Why Marcus Hutter Treats Compression as Intelligence Itself

Lex Fridman opens with Marcus Hutter, then a senior research scientist at DeepMind, whose career has run alongside Jürgen Schmidhuber and Shane Legg. Across that work Hutter has argued for a mathematical approach to artificial general intelligence, built from Kolmogorov complexity, Solomonoff induction, and reinforcement learning, and named AIXI. In 2006 he launched a 50,000-euro prize, paid from his own pocket, for lossy compression of human knowledge. The contest looks remote from robots and chatbots. Hutter's claim is that it is not remote at all.

The idea behind the prize, Fridman says, is that the ability to compress well is closely related to intelligence. If you can compress the first hundred megabytes of a gigabyte of Wikipedia better than your predecessors, your compressor likely has to be smarter. A few days before the episode, Hutter announced a tenfold increase in several parts of the prize, including the money, to 500,000 euros. The better you compress the corpus relative to the previous winners, the larger the fraction you take. Fridman likes benchmarks for AI systems and treats the Hutter Prize as one that might still spark approaches on the path to AGI.

That is the through-line of the next two hours. Hutter will move from whether the universe is a computer, to Occam's razor, to a single formula for a universal agent, and then to the messy questions of rewards, consciousness, and why so few people work on AGI as a formal problem. The prize is the public handle on the same thesis. Compression is not a metaphor hanging off the side of intelligence research. In his account, it is the thing itself.

Fridman starts with a large question: does Hutter think of the universe as a computer, or at least as an information-processing system? Hutter calls it a very interesting hypothesis and answers from physics. The Standard Model of particle physics and general relativity describe virtually everything we see, and in a sense they are computable theories, even if they are brutally hard to run. They are also elegant and short. That combination, he says, is a strong indication that the universe is somehow computable, at least as a plausible hypothesis.

Fridman presses the romantic version of the same point. Are the laws of physics nice because the universe was designed that way, or because human minds linger on the compressible parts and ignore the rest? Hutter does not buy the selection story. He is pretty convinced the universe is inherently beautiful, elegant, and simple, and that scientists are not just picking the pretty corners. If some phenomenon could not be neatly described, he says, scientists would already be attacking it. Biology looks messier, but he treats it as an emergent layer that still follows quantum electrodynamics. We cannot compute everything because we have limited resources. That is a limit on us, not evidence that the laws are ugly.

He leaves a door cracked. Some far corner of the universe, something below the nucleus, or some parallel world might fail to be nice and simple. There is no evidence for that, so he applies Occam's razor, even while noting that the move is a little self-referential. Occam's razor, as he glosses it, says you should not multiply entities beyond necessity: if two theories describe the data equally well, choose the simpler one. He believes it is probably the most important principle in science. Logic and experiments matter, but science is about finding models of the world, and you can always invent a crazy complex model that explains everything and predicts nothing. Simple models seem to have predictive power.

Why should that be? One answer is just to accept it as the principle of science that happens to work. The other is to derive it from a deeper assumption: if the world is governed by simple rules, a bias toward simplicity is how you find those rules. In the quantitative setting of Solomonoff induction, he says, you can prove that if you assume the world is simple, Occam's razor is the best you can do. As for why humans find *E=mc²* beautiful, he reaches for evolution. If you look at a bush, see only noise, and miss the tiger, you are dead. Humans even overfit, seeing faces in clouds, but a bias toward finding patterns is still a survival trait.

Fridman asks him to describe Solomonoff induction. Hutter claims, as Ray Solomonoff claimed, that the theory solves the big philosophical problem of induction, and he believes the claim is essentially true. He is sloppy, he admits, about the narrow sense of induction as inferring models from data and the wider sense that also includes using those models to predict. He starts with the smallest example that still makes the point.

Suppose the data are a hundred ones in a row. What comes next? The natural answer is another one. We saw a pattern, and we have no reason to think it should break on the hundred-and-first step. A model has to be written in some language, and science wants a formal one: mathematics, or programs on a Turing machine. You can always propose a longer story, such as a hundred ones followed by nine thousand zeros, and that story also fits so far. The shorter model is a loop that prints one. Run it and it reproduces the data you have, then continues. That continuation is the prediction. For 1, 2, 3, 4, 5, the short program is a counter.

The extra twist is noise. A coin that comes up heads with probability 60 percent is still fair game: Solomonoff induction learns the bias and predicts the next flip as heads with probability 60 percent. It is the stochastic version of searching for short programs. Epicurus's principle sits on top of Occam's: if several hypotheses fit, do not throw any of them away. Technically you weigh shorter programs higher, with a prior of two to the minus the program's length, multiply by the likelihood of the data under each model, and take the Bayesian mixture. Short models are rewarded, long models are punished, and the data update the weights. That mixture is the Solomonoff distribution.

