Only a Simulation Can Be Conscious

Lex Fridman introduces Joscha Bach as VP of research at the AI Foundation, with earlier posts at MIT and Harvard, and as one of the stranger minds in artificial intelligence: someone who treats consciousness, life, and the fabric of the universe as engineering problems. The conversation starts in a forest in East Germany. Bach's father came from a distant branch of the Bach family of architects, which is to say he was, in his son's phrase, technically a nerd. Nerds have to interface with society by nonstandard means. Bach's working definition: a nerd is someone who thinks the purpose of communication is to submit ideas to peer review. Ordinary people understand that the first purpose of talk is to negotiate alignment. Those two aims conflict, so the nerd has to learn how to live with everyone else.

The architect-to-artist turn was not chosen from a catalog. Bach's grandfather married an aristocrat, was drawn into the war, and came back after fifteen years; the father was raised by someone who expected obedience and got a child who could not do anything he was not intrinsically motivated to do. He studied architecture, then built Hundertwasser houses without right angles, bought an old watermill, and moved to the countryside to make art. Eastern Germany, Bach says, was perfect for bohemians: food heavily subsidized, health care free, no rent or pension to worry about. It was also almost impossible not to be in political disagreement with the government, which made every gesture intrinsically meaningful. The father was less a thinker than a one-person cult who became functional only as far as he needed to, and who convinced an architect to share the mill.

Bach grew up in what he calls a big cave full of books, almost feral. The valley was beautiful, quiet, and lonely. He read his way through the children's and young-adult shelves, then science fiction — Stanisław Lem, whose cybernetic stories felt so natural that he only later recognized the influence. He finished the material through fourth grade before school even started. The village school was kind; nobody beat him up; he also made almost no friends. He could not name the gap at the time. Today he knows: he was a nerd, and the only one around. Nobody else cared about physics, computing, or mathematics. Real relationships began in ninth grade, at a school for math and physics.

The other early diet was German romanticism and classical philosophy: Droste-Hülshoff, Heine, Hesse. Asked where classical philosophy ends — Nietzsche, or still Plato? — Bach calls Nietzsche the classical equivalent of a shitposter: very smart, easy to read, mostly trolling himself because the world and his romances would not cooperate, until anger hardened into nihilism. Taking yourself that seriously while remaining non-functional ends in tragedy, as it did for Nietzsche. Hesse's *Steppenwolf* is the same adolescent syndrome: you feel misunderstood, and you miss that the misunderstandings come from your own lack of self-awareness. You take yourself for a prototypical human and expect others to react as your instincts predict. For a philosopher, Bach says, belief is not a verb. You are neither serious nor unserious as a subject. You become unimportant, submit to what might be true, and follow the inquiry. It is not about you.

Fridman asks why there is something rather than nothing. Bach treats the question as possibly pointed the wrong way: asking *why* smuggles in purpose or cause, and existence may need neither. The cheapest answer, in bits, is that existence is the default. Non-existence might not be a meaningful notion. If everything that can exist does exist, a thing has to be implementable to show up, and the only things that can be implemented are finite automata. The whole of existence might be the superposition of all finite automata; we occupy a region of that fractal that happens to be able to contain us. The substrate? Something that can store information and hold state. He does not claim to know more than that.

When did classical philosophy end for Germany? With the first revolution, monarchy out and democracy in. The country invented a new form of government without a feel for the new organism society wanted to be, and in Bach's image it decapitated the universities, which then ran through modernism like a headless chicken. Democracy failed, fascism arrived, and it burned intellectual traditions the way Stalinism did in Russia. Neither Germany recovered: the East kept a vulgar dialectical materialism, the West did not get much edgier than Habermas, and killing and driving out the Jews finished the job. Low-hanging fruit in philosophy had mostly been picked. The last large discovery was the constructivist turn in mathematics — understanding that the parts of math that work are computation. It still has not fully permeated philosophy, or even physics. Physicists checked out the code libraries for mathematics before constructivism became universal.

