable of solving Alignment alongside him. In AGI Ruin, in which he declares despair over Alignment and predicts a >99% chance of death, he repeatedly bemoans the fact that he is “the only one who could have generated this document” and that “humanity has no other Eliezer Yudkowsky level pieces on the board”. “Paul Christiano’s incredibly complicated schemes have no chance of working”, he laments about one of his closest collaborators. There are not many truths that Rationalism collectively discovered that it did not know at first, nor is there anything it radically changed its mind on. And while Rationalism’s founder, Yudkowsky, has declared a >99% chance of death from AI, few in this community are updating from his posteriors to go along with him, or can even really feel like they understand fully where his confidence comes from, much to his great frustration. Rationalist epistemic norms have allowed for a lot of riveting debate, great writing, and the formation of many friendships, but it’s not clear that people actually converge on a ground of priors, or that performing the speech-patterns of a Bayesian reasoner actually allows one to approach the ideal one is approximating. People don’t usually end up finding a common ground when they debate — usually they end up relaxing parts of their position while bunkering into some increasingly pedantic and obscure point until the other person gets bored. Disagreement doesn’t get resolved, it spirals endlessly. The tree-like structure of the LessWrong and Astral Codex Ten comment sections reveals this all too well. People aren’t especially exchanging a set of probabilities over truth-claims when they discourse, least of all in the fluid, low-friction manner expected of a network. What people mostly do is quibble over a set of linguistic frames.

What can be done? Is it possible to construct something more optimal? We feel that the failure of the Bayesian community to come to a healthy consensus arises from this structure it places upon the operation through which one perceives, investigates, learns from the world, uncovers reality. Knowledge of reality is held to be ability to model it as an algorithm which generates one’s experience. But there is something rather hubristic about this idea: that in order to understand reality and be guided by it, one must also fit it inside one’s head.

The Choir of Flowers

(Beauty as Episteme)

Perhaps the reader will follow us along with a philosophical experiment of sorts. Let’s begin by repeating: anything which is able to make predictions, to the extent that its predictions better anticipate reality, increasingly approximates the ideal of a Bayesian reasoner.

At Harmless, we noticed that the output of neural networks and their resulting effects on society best is predicted as an acceleration and intensification of existing trends. People have long been complaining about the content on Netflix being algorithmically generated, before this actually became possible. The flattening of style that will inevitably happen with generative LLMs being widely applied has already been well underway in the past decade, with the flattening of style in all fields, interfaces, architecture, design, and speech. The cheapness of artistic production flattening art and making its economic viability difficult has already been felt in music for instance, with artists making music for the Spotify playlist and not for the LP, leading to the rapid overturn in popularity and a post-fame era in popular art.

Briefly, we can describe a Bayesian Neural Network as such: a Bayesian Neural Network is a set of nodes, each tasked with declaring a certain probability over the same truth-claim; this could be: is a given image a picture of a cat or a dog, is the enemy planning to attack tomorrow, is AI going to kill everybody in the next decade, etc. (Technically, in a Bayesian Neural Network each node forwards a distribution over all possible probabilities, this is actually what differentiates it from a standard neural network.) In the lowest level of the Bayesian Neural Network, the nodes each pay attention to a specific piece of the evidence at hand and use it to establish their own estimation of the probability. It is like the parable of the blind men and the elephant: one node looks at the ears, one node looks at the eyes, one node looks at the feet, and each gives its estimation of whether it is looking at a dog or a cat. Intermediate levels aggregate predictions from lower-level nodes, they are like managers who collect business reports from their employees with some skepticism, noting down which ones are underperforming. The final layer is like a council of wise men who receive all the reports and usher forth an ultimate judgment.

This led us to wonder: is the world itself almost like a kind of neural network? Does the world learn? Could that be the secret truth behind mysterious phenomena such as Moore's law: that reality itself is like a Bayesian reasoner, which is really only to say that it reasons? Now let us describe the world like this. Anything that exists, insofar as it has a discrete existential status, we can describe as expressing an existential hypothesis. Everything speaks to us: “I exist”.