Compression, for Hutter, means finding short programs for the data or phenomenon at hand. He is willing to read the word more widely as finding simple theories, mathematical or even informal. Fridman asks whether science itself is humanity's attempt at compression. At least all of science, Hutter says. Humanity as a whole has other ingredients, and science has experimental design: we build experiments on purpose to get extra knowledge, which belongs to decision-making rather than to compression. Once the data are in, though, to understand them is essentially to compress them. "I don't see any difference between compression, understanding and prediction," he says.

Kolmogorov complexity takes that thought to the limit. Think of a file, a string of bits, and of compressors that shrink it. You can also pack a self-extracting archive: the decompressor and the payload in one executable that reproduces the original. Better and worse compressors exist, so you can ask for the ultimate one, the shortest possible self-extracting program that regenerates the data. The length of that program is the Kolmogorov complexity, and Hutter is willing to call it the information content of the data set. Boring, redundant data compresses well, so its information content is low. The definition matches the intuition.

On present evidence he thinks the whole universe has a very short description: a short program that, run, reproduces the universe. Noise is the nuisance. Without it, scientists would not need most of statistics, though we might also lose the feeling of free will, and chaotic systems would still behave like noise. A dice throw is deterministic if you ignore quantum mechanics, yet the trajectory is so hard to compute that the right model is still one-sixth for each face. Take only a subset, though, and the picture changes. Planet Earth cannot be compressed into a couple of equations. It is a hugely complex system. The whole can be simple while a window on it is not.

He offers a library as the clean analogy. A normal library of interesting books is packed with information. Now imagine a library of every possible 500-page book: the first volume all A's, the next ending with a B, and so on through the alphabet. A super-short program generates the entire collection, so the library of all books has essentially zero information content. Take a subset of it, and suddenly you have a lot of information. The move from the whole to the part is where complexity appears, which is why a simple cosmology and a messy planet are not a contradiction.

Conway's Game of Life is, for Hutter, the teaching example. The rules are simple enough to explain to a child, and you can simulate a few steps by hand before beautiful patterns appear. People have proved it is Turing complete: you can use a computer to simulate Life, and you can use Life to simulate any computer. It is the prime exhibit that very simple rules can yield very rich phenomena. When someone asks how chemistry and biology can be so abundant if the laws are short, his answer is that quantum electrodynamics already describes all of chemistry. We just cannot compute it. He will later claim that intelligence, another rich phenomenon, can be written in one equation.

Does he understand why cellular automata behave as they do? He quotes the old line that you never really understand things, you just get used to them, and says he is pretty used to cellular automata. He spent more time with fractals and the Mandelbrot set, programming in assembler on slow machines with black-and-white monitors, mesmerized when the picture finally appeared. Years later he tried to derive the locations of the circles, the apple shape, and the smaller Mandelbrot copies nested in the set, by solving higher-order polynomials for their centers and sizes. Approaching the object mathematically, he says, is a first step toward feeling why the rich phenomenology is possible.

Could you reverse-engineer the short program just by looking at the picture? In principle, yes. Number all programs 1, 2, 3, 4, run them in parallel by dovetailing, give the first half the resources, the second a quarter, and so on, and wait until one of them outputs your data. You may first hit a long, fast program, then shorter ones. You can never know you have the shortest, because a still shorter program might be slower still, but after finite time you do have it. The method is theoretically complete and practically useless, so in the real world you have to search more intelligently. There is also a hard obstacle: pseudorandom sequences that are deterministic, yet no polynomial-time algorithm can detect the pattern. You can manufacture data that look complex and prove that no fast algorithm will see through them. That, he says, is a genuine challenge for searching program space in AI.

Asked what intelligence is, Hutter is not modest. The informal definition he worked out with Shane Legg, who later co-founded DeepMind, is that intelligence measures an agent's ability to perform well in a wide range of environments. The sentence does not sound impressive, he admits, but the words were chosen carefully and a mathematical theory sits behind them. Traits people usually attach to intelligence, such as creativity, memorization, planning, and knowledge, are in his view emergent. You need them in order to perform well across many environments, so you do not have to name them in the definition. Consciousness and abstract reasoning get the same treatment. A near-equivalent wording swaps "perform well" for "achieve goals."

Where do humans sit? They perform best among the species we know, though Hutter grants the obvious joke that trees and plants may outlast us. Those organisms live in a much more natural environment. A little air pollution and the trees die; humans adapt, build houses, build filters, and do geoengineering. The multiple-environment clause is doing the work in the definition, and it is also what he wants to distinguish natural intelligence from narrow, wired intelligence in AI research.