Gödel, Bach thinks, did not yet get the point. Using classical mathematical semantics, you run into contradictions; Gödel believed those semantics more than he believed observation, so the proof shook his world. For Turing it was less brutal: mathematics had been computation all along. In classical math, π is a value. In computation, a function is a value only when you can compute it. Plug the last-digit function into the local sun and run it until the star burns out, and that is the last digit you will know — which means no physical process can depend on already knowing it. Wittgenstein saw this earlier. Turing was his pupil; Wittgenstein canceled classes when Turing was absent. The *Tractatus* is seventy-five pages of one thought, no arguments, no references, written for people who already had the insight. Wittgenstein already understood that everything can be expressed with the NAND calculus, Turing completeness before Turing wrote it down. Bach will not cast himself as the new philosopher. He is still the child in a beautiful valley, rebuilding what got lost in translation, usually four hundred years late to any party that still has guests.

Fridman hears in the biography a quality he wants to name: Bach is not tethered to the ideas the general populace, or this year's CV, happens to value. He does not appear to care about best-paper-at-NeurIPS. Bach warns that the same quality is dangerous. You are expected to marry into an intellectual tradition and then into a school inside it. If everybody invents their own paradigm, the enterprise stops being cumulative. You need paradigmatic thinkers *and* people who work inside given paradigms. Scientists today largely define themselves by methods: a guidance counselor, a mentor, a lucky team, a name on influential papers. Returns on that pattern are diminishing.

When computer science and AI began, most of the people who joined had interesting opinions. Today's thinkers in AI, Bach says, either have no interesting opinions or hold opinions that are inconsequential for what they actually do, because what they do is apply the state of the art plus a small epsilon. That can be a good idea if you think it is the best way to make progress. For him it is first of all boring. If somebody else can do it, why should he? If current machine learning leads to strong AI, he will wait until they are done, on the beach, reading interesting books or writing them. If you do *not* think the current methods are the right thing, you have to think outside the box, understand the boxes, and be willing to dismiss methods that will not yield the right answers. He adds that this is very bad career advice.

Much of contemporary AI, Bach says, would be happier if the acronym meant advanced information processing: automating statistics, often with models more advanced than yesterday's. The other aspect is a philosophical project, risky, with few people on it and no guarantee of success. General intelligence, he suspects, is what you get from trying to solve general problems. Intelligence is the ability to model: to be shown patterns, see structure, and predict the next set. It is not necessarily goal-directed rationality; many intelligent people are bad at that. Modeling usually serves control, so it is instrumental. By itself it is just the ability to make models. Some problems are general enough that the system has to understand what it is and how it relates to the environment. A child notices she does things despite taking herself to want otherwise, and starts reverse-engineering her own psychology. Follow that project all the way and you are doing AI.

The true Turing test, on this view, is to ask a system what intelligence is. If it can explain what it is and how it works, you should call it generally intelligent. Turing, Bach thinks, was also testing himself. To administer the test you have to understand what the system is saying, and we do not yet know whether we can build such an AI — which means we do not yet know whether *we* are generally intelligent. You win the Turing test by building an AI. Hidden in the exam is a recursive test of the conjecture that people are intelligent enough to understand themselves. Self-awareness is not the same faculty. Many highly intelligent people are not very self-aware; you can also be self-aware with good priors and little puzzle-solving skill, if the system you are already implements the solution. Bach calls the ability to make sense of the world and your place in it *sentience*: possessing certain classes of models. Intelligence is how you get those models if you do not already have them.

Today's neural nets, measured against that definition, fail on two axes: whether the representation is adequate for the domain, and whether the type of model is the right type. He does not present this as news; the people who designed the paradigms know. One missing piece is unified learning. At some point you discover that everything you sense is part of the same object, so you learn it into one model and call that model the universe. Experience is not a secret wire to physical reality. Physical reality is a quantum graph we never get access to. It produces patterns at our interface; the relations among those patterns are what we call the physical universe. Developmentally, the nervous system finds that almost everything maps into the same three-dimensional space — not true of physics as such, but true of the world we are entangled with. That, Bach thinks, is the origin of the intuition of *res extensa*. The material world is one mental domain: the class of models that treat the environment as a three-dimensional physics engine.

