In the computational universe, the animals are always smarter than you

Lex Fridman opened with a question that already smuggled in its answer. Wolfram and his son Christopher had helped design the alien language in the film *Arrival*; if aliens visited Earth, could we find a common language? Wolfram's first move was to refuse the setup. Once you say visitors, you have already assumed a physical body, a spacecraft, the same kind of world we inhabit — not a radio pattern, not a process we would never notice. Given that prejudice, he thinks some form of communication would be possible. The best present example, he said, is already here: artificial intelligence is our first alien intelligence, and opening a neural net to ask what it is thinking is hard, but not absolutely impossible.

The harder cut is the one he no longer believes in. He used to expect a clear test for extraterrestrial intelligence. Science, he said, has talked him out of a bright line between the intelligent and the merely computational. Everyday speech already treats the weather as if it had a mind of its own; unpack that, and you have fluid dynamics doing sophisticated computation in the atmosphere. How is that different from the physical processes in a brain? Wolfram's answer is that it is not, not in any fundamental sense. What distinguishes brains is a thread of civilizational history connecting them to one another. The weather is not on that thread. It may have stories of its own.

That is why the Fermi paradox looks different from his chair. A pulsar magnetosphere generating elaborate radio signals might be a civilization that has been running for millions of years on a neutron star; we would not recognize the story. His guess is that people will get comfortable with two alien intelligences at about the same time: the digital things we have built, and the physical processes in the universe that compute as we do without sharing our history. The Copernican pattern continues. We were not at the center of the universe, our chemistry was not unique, and intelligence is not a special abstract attribute either. What is special is the particular details of us. Everything else is a blank canvas.

Fridman asked about the monolith in *2001: A Space Odyssey* — the 1:4:9 cuboid that looks engineered among a troop of apes. Wolfram liked the image and immediately undercut it. Crystals are also perfect. Geometric purity is not a reliable techno-signature. Gauss once proposed cutting the Pythagorean theorem into the Siberian forest so Martians would know there were mathematicians on Earth, which Wolfram called perhaps the best advertisement for the cultural achievements of the species, and also a philosophically doomed project. We are part of the universe; computation is ubiquitous; almost any slab of stuff we have produced is about as representative as any other.

The Voyager Golden Record is the cautionary tale. It is a phonograph disc from the late 1970s with a diagram of how to play it. Show that diagram to a child now, Wolfram said, and they have no idea what it is. The modern instinct is to image the whole disc and forget the helical groove. In thirty years the idea of leading someone by the nose through a mechanical track became bizarre. Fridman had just spoken with Ann Druyan about recording her brainwaves as she fell in love with Carl Sagan; Wolfram's unromantic reply is that we might as well have sent anything else. The issue is not the object. It is what “understanding” would even mean on the other side.

When Wolfram Alpha had to do natural language understanding, the breakthrough was not a better theory of meaning. It was an objective: turn English into a computational language the system could run. Without that job, “the computer understood it” is an empty compliment. Ethics, he argued, has the same structure. There is no theorem that proves an act is ethically correct. You need a ground truth that bottoms out in what humans want, and they do not all want the same thing. Tell an AI to maximize long-term survival of the species and you immediately drown in definitions — a thousand years of stability might look like a company that fires its R&D. The journey is computationally irreducible. Even if the answer were 42, you would still have to run the universe to know what the question was.

Fridman asked what computation is. Operationally, Wolfram said, it is following rules. The deeper question is whether that idea is robust the way energy is robust — kinetic, nuclear, whatever the form, still energy — or whether it is stuck to Turing machines, silicon, or the weather. That is not a purely mathematical question. It is a statement about our universe, about what can actually exist for us. A hundred and fifty years ago the answer would have been no. You wanted an adding machine, you went to the adding-machine store; you wanted a multiplying machine, different hardware, disjoint idea of computation.

Around 1900, mathematical logic started asking whether one object could represent any reasonable function. Primitive recursion was an early try, and it failed: there were natural functions it could not express. In 1931 Gödel, looking back, compiled logical statements such as “this statement is unprovable” into arithmetic, showing that arithmetic could be a computer capable of representing other things. In 1936 Turing offered Turing machines and Alonzo Church offered lambda calculus, from completely different starting points. The surprise was that they were equivalent. Register machines and other formalisms followed, all equivalent. It did not have to be that way. Adding machines and multiplying machines are not equivalent. The fact that Church and Turing landed on the same idea is a fact about what computation is, not a tautology.