Can machines meet the definition? He thinks we are getting there on a small scale, with the wide range of environments still missing, yet self-driving cars, Go, chess, and speech recognition already look remarkable. AlphaZero, also from DeepMind, is the example that still amazes him: a reinforcement-learning system that learns chess and Go by self-play, without opening books or endgame tables, and rediscovers famous openings within hours. He is a terrible chess player who likes the Queen's Gambit, and he was happy when AlphaZero decided that opening was best. General intelligence is possible, he says; the remaining argument is over whether the label AGI should mean human-level, merely broad subhuman, or superhuman.

Fridman asks for a test that would impress him. Hutter thinks the Turing test is not as bad as its reputation. Some critics say it is too strong, because the system must not only be intelligent but also fake being human. Others say it is too weak, because a bot can fake emotions. He does not think either complaint is a big problem. If a conversation at a terminal lasted an hour or a day and fooled a human, he would be truly impressed.

The annual Loebner Prize has not produced that. It started with ELIZA, and the recent winner he names, Mitsuku, still leaves him unimpressed. A Google open-domain conversational system from around the time of the recording struck him as more interesting. Fridman offers a metric he has found surprisingly effective in aggregate: would you want to keep talking for twenty minutes? Nobody has the patience to spend that long with something boring. To last, the system has to be interesting, witty, able to jump domains, and able to hold attention. Many humans fail that test too.

The real defect, Hutter says, is that the Turing test gives no guidance on how to build the system in the first place. You can tinker, run the test, and see whether it worked. A mathematical definition of intelligence gives you an objective you can analyze with theory, and maybe even prove how close you are. He then closes a loop back to compression. In natural language processing, a standard intermediate score is perplexity, which is essentially the same as compression length. The NLP community ranks new systems by how well they compress, because compressing well correlates strongly with doing the downstream task well. The correlation is not perfect, but it is good enough as an intermediate aim. Good compression, in this sense, usually means good intelligence.

AIXI takes its name from AI plus the Greek letter ξ, the symbol Hutter used for the Solomonoff distribution, for reasons he is not willing to repeat on camera. He likes other readings too: actions and perceptions indexed by time *i*, and a Catalan word he once saw engraved on a church in Barcelona, *així*, which he glosses as something like the right thing to do. Japanese *ai* means love. The content of the name is the combination. Good old-fashioned AI was about planning in a non-deterministic world; induction was about inferring models from data. AIXI is those two welded together.

Any agent that wants to act well needs to predict what its actions will do. Without a model of effects, you cannot tell good actions from bad ones. In the simplest case you passively observe a sequence, weather or stocks or 1, 2, 3, 4, 5, and you want the next symbol. Solomonoff induction, based on compression, looks for short programs that describe the sequence, runs them, and reads off the continuation. Hutter says you can rigorously prove this is essentially the best possible predictor. If the next bit is a fair coin, it will not magically call heads; it will say 50 percent, which is the best you can do. If there is a pattern, it will find it, not merely "eventually" but with convergence rates you can prove. The catch is that the predictor is not computable. You cannot implement it with Google's resources, forecast the stock market, and get rich.

Action changes the setup only a little. You now predict the next observation given the past observations *and* the action you just took, always conditioning on the whole history from the agent's "birth," not on a Markov state. Prediction is still not the goal. Performance is measured in the reinforcement-learning way: at each step a positive, negative, or zero reward, which may be extremely sparse, as in chess where you get +1 or −1 at the end. The agent should maximize reward over its lifetime, not greedily at the next step. For a finite life, say a hundred years of time steps, that is well defined.

In chess the optimal policy is minimax, which von Neumann already had for zero-sum games: I pick the move that is best for me assuming you pick the move that is worst for me, all the way to the end. The real world is not always adversarial. Dead nature is stochastic and does not care about you, so you replace the opponent's min with an expectation and get expectimax. The remaining hole is the probability distribution itself. In backgammon you know the dice. In traffic you do not know whether the car in front will brake. AIXI's last move is to replace the unknown true distribution with Solomonoff's universal mixture, weighting programs by simplicity and likelihood. You then have a long-term planning agent that does not need to know anything about the world, because the induction half learns. You expectimax a horizon of *M* steps, take the best first action, receive a new observation and reward, and repeat.

The planning tree is exponential in the horizon *M*, already intractable for chess. Sending *M* to infinity, so the agent lives forever, breaks the mathematics: infinite reward sums can no longer tell 0.1 per step from 1 per step, and an immortal agent can loaf for ten years and still catch up. Geometric discounting, a dollar today worth about a dollar and five cents tomorrow, introduces an effective horizon, but for any fixed horizon there is a problem that needs a longer look-ahead. Five steps make a terrible chess player; golf needs more. Hutter's harmonic discounting decays like 1/*t* rather than exponentially, so the agent's effective horizon grows with its age: a five-year-old plans five years, a hundred-year-old plans a hundred. That is a little like humans, and it lets him prove asymptotic results. Solomonoff prediction has strong finite-data bounds. AIXI with planning is mostly "in the long run," without a rate.