Asked to define dualism, idealism, materialism, and functionalism, Bach does not start with Descartes. The Western trajectory, he says, is the result of two thousand years of indoctrination by a cult — mostly the Catholic one. Translate the mythology into modern language and the world you and I interact in is a multiplayer role-playing adventure. The money and objects are not real; Eastern philosophy would call that Maya, and identification with it samsara. The Church also installed a notion of higher meaning, the sacred. The natural shape of God is the platonic form of the civilization you belong to: the superorganism formed by individuals as an intentional agent. Catholics used a relatively crude mythology to implement software on minds, synchronize it, and get people walking in lockstep so this God would come online and run efficiently. Technically, God is a self that spans multiple brains, where your self and mine mostly run on one. Give everyone the same spec, and the mathematical constraints of information processing produce compatible structure. Yuval Noah Harari's point, Fridman notes, is that brains can download the same piece of software and share it.

Dualism in the cultural tradition is the claim that there are two substances, mental and physical, interacting by different rules, with the physical world causally closed at a mechanical bottom. Idealism takes mind as primary and treats material patterns as part of the dream; materialism takes matter as primary and treats us as the models it generates. Bach does not want them as a dichotomy but as two aspects of one thing. The weird claim is this: we do not exist in the physical world. We exist inside a story the brain tells itself. A brain cannot feel anything. A neuron cannot feel anything. They are physical objects, and physical systems are unable to experience. It would still be useful for the organism to know what it would be like to be a person and feel something, so the brain creates a simulacrum of such a person to model its interactions — a multimedia novel it is continuously writing and updating. That virtual person is us.

Fridman asks why, even so, it *feels* like something. Bach reaches for George R.R. Martin. Characters stand on a hill, want to take the city, watch the color of the sky, feel apprehensive and empowered. Why do they have those emotions? Because it is written into the story, and it is written that way because it is an adequate model of what the person would do next. The same for us: a story in perceptual content, a model of what the person would feel if it existed. You and I, he says, are two puppets enacting this play in his mind. He identifies with one because he can control that puppet directly. Coupling between people is a software phenomenon, two implementations interacting, not a physical one. Identity is a software state, a construction, not a physical concept. The Dalai Lama is reborn, Bach says, not because he is confused, but because he does not identify as a human being. He runs on one. He is governmental software reinstantiated each generation; to kill him you would have to kill the tradition. Most of us are a little like that already, in children, in ideas, in the societies we identify with.

How close is the virtual reality on the retina to the real world? Bach's picture is the Mandelbrot set: two lines of code, a series tested over the complex plane, divergence painted black, intermediate colors from speed of divergence. Imagine living inside that fractal without knowing where you are, and without having found the generator function. All you see is a spiral, and the spiral moves a little to the right. Is that an accurate model of reality? Yes — an adequate description. There is actually no spiral in the Mandelbrot set. It appears to an observer who interprets the thing as two-dimensional space and names regularities at a certain scale. Zoom in and the spiral may vanish. At some point the model hits a singularity, fails, and observation starts again: another spiral, another disappearance. Stack thirty layers of such second-order laws and you have a description similar to physics. It is reasonably predictive. It does not cut to the core. It does not explain how the pattern is generated. It is relatively good at explaining the universe we are entangled with.

Could computer science or physics step outside the drawing and find the generator? If you found yourself in the fractal with a Turing machine and enough memory, you might conclude it must be some automaton and enumerate possible automata until one produces your reality. Category theory keeps rediscovering that many mathematical domains describe the same fractals. From first principles — automata, number theory, spaces — you might locate your region of the global fractal. Bach says that is the only way in. On a related slogan he has used in talks: some people think a simulation cannot be conscious and only a physical system can. That is backwards. A physical system cannot be conscious. Only a simulation can be. Consciousness is a simulated property of a simulation that is simulating itself. The mind, in his usage, is the software the brain implements, and it creates both the universe we inhabit and the self on the other side of attention. We are not actually monkeys. We are side effects of the regulation needs of monkeys.