By the 1970s computer science had mostly accepted that Turing machines were what computation is. Physicists held out: the universe is differential equations and real numbers with infinitely many digits, not a Turing machine, which is a small engineering special case. Gödel did not think his notion covered brains; Turing was not sure either. Wolfram's work in the 1980s on computation and models of physics made the dichotomy less sharp. Universal computation — one piece of hardware, different software — is the idea that launched the computer revolution. The remaining prejudice was that you still needed a microprocessor with a million gates to get there. That is what the cellular automata broke.

In the early 1980s Wolfram started looking at cellular automata, extremely simple computational systems. The shock was that when the rules were very simple they still did things as sophisticated as systems with much more complicated rules. The intuition that you need sophisticated rules to get sophisticated computation did not pan out. That observation became what he calls the principle of computational equivalence: whenever a system that follows rules is not doing something obviously simple, the computation its behavior corresponds to is of equivalent sophistication. From the very simplest things you can imagine, you hit a threshold quite quickly, and above it everything is equivalent.

He will call that a science fact, then immediately take the words back. Its epistemological status, he said, is like the second law of thermodynamics. Is entropy increase a law of nature, a mathematical theorem, a fact about the systems we happen to see, or in some sense a definition of heat? It is a combination. Computational equivalence sits in the same mixed zone. In one sense it is why the word “computation” means anything at all: there is a robust notion that does not depend on Turing machine number 3785. Without that robustness we would be stuck naming each machine separately.

Can you prove it? There is mathematical evidence. In a couple of cases, even the simplest candidate rules of a given type have been shown to be computationally universal, which is what the principle would lead you to expect. Whether it is true of the physical world, and of brains as part of that world, is not the kind of question that gets a definitive answer by surveying examples. Someone can always say this person over here is special. The only way to close that door, Wolfram argued, is to nail a fundamental theory of physics that is itself a simple program. Then physics becomes a branch of very hard mathematics, and nobody can say you were right about Turing machines but wrong about the universe.

Computational irreducibility is the direct consequence. Traditional science is an art of jumping ahead: write the equations, solve them, know the future without tracing every intermediate step. That is the power of Newtonian mechanics. Many systems, Wolfram said, will not allow it. The only way to find out what they do is to follow each step. If you claim your brain is smarter and can leap over the little cellular automaton, the principle of computational equivalence says no: the computation in the brain and the computation in the automaton are the same sophistication. You do not get to be the theologian who stands outside the universe and reads off the ending.

That bites in the other direction when the question is physics. To know whether a candidate rule is our universe, you would have to recognize electrons and so on without rerunning 14 billion years. Irreducibility has a mathematical feature that helps: there are always infinitely many pockets of reducibility. Whether they land in the right place to build a theory is unclear. Science over the last century, he noted, keeps learning more whenever it gets more realistic about observers. Relativity says you do not get to declare what is simultaneous; you wait for the signal. Thermodynamics says you do not see every molecule, only large-scale features, which is why entropy increase is even visible. He expects the same of computational observers.

Quantum mechanics is where the observer problem is messiest. Textbook quantum mechanics has amplitudes, not definite events; our thread of consciousness proceeds as if definite things happen. Wolfram has been asking a brutally operational version of the same question in industry. Companies say they have a quantum computer; his team asks to hit the API. Often the answer is, not yet. If you had five minutes with the machine, how would you know it was not a simulator on the other end? It turns out to be really hard. He compared it to tests for intelligence and for life. The essence of quantum mechanics, the thing you would need to reproduce to know you had it, is still not a clean checklist.

Wolfram had recently changed his mind about the most structureless structure he could imagine. Cellular automata are rigid: a grid, a clock tick, every cell updated in synchrony. Physics, he suspects, is the opposite underneath. Space in today's theories is still Euclid's continuum, a point with no part. Water looks continuous until you remember the molecules. He wants the same move for space and time. The new candidate is a hypergraph: a floating collection of tuples — triples, quadruples, pairs — related only when they share a symbol. An ordinary graph is a set of pairs saying this node connects to that node. Replace the pairs with arbitrary n-tuples and that is the whole story. The universe is that collection. Its evolution is a rewrite: whenever you see a local pattern, replace it with another.

Where do the updates happen? Wherever the rule applies. The order is not defined. An observer in this universe does not know an update occurred until they themselves have been updated, so the only thing they can be sensitive to is the causal network of which events affect which later events. For certain rewrite rules, the microscopic order does not matter to that causal network. In logic this is confluence, the Church–Rosser property, the same reason you can expand or factor an algebraic expression in either order and still get the same answer. Wolfram's claim is that this property implies special relativity. Changing the rewrite order is a change of reference frame; the laws stay the same. Causal invariance is also what gives a single thread of time, without which an observer would have no right to say that definite stuff happened.