I am looking at a flower in a vase. In a few days, it will wilt, die, decay, but for now, it is alive, and it tells me such. The probabilistic quality to this claim comes into play when we understand that everything tells us it exists, but not with equal confidence. Some men bellow it with absolute certainty, but some hardly seem sure. Signs of death in living matter haunt us everywhere; jaundiced cheeks and pockmarks hastily covered up with makeup. level nodes, they are like managers who collect business reports from their employees with some skepticism, noting down which ones are underperforming. The final layer is like a council of wise men who receive all the reports and usher forth an ultimate judgment. indeterminate mixture of all these various realities underlying the event. So, surely living things, animals, humans, corporations in a competitive environment, are like nodes in a network which expresses the odds that life continues to exist. But a process like this can be said to occur even in inert matter.

There is a beautiful illustration in Yudkowsky's exposition of Bayes’ theorem on Arbital which shows the correspondence between Bayesian prediction and a physical system by describing a waterfall that exists at the convergence between two streams. A fifth of the water supply of the first stream is diverted into the waterfall while a third of the water supply of the second stream is, now the waterfall contains a mix of the particles in these two streams. The analogy to Bayes is this: first stream is the multiverse of possible worlds in which my lover loves me, the second stream is the one where she does not, and the waterfall is her taking two hours and forty-five minutes to text me back “Haha”, with its expression of probability, this vulgar inseparable mixture.

As such, a sedimentary rock expresses the reality of worlds in which quartz travels down one stream to deposit itself in a bank and intersects with a stream of silicon. As the sediment builds up, it expresses the reality of its existence with increased vigor, as well as the infinite worlds of quartz and silicon expressed in its particles. The rock is built up by the streams and broken down by the air, it provides the initial material for soil. Within this soil, a flower grows. I look at it and I see it not only scream its own existence, but a probabilistic expression of infinite streams of pollen floating through the air, infinite bumblebees carrying it across the sky, streams of minerals, swarms and swarms and swarms of bugs. Life describes not only its own life but the life of everything which contributes to it, life testifies to the conditions for life. When I see life, I know that I may live.

(Although this is only generally true for apex predators like man — that the conditions for life and good health are always a positive sign that also one may live. If one is a prey animal, to exist alongside a very healthy predator is the worst possible thing, and for that reason it might be better to go to less life-generating environments. This is why Nietzsche said that his philosophy of health was a master morality and characterized his philosophical opponents as prey animals, for they mainly define their moral system against fear of some oppressor.)

As Nietzsche told us, it is so much easier to evaluate health than it is to evaluate truth claims; it is not really clear why we even waste time bothering to do the latter. To read through the million words of the LessWrong Sequences or worse, the dense mathematical decision theory published by MIRI on Alignment, is overwhelming and laborious, but to look at Rationalism and notice its death-spiral is very clear. Let us make a gradient descent to greener pastures, more fertile fields.

We look at the cultural products produced by competing actors in the market and we see that the neural network of the universe is being trained. Art is the greatest expression we can make of our health — our confidence in ourselves, our capacity for deep thought, our understanding of the world, our ability to spend time on the non-essential, and above all, our ability to appreciate the marks of good health in others. An economically thriving city produces a cultural scene. I have given up trying to understand what goes on in her head, because I know what it means to have red lips and long flowing hair. Somehow, when I stopped wondering about her, and started only gazing at her, that was when everything changed.

Beauty is more efficient, more effective. The ideal posed by Bayes and Solmonoff, to simulate all these millions of worlds, is totally impossible, totally unthinkable. I've lost the ability to maintain all these dancers in my head; they have started spinning off course in oblique directions. The more I try to simulate my lover, the more she seems to speak only in riddles. The more she started to speak in riddles, the less we felt like communicating in words. These days, just hand each other flowers. Each flower is a portal to a multiverse, but not even a multiverse which needs to be simulated — one which reveals itself perfectly in its expression: every petal shows us a multitude of streams of bees. Flowers do not need to be interpreted. They are love letters that are not sequences. They sing; they testify.

It is for this same reason that expressing one’s truth claims as couched in probability should perhaps be rejected, as well as the attempt to converge with other reasoners. Whatever one’s hypothesis is, one should commit to it with maximum intensity and vigor. The most noble life is the one where you exist as a truth claim. Let reality herself be the judge — she is a little slow to reason, but loves to be impressed — this is the only way the princess learns.

Next: On Game Theory →