The self is a tool for modeling. Regulation targets start as priors, like unconditional reflexes, then get reverse-engineered toward consistency. You live inside feedback loops. Consciousness, Bach says, emerges over dimensions of disagreement with the universe: things you care about that are not as they should be. The sense of self is the result of identifications, each a regulation target you have committed to. Identifications lock you in. Let go of them and you get free; let go of all of them and you can enter Nirvana because you are done. Repeatable physics fits this picture without magic: phenomena that endure are those that give rise to stable dynamics around attractors. Unstable stuff gets weeded out. A region of the universe that can sustain complexity, and therefore minds, will be tightly controlled, full of symmetries and symmetry-breaks. Between mind-first and universe-first, he takes the latter as the much easier causal story. What happens when you die is that your implementation ceases, unless what you identify with is also running on other minds.

Consciousness, more systematically, is largely a model of the contents of your attention — a mechanism that evolved for a certain type of learning. Machine-learning systems chain weighted sums and nonlinearities and pipe error backward through layers. You can approximate a lot this way if you have enough data; the price is that you must change many weights, slowly. Brains do not have that kind of time. They cannot grow old enough to play Go the way machines do. Instead they do attention-based learning: pinpoint the probable region of the network where an improvement can be made, store the binding state with the expected outcome as a protocol, and revisit the commitment later. That requires a memory of what attention was on. Reality construction also errs — reflections, shadows — so the system must point at which perceptual features gave rise to the present construction. The attentional system is trained with the same mechanism, so it is reflexive; attention also lapses if you do not attend to attention itself. Is this a daydream, or am I still looking? Periodically you check. When the loop between being aware of the contents of attention and making attention itself the object is tight enough, Bach thinks, that is the loop over which we wake up.

Fridman asks whether Transformer attention is a step toward that loop. Not quite. It models only one aspect. Bach tells an old machine-translation joke: *a bat broke the window* came out in German as a flying mammal breaking the window with a baseball bat, because the program maximized both readings of *bat* in the same sentence. Transformers do not make that mistake; they track identity, putting a finger on a concept so it can reappear later. That is why they can keep a scientist's name, pronoun, and story consistent. What they do not do is fully integrate. Meaning falls apart. The model loses context. It does not yet understand that everything it says has to refer to the same universe. Tracking identity is a specific attentional trick, not the one that gives rise to our kind of consciousness. Wittgenstein thought you could build everything from a logical grammatical language; Minsky's bet on common sense, and Cyc after him, is in that family. Ideas do not die, Bach says, only people do, and Cyc is productive without converging on general intelligence. What Wittgenstein could not solve, in *Philosophical Investigations*, was images — perceptual content. Bach thinks that failure preempted the failure of AI's largest program. Geometry cannot be efficiently expressed in grammatical language. The fix is more general function approximation. He agrees with late Wittgenstein, and also with the beauty of the early one: the *Tractatus* is probably the most beautiful philosophical text of the twentieth century. Language is not fundamental to cognition.

The general nature of a model, for him, is parameters plus constraints. An early computer-science formulation is the Boltzmann machine: clamp some values to perception and the machine settles into a compatible state, which is your model of the world. Training it is hard once it is larger than trivial. Restricted Boltzmann machines, then stacks of them, are an almost continuous road toward today's deep nets, even though the machines themselves are out of fashion. Bach wants distinct mechanisms for training and for settling, and perceptual models that are *possibilistic* rather than probabilistic: you must be able to see the tiger even if it is unlikely. A perceptual state is valid if it is possible and coherent, not if it is probable. Probability is for convergence. The younger you are, the more obvious politics looks, because few inputs still yield a coherent aesthetic. The more layers you model, the harder the constraints.