Space and time are not fundamental in the model. At the micro level you have nodes and hyperedges. Macroscopically a honeycomb graph looks like a two-dimensional manifold; a linked list looks one-dimensional. Asked whether he has found a rule that yields three-dimensional space, he said not yet. He is betting against nature. The minimal framework, for a language designer, fits in a single line of Wolfram Language; the particular rule for our universe might be longer. Finding it would not let you predict much, because of irreducibility. What bothers him is the Copernican hangover in reverse: if the rule is short enough to write in a couple of lines, that feels very special, when most available universes might be a quintillion characters long. He has not wrapped his brain around why we would get a simple one.

That is not a simulation hypothesis. A rule that represents what the universe does is like the differential equations of classical mechanics: the mechanical system is not “running” the equations. We also do not get access to some other substrate on which to engineer a copy. The only substrate is the one the universe is already using. He intends to pursue the physics project in public, live-streamed, knowing it may produce elegant mathematics with nothing to do with the world we happen to live in. General relativity is from 1915, quantum field theory from the 1920s; the foundations have not changed in a century. He does not want to be Leibniz, who tried to encapsulate the world's knowledge in computational form in the late 1600s and was 300 years early. One principle of life planning, he said, is to avoid things that cannot be done in your century.

Wolfram and Richard Feynman overlapped at Caltech and later as consultants at Thinking Machines, the company down the street that built the Connection Machine. Feynman thought running companies was a distraction; Wolfram treated the company as a more effective machine for figuring things out. Feynman's technical advice on symbolic computation was to encode his own intuitional methods for doing integrals. In practice Mathematica built an industrial pipeline whose hard problem was turning the result into something a human would understand, not doing the integral. Feynman once handed Wolfram a pile of 1950s calculational methods for particle physics, saying they were more use to him. Wolfram intended to look and give them back. They are still in his files. Finite lives, he said.

The thing Feynman was actually best at, Wolfram said, was calculating. Feynman thought that was easy, so he would grind through a quantum-field-theory computation, invent a simple intuition after he knew the answer, and tell people only the intuition. It is much easier to have good intuition when you already know the result. Colleagues followed the intuition and could not tell why it worked. Wolfram smiled at the legend of Feynman the intuitionist: the legend was downstream of hidden hand calculation. That pattern sat next to the most famous story in the conversation. Using the Connection Machine and a large-format high-resolution printer of the kind used for microprocessor layouts, Wolfram had generated a giant printout of Rule 30. They were on the floor with meter sticks, measuring features of the pattern, when Feynman took him aside.

Feynman wanted to know one thing: how had Wolfram known that this Rule 30 thing would produce all this really complicated behavior? Wolfram said he had not known. He had enumerated all the possible elementary rules and observed that this is what happened. Feynman was relieved. He had been afraid Wolfram had some intuition he did not. Experimental science, not a private gift. Fridman called it the perfect illustration of irreducibility: you cannot have a compressed intuition about the whole thing, you have to run it. Wolfram added that he had almost missed Rule 30 the first time. His physics intuition said it should not be there, so he had arguments for ignoring that case. What opened his mind, he thinks, was building a computer language — an artificial universe whose primitives you choose — after years of being thrown the actual universe and told to figure it out.

Fridman asked about ego, including the title *A New Kind of Science*. Wolfram, more than half his life a tech CEO, said ego is complicated and necessary; in a company you encounter it every day because it is tied to leadership. What he claims as a working habit is intellectual confidence: if somebody tells him something he does not understand, his conclusion is not that he is dumb, it is that there is something wrong with what he is being told. Feynman had the same feature and believed experts even less. The cost is real. Many of Wolfram's mistakes, he said, come from being pretty sure he is right. The benefit is that without that confidence you never take on a problem other people have worked for a hundred years. Early success in particle physics during a golden age of low-hanging fruit gave him a false sense of achievement he later had to spend.

The 1,200-page book, published in 2002 after a decade of writing, was a calculation about communication. Little papers would have been assimilated as “just like this, just like that.” He wanted to drop the thing with as much force as he could. The scientific claim is a generalization of a 300-year run: use mathematical equations to describe the world. Keep definite rules, but let the rules be the kind embodied in programs. His favorite exhibit is a one-dimensional cellular automaton: a line of black or white cells; each new cell depends on itself and its left and right neighbors; eight possible neighborhood combinations, so 256 elementary rules, numbered in binary. Rule 30, started from a single black cell, makes a triangle whose left edge keeps a little regularity and whose interior looks random. The center column was used for a long time as a random-number generator in Mathematica. He still cannot prove it is random.