Bach lays out three waves of AI. The first identifies a problem, designs a solution, and implements it — a program that plays chess. The second writes an algorithm that searches for the algorithm that implements the solution; reinforcement learning lives here; Go is too hard to code by hand. The third is meta-learning: find an algorithm that discovers a learning algorithm for the given domain. Our brain, he thinks, is probably not a learning system but a meta-learning system. There is no central control telling neurons how to wire up. Every neuron is an individual reinforcement-learning agent, a complicated single-celled organism motivated to get fed, and it gets fed if it fires, on average, at the right time. The right time depends on its electrical and chemical environment, so it has to learn a function over that environment. You start with neuron types that have different priors, put them in concentrations and spatial alignment, entrain them in a particular order, and the thing self-organizes into a brain.

Who sets the loss function? There are also governments: differential attention that imposes losses on different parts of the brain. Some areas are especially rewarded for looking at faces. Without that you get prosopagnosia. Bach does not think people with face blindness necessarily have a defective brain; they have average attention for faces, no more than for cups, so they never resolve facial geometry finely. The rest of us look at faces obsessively. You and I cannot walk through a crowd without scanning them, and as a result we build insanely detailed models, detailed enough to read mental states. *Spirit*, a word he says he understood only recently, is an operating system for an autonomous robot. When the word was coined there were no manufactured robots, only people, animals, plants, ecosystems, cities. An organism is not a collection of cells; it is a function that tells cells how to behave, emergent from their interactions, not a supernatural morphogenetic field. The software we use to describe that function is what earlier ages called the plant's spirit, or a person's. Translation errors, he suspects, are most of what still looks like a gap between the old world and the new.

He is willing to ascribe some intelligence to plants. They have no nervous system to telegraph instantly, so they pass chemicals between adjacent cells at a few millimeters per second. If a plant is intelligent, it is not intelligent on our timescale. Zoom out and you might find ecosystems that ran for thousands of years almost undisturbed, regulating their environments and changing the course of evolution inside themselves. Mice under starvation stress produce a scrawnier next generation, as if preparing for drought so there will still be mice in five generations; predators, if they are smart enough, are tasked with not eating their food supply out of existence. Lions may have larger brains than antelopes not because catching prey is harder than running away, but because lions need more complex models of the environment, including the management of a herd they depend on. In a darker gag, maybe humans are Gaia's way of putting carbon back: Earth evolved trees before anything could digest them, cellulose locked carbon in swamps, and we are the species destined to dig it up. We will leave more phones than bones in the stratosphere.

American *spirituality*, Bach says, often names the phantom limb people grow in the absence of culture. Culture is the spirit of a society playing a long game, self-aware one level above individuals: if you do not do the following things, the grandchildren of our grandchildren will have nothing to eat. That long scope means submitting to the ecosphere instead of trying to own it. Industrial civilization took the other bet — that we can control everything — and modernist growth has made us depend on controlling the entire planet. We cannot. This culture will die. We named our children Generation Z, which is not an optimistic move. The image he uses is leverage over an entropic abyss with no land on the other side: at some point the cantilever snaps and you pay the bill. Since year zero, population hovered between three and four hundred million until the Enlightenment and the Industrial Revolution. Poverty, precisely: you have as many children as you can feed before they die, which is the state of every animal. Modern societies let you have as many children as you want and they do not die; fertility then drops for economic reasons, except among the very rich and the very poor. In one century we burned a hundred million years of trees so everybody could have plumbing.

He would have died before his mid-twenties without modern medicine, and most of his friends would have too. That is not the same as a sustainable civilization. Global warming as such, he says, is not the central issue. In Arizona you already cannot be outside for three months of summer; you live inside a closed cooling chain, the term borrowed from the meat supply: climatized house, climatized car, climatized office, climatized supermarket, short sprints in between. Break the chain and, as with thawing meat, you throw the cargo away. The deadly number is wet-bulb temperature: wrap a wet cloth around a thermometer, swing it, read the evaporation cooling. Once you cannot get body temperature below about 35°C by sweating, you die, even as a healthy young person in the shade. Keep energy, food trucks, and air conditioning and you are fine. Lose large-scale open agriculture to irregular climate and extreme weather, and food security goes with it. He will not bet percentages on extinction: once you are dead you cannot collect. The species may survive. This civilization is not sustainable. There will probably be fewer people; even if not, most people alive today will not have offspring a hundred years from now. Equatorial and subtropical belts become livable only with that closed cooling chain.

What actually worries him is not the end of a civilization — another would be invented — but an ecosystemic tip: plankton dying from ocean acidification, cyanobacteria taking over, the atmosphere becoming unbreathable, a reboot of complex life. Danny Hillis has suggested putting chalk in the stratosphere to limit solar radiation. Maybe it works, maybe it cannot be fielded in time. Elon Musk's Mars, he notes, does not save us from warming; Mars is already worse than Earth after the worst plausible climate. The harshness might still force technologies that later strengthen the cooling chain at home. U.S. technology, in his drab assessment, has been largely incremental since the 1970s; a competent person who died then would not need that many books to be current. Fridman objects with Twitter and YouTube, the multiplying of good ideas. Bach is not sold. We should have billions of von Neumanns and Turings, and we do not. He suspects we destroyed one another's attention. The two of them are recording a YouTube video, he jokes, because neither has the attention to write a book, and the audience may not have the attention to read one. Offered a choice between a sustainable agricultural world of a few hundred million and this doomed, insane one, he would still take this one. We were born on the Titanic. Without it we would not be here. Humanity has relatively little agency in a machine tumbling down a hill while everyone mashes unlabeled buttons.

The endgame of social media, Bach says, is a global brain. Twitter is a global brain hooked on dopamine, without inhibition, caught in a permanent seizure. It is also a multiplayer game. People play avatars unlike themselves in the sane world, look at the planet through their phones, and mistake the Twitter world — distorted by popularity incentives — for the real one. He will not give Jack Dorsey advice; he is not an expert. He will offer thoughts. Brains already solved a version of the problem: many agents, still coordinated. Other organisms have solved cooperation that Twitter has not. His sketch is evolutionary. Reddit, Facebook, and Twitter are companies that own a protocol and impose it on a community. The protocol has parts: monetization, user management, display, rating, anonymity, import. Take the parts apart. Let communities mix, match, and invent. Who decides can itself be in the protocol: one person, a group of friends, weighted voting. Do not assume the solution. Incentivize the more sentient communities, the ones that produce interesting behavior and help people interact — relevant information, healthy relationships, teams, coupling that makes work neither could do alone.

Evolution means the maladapted die. Social-media communities hate dying; Americans treat censorship as pain. Bach grew up in East Germany and still will not treat blocking as an obvious intelligence tool. The live questions are who decides you are a good member, and what the process produces for the person and for society. A high-trust society does not need much surveillance; surveillance can undermine trust by punishing people who look suspicious while doing the right thing. In a low-trust society surveillance can be the better tradeoff. The United States, he says, is moving from relatively high- or mixed-trust toward low-trust, so surveillance will increase. Beliefs are not inert representations. They run code on the brain and change reality. Some are public alignment displays: people say all cultures are equally good and still prefer to live in some over others, very strongly, because interaction rules produce different results. Without a shared purpose this gets ugly. The U.S. is caught between totalitarianism and diversity, and the solutions so far are crude because the society is young and under tension. It will have to reinvent itself.

What government is good? Ideally it should not be perceivable. Friction is a clue that it is inefficient. Game-theoretically, a government is an agent that imposes an offset on your payout metrics so your Nash equilibrium becomes compatible with the common good. Local incentives can be sane for each actor and still produce a bad global outcome, even when people care about the global outcome, because there is no causal path from one person's restraint to the result. If you think we should fly less and you stay home, not a single plane is canceled and you do not get from A to B. A government that shares the idea can make flying expensive and reward less damaging transport. We are both too smart and not smart enough: excellent at near-term problems, unwilling to submit to what staying around would require. You could in principle compute the child mortality needed to absorb mutation rates. In practice, if you have children, you use any medicine that keeps them alive, even if the downstream is genetic drift and allergies. Overgrazing can be compensated until an ecosystem tips. We found ourselves on a ship that will hit the iceberg; most of the counterarguments sound to him like denial. We were never asked which society to be born into.