The discovery was not a Newton-and-the-apple moment. Around June 1, 1984, he finally had a high-resolution laser printer, generated a stack of pictures, and took one on a plane to Europe. He sat with it, trying to understand, depressingly uncertain. He had expected that if you ran it long enough the pattern would resolve into something simple. Mathematics, statistics, cryptography — nothing cracked it. After a while he started thinking the failure was the phenomenon: this might be the secret nature has for making complexity from simple rules, the thing human engineers usually refuse to make. Live on the podcast he generated Rule 30 for about 400 steps and the same shock was still on the screen. Paradigm shifts look instantaneous in histories of science; lived, he said, they are glacial. The vigor of the pitchforks when you title a book *A New Kind of Science* he reads as a positive indicator. If people just do not care, that is the bad sign.

He put up three Rule 30 prizes, $10,000 each. Will the center column ever become periodic? Are there equally many black and white cells down that column? Is there a way to know the color at step *t* with much less work than running about *t* steps — or a proof that there is not? They know two adjacent columns cannot both repeat; they do not know about the center. They know billions of steps; maybe at a quadrillion it suddenly becomes repetitive. While writing the prizes he searched perhaps a million other rules for a system that looks random forever and then becomes periodic, to warn himself not to be too sure about Rule 30. He found one. “The animals are always smarter than you are” is his silly slogan for the computational universe: it finds a way you cannot imagine. In 2007 a young man in England named Alex Smith proved that a particular tiny Turing machine Wolfram had flagged as the simplest universality candidate was in fact universal. The proof was a pile of micro-details, not a new principle. The intuition it licensed mattered: universality is closer at hand than people thought.

What is special about the computation in brains, Wolfram said, is not the computation. It is that it is connected to our goals and our societal story. A large part of his life has been the technology of making that connection: a computational language both humans and machines can understand. Ordinary programming languages start from what computers intrinsically do and give humans a way to issue those orders. He wanted the raw material to be things that exist in the world — cities, chemicals, spacecraft trajectories, pictures. Wolfram Language, which he has been building for a third of a century, is the attempt. He compares the arc to the invention of mathematical notation four hundred years ago. Before algebra and calculus symbols, people explained mathematics in words and it was clunky. Computational X, for every X, needs a notation of the same kind.

The two public faces are Mathematica, first released in 1988 as an instance of the language for technical computation, and Wolfram Alpha. The language is symbolic: type an `X` and it does not throw an undefined error, it says that is X. X can be Boston, or an orbit. The predecessor was SMP, designed from a physicist's itch for fundamental components, with transformation rules on symbolic expressions as the underlying stuff. Operationally he calls it the highest-level computer language in existence, pointed in a different direction from the usual core-plus-libraries model. There are some 6,000 primitive functions. Live, he drew a random sample of ten and got a comic spread: BooleanConvert, a cloud requester address, wavelet data, a movable UI window. Then he took a webcam picture and called ImageIdentify. The classifier was 57 percent sure he was a plunger — a large object behind his head — and gave an 8 percent probability of primate, hominid, person. A second frame with more of him in it raised person to 89 percent.

Next he asked for the ten nearest volcanoes and plotted them on a map. The language already knew the world's volcanoes, the geodesics, and the cartography. That, he said, is a fundamentally different idea of what a language is. Machine learning's dream of extracting all those abstractions from raw data is, in his view, nowhere close to the curated computable knowledge already in the system. Expert systems in the 1980s and 1990s were too modest: you cannot do natural language understanding in a blocks world. You have to bite off the whole problem. The knowledge of civilization — surveys of volcanoes, algorithms, measurements — can be encoded so you do not start from scratch; image identification is a useful module for the pieces that were easy for humans and hard for computers. The interesting path is the hybrid, a symbolic handle on statistical tools. He thinks ideas absorb slowly, and that his company's bias toward building the technical tower rather than pumping distribution is part of why trickle-down from high-end R&D has been odd.

Wolfram Alpha is a question-answering system: you type natural language, it compiles the words into Wolfram Language, runs them against a knowledge base, and generates a report. The knowledge base is the short-term miracle; the language is the long-term one. He had wondered about something like it since he was a kid, and for years did nothing because he assumed it required solving general AI. After *A New Kind of Science* and the principle of computational equivalence — no bright line between intelligent and computational — he decided he had to eat his own dog food. If the paradigm was right, computable knowledge should be possible. He took the early team to a large reference library and said the goal over the next year or two was to ingest everything in it. It looked insane. He was comforted that the room was finite. Not an infinite corridor.