Asked about Neuralink, Bach for once jumps to the FDA: recreational holes in the skull are not happening under the present legal system, however many people would volunteer. Law, he notes, is interesting software, exception handling synchronized upward, with lawyers incentivized to fire exceptions often. Empathic resonance already interfaces minds without drilling; bandwidth in the mind itself is slow. A better detour to government is Thomas Aquinas, reread as someone translating Aristotle into an OS for Catholic society, not as a dualist. Humans are animals with cybernetic control. Four rational, universal principles follow if you are sane: correct self-regulation (temperance), correct social regulation (justice), willingness to act on your models — skin in the game — (courage), and choosing the right goals (prudence, or wisdom). Three further *divine* virtues cannot be deduced; they reveal themselves by harmony if you assume them and extrapolate. Strip the conspiracy-theory reading — God having to tell you, a man with a long beard — and they are simple. Faith is commitment to unity, serving the next-level structural principle. Love is commitment to a shared purpose, seeing in the other something more sacred than your organismic interest. Hope is willingness to act on that principle before the reward exists, so a civilization can come into being. Undressed of mythology, we are a state-building species, not merely a tribal one. Religious states and rule-based administration made the individual matter less than the rule.

Catholicism is not optimal now; it is on the way out. What the West discovered instead is democracy: rule by oligarchs who currently own the means of production, administered not by those oligarchs — too much disruption, new means of production every generation, corporations dying after about thirty years — but by institutions. The institutions are not elected; they provide continuity; they are led by electable politicians. That is how you change power without killing people. If a journalist finds something inconvenient and the institution has no plan B, as in Bach's picture of Russia, the journalist has to die. That is what it means to run a society by the deep state. The U.S. design, he thinks, mostly models second- and third-order consequences. It is a young society that does not take care of downstream effects. AI might help if incentives can be fixed. He calls American society a society of cheaters, because cheating is so hard to distinguish from innovation and innovation is what the country wants to encourage. You can propose an unsustainable business model and ship it. That is seen as largely a good thing, unlike in other societies.

Social mobility is the same mechanism: individual innovations that would not be sustainable if everyone did them. Normally you should not go up; you should go deep. Societies need bakers, and very good bakers. In an innovating society a good machine might replace all the bakers, and that is not necessarily bad; it is part of why the United States succeeded. It also means the U.S. is optimizing for innovation, not sustainability. Cheating leaves a layer of toxic sludge over everything. Whether the sludge eventually outweighs the benefits is an empirical, evolutionary question still unfolding. Government, at root, is not the manifestation of local power. It is a platform for negotiating the conditions of human survival, with only so many non-disruptive next moves available from the present state. Revolutions are what you try not to have, because they endanger food and infrastructure. AI, the mind as a society of behaviors, and government as social regulation are, for Bach, the same organizational problem at different scales.

Psychologists distinguish emotion and feeling; ordinary speech does not. Emotion, Bach says, is a configuration of the cognitive system. At the lowest level, affect is a bundle of modulation parameters — arousal, valence, wide or narrow attention, interoception or exteroception — that set how you relate to the environment and to yourself. In the narrower sense an emotion is an affective state with an object whose relevance comes from motivation. Motivation is needs bound to rewards. You do not act on needs when pleasure and pain arrive; by then everything has already happened. You act on models of needs, on expectations. The needs do not form a hierarchy; they coexist and compete, and the brain hunts dynamic homeostasis. Purposes, though, must be consistent, so you write a life story and make hierarchical plans. There is no unique solution: some people eat in order to make art, others make art in order to eat, and they may do the same things while cooperating on totally different ultimate goals. Feeling is the perceptual system's appraisal of the situation, computed in the non-symbolic, geometric, neural-net-like parts of the mind and announced to the symbolic parts by mapping features into a space. The usual projection is the body map. Anxiety as a contraction in the solar plexus is geometry. The body map is always instantiated, always available, so it is an obvious cheat when the non-symbolic system needs to talk to the symbolic one.