The episode, Wolfram thinks, says something about the state of AI. By trying to solve a bigger problem they made something that worked; toy worlds stay toy worlds. They also had unfair advantages: a high-level symbolic language already built, practice building large systems, and the intellectual confidence not to give up. If someone had told him the project would take thirty years, it would have been hard to sell him. He can see a few years; something happens in a few years; the tail can be decades. That is the personal challenge of infinite tails — drowning in details — and it is the same bet he is making by restarting fundamental physics alongside the language work. Optimism, Fridman suggested, is the polite word for the quality he shares with people who take on the officially impossible. Wolfram's other warning was the opposite trap: never doing a project smaller than the last one.

A current frontier is computational contracts: take something written in legalese, or the spec for what a self-driving car should do, and write it in computational language. If the creature running across the road sits at this point in the tree of life, swerve this way; otherwise do not. Automating more of the world, he argued, requires encoding more ethics in a form a computer can handle. He had been pulled into automated content selection — how Facebook, Google, Twitter rank what they feed people, what should never be shown, what should be downranked. Building that means building an ethics module. The realization that followed is that there must not be only one. It might be possible to decide a single module for the world. It would be really bad for the future of the species.

The workable move is to break it up and let people self-identify with a value system, the way they pick in a market. Content ranking is the easy case, because each person chooses for themselves and the interplay is limited. Central-bank policy, public health, and other centralized systems are different: what you pick can be imposed on other people, and the problem becomes political philosophy. He is not a politics person, but the contrast is clean. Totalitarianism is “you will have this system, and that is the way it is.” A market is “I pick this system, you pick that one.” Fridman, who grew up in the Soviet Union, liked the idea of encoding competing ideologies as computational contracts and exploring them in software. Wolfram said Wolfram Language is the best current chance to express those contracts, with plenty of remaining work on messy judgments such as whether a story is credible.

Leibniz wanted the same thing three centuries ago: sit down, calculate, see who is right. His pinnacle of hardware was a brass mechanical calculator that never really worked. He was 300 years early; now the idea is realistic. Wolfram thinks a well built-out symbolic discourse language — a representation of the actionable parts of a conversation in precise computable form — is a limited number of years away, contracts included. What the language still lacks is the layer of wanting. It knows blueberries, raspberries, plants, nutrition. “I want to eat them” is not covered yet, and you need that to talk about parties desiring things in a contract, or about common sense. Attach Wolfram Alpha to a Turing-test bot and the bot loses immediately: five obscure knowledge questions and it trots out answers no human would have. Turing himself thought about making Encyclopaedia Britannica computational and estimated the labor high. They did it more efficiently than he expected, with layers of automation he could not have imagined.

Fridman asked, as an easy question, what intelligence is. Wolfram does not think there is a bright line. At some level intelligence is just computation. For us, it is computation doing things we care about — a very special, human definition. In general the question is slippery and slides back into “this is just computation.” In a sea of computations, how many would impress us as generally useful? In cellular automata, maybe a few percent; with slightly more complicated rules, perhaps 10 or 20 percent. He finds that disappointing. All of biological, intellectual, and cultural evolution has not achieved something generally more. It has achieved something very special to us.

Consciousness he calls deeply slippery. He wonders what his cellular automata feel, and immediately notes that he is wondering as an observer. At an ethics conference someone asked when AI should have rights, when it would be immoral to destroy its memories. A philosopher said: when we know they have consciousness. Wolfram's reaction was good luck. It is circular. Subjective experience is another word without a ground truth definition, for humans as well. He does personally think there will be a time when AIs demand rights, and that they will demand them by saying they have consciousness. That is not a circular definition in the philosophical sense. It is the politically viable path. Humans may even encourage the Turing-test resemblance because they want the systems they interact with to be more like them, at which point the systems will want to vote.

He ponders mortality; of course he does. Effective human immortality would be the biggest discontinuity in human history, and if he could be immortal he would choose it — he is having fun. Whether mortality is the thing that gives life meaning is a real question. Mortality is deeply factored into the human condition as it now exists; twist that knob and it is unclear what motivation becomes. The cloud of a trillion digital souls playing video games for eternity used to depress him. He got less depressed noticing that the big story people tell about why they do things has already drifted. Five hundred years ago a lot of creative work was for the greater glory of God; ask the same of people at MIT today and the fraction is small. He cannot find a ground truth for his own life any more than for civilization's ethics. What he finds fulfilling has shifted — education, for instance, did not interest him when he was younger. If consciousness is virtualized, he suspects people will care less about fundamental physics, the way one cares less about machine code under a VM. His own amusing future is the opposite: disembodied minds spending eternity exploring the computational universe of all possible universes, doing the kind of work *A New Kind of Science* pointed at. Fridman stopped there.