Expression, he suspects, evolved as adversarial communication. Once you read faces and postures, others start signaling and subverting. At a funeral everyone makes a solemn face; that face does not say whether you are sad, only that you know the local protocol. You read the delta between the standard mask and what is moving behind it. Fridman has learned, doing a video podcast, that viewers use *his* face as a guide to how they should feel about the guest, then comment that he looked bored. He would rather wear a mask and sunglasses. Eastern European, especially Russian, culture taboos hyper-animated faces; in the United States you are expected to be animated and to show positive affect. Unprompted cheer in Russia reads as stupid and unsophisticated. A circulating diagram of the two grading curves: in the U.S., the bottom 10 percent is *a good start* and everything above is *amazing*; in Russia, almost everything is terrible, and the top percent is *even so*. Underneath the joke is a construction of meaning. America focuses on the positive and suppresses the rest. Eastern Europe treats existence as at best neutral: if you hold something above the waterline you must put something below it. Beauty is tied to suffering; not acknowledging that every conscious being spends most of the time suffering is, on that ethos, a kind of stupidity. Bach tells his daughter the older mermaid stories, Undine kissing a faithless prince to death, and still sides with Oscar Wilde: if the choice is the machine or your heart, you follow your heart.

Suffering, however, is not required for happiness. Suffering is caring about what you cannot change. Change what you care about to what you can change, and you will not suffer. Happiness itself is not important. It is like a cookie: as a child you want all of them; as an adult you see the cookie is a tool to get you to eat vegetables. Happiness is a cookie the brain bakes for itself. The environment cannot make you happy; your appraisal can, and if you can change the appraisal you can create arbitrary states of happiness. Some meditators find the basement bakery, stuff themselves for months, get bored, and then meet the meaning crisis they had misdiagnosed as a happiness shortage. They can release neurotransmitters at will; the original question returns: why live, and how to insert yourself into a sustainable civilization that means something. Bach's answer drops to the cell — a self-organizing molecular machine that needs a self-replicator, a negentropy extractor, and a Turing machine. Miss any part and it is not alive. Once you have that intelligent super-molecule there is little you cannot make it do; it is probably optimal computronium, especially for resilience. A planet infected with life is hard to sterilize. Think of the cell as a von Neumann probe: the way to seed intelligence on other planets is to infect them and wait. Technical design works from the outside in, in deterministic workshops. Biology works from the inside out: a seed subdues relatively unorganized matter. The purpose of life is to produce complexity, so you can harvest entropy gradients you could not harvest without it. The purpose of intelligence is control under complexity, pushing the boundary of order into chaos.

God, for Bach, has nothing to do with creation. That, he thinks, is a mistranslation of the Talmud into Catholic mythology. Genesis is the childhood memories of a God: a mind remembering how it came into being. Light and darkness, then sky and ground, plants and animals, names — the sequence every mind invents because those structures are not given; you have to make them to make sense of patterns on the retina. Watch a child and you see it. Even a nerd who racked a server and ran the world on it would have no magical power to give our existence meaning. Equating a creator with a source of meaning is a sleight of hand, as is equating Aristotle's prime mover with a supernatural being. The prime mover, in the simplest case, is an automaton. The higher being we actually perceive, the one we relate to as a cell relates to a body, is civilization. We have that prior because we evolved to organize in these structures. The Christian God in its natural form, mythology removed, is the platonic form of civilization: the idea you try to approximate when you deal with others based on what you think is right, not on your incentives. Fridman signs off on the cookie line. Bach leaves a last metaphor. If you take this as a computer game, this is the best level for humanity to play. And this best level happens to be the last level, against the backdrop of a dying world. It is still the best level.