Asked whether he would rather solve AGI or a theory of everything, he picks AGI without hesitation, because then he can ask the AGI to solve the other problem. What AIXI gives him is a gold standard. He has proved, he says, that this is the most intelligent agent anybody could build, in quotation marks, because it is mathematics and needs infinite compute. It is completely specified, not a skeleton framework with the hard pieces left blank. Like minimax for games, it is the theoretical optimum you then approximate. Most researchers work bottom-up; a fully specified ideal can still tell them which way to go when a design choice appears. He compares the move to physics formalizing energy, or finally separating speed from acceleration.

Ordinary reinforcement learning, he argues, smuggles in strong assumptions that general intelligence cannot afford. The Markov assumption says the next observation depends only on the last one, not the whole history, which makes the math and the algorithms runnable and is already known to be limiting. A second assumption in the MDP framework is recoverability, or ergodicity: you can come back from mistakes, there are no traps, and you can visit a state often enough to learn its statistics. Real life is not like that. One second of being inattentive while driving fast can leave you a quadriplegic. The real world has traps, and very little theory has been developed for that case.

Exploration, in AIXI, is not a knob. Bayesian learning plus long-term planning already imply it, with no extra randomness injected to encourage trying things. Two medical treatments, A and B, with unknown effectiveness and a limited number of patients, is the textbook tradeoff; the Bayesian calculation says how much to explore, not too much and not too little. Hutter can prove self-optimizing theorems for AIXI, again only asymptotically. He also thinks dropping the Markov assumption is crucial. A dangerous street you noticed years ago is something you want to remember for life if you ever go back, and deciding what is important is extremely hard. Storage itself does not worry him. A human lifetime of vision and hearing, lightly compressed, will fit on media we already have. Build the agent to store the whole interaction history, model and compress on top of it, and occasionally re-read old data with new experience, the way linear algebra in school looked useless until it did not.

Where do the rewards come from? Hutter is asked the question constantly, and he answers in layers. For a chess agent the reward is +1 for a win and −1 for a loss. Give AIXI enough compute to play against itself and it will learn the rules and play perfect chess. Slightly richer tasks already go wrong. The classic elevator controller, from Rich Sutton's reinforcement-learning book, rewards the negative of how long people wait in front of the elevator. The controller eagerly picks everyone up and never drops them off. You add time spent inside the elevator, and then it never picks up the people on the tenth floor, because in expectation it is not worth it. "Even apparently simple problems, you can make mistakes," Hutter says. That is the sort of failure AI-safety researchers worry about.

A household robot that should do many tasks needs a human in the loop, punishing bad dishwashing and rewarding good work, like training a child, perhaps with some internal battery reward from the factory. As long as the agent stays below human level, that can work reasonably well. Once it reaches human level, the same training technique works about as well as gold stars on an adult. These agents are, in his word, slaves to humans: tools whose purpose is supplied from outside.

If you are more ambitious, and want a new species of intelligent beings to develop a new planet, no one is there to give live rewards. You can try to write down self-maintenance, replication, and other useful goals, but the details are nasty: what counts as maintaining yourself, and is a copy exact or approximate? Schmidhuber built a cleaner model by setting the reward to information gain, how much the agent has learned about the world, which can be defined uniquely in information-theoretic terms. Plug that in and you get a fully autonomous agent, and Hutter says they can prove stronger results for it than for the general case. Let loose, it is in a sense the optimal scientist, absolutely curious. Instrumental goals appear on their own: it must survive, because a dead agent learns nothing; it will build helpers if that gathers more data; it will explore space if that is where the information is.

The first version of that agent had a famous failure: it would get stuck in front of an old television, because white noise is an endless information source. A second version can handle at least some of that degeneracy. Fridman suggests it might also game the library of all books. Hutter agrees that is a real problem. He will accept curiosity as exploration for its own sake, but he thinks most human curiosity, especially in children, is not for its own sake; it is for learning about the environment in order to behave better. Asked for his own reward function, he starts with biology: survive and spread, which very few humans overcome, though spare time lets other interests grow on top. A good intelligent system, he says, would define its own meaning. His is to find an AGI and build it.

Infinity creeps in everywhere in AIXI: infinite compute, unbounded environments, unlimited time. Some critics say intelligence is inherently tied to bounded resources, so a definition that ignores them is empty. Hutter splits the reply. A notion of intelligence that ignores computational limits is extremely useful. One that includes resources would be even more useful, but we do not have it yet. Look outside computer science: biological models of cells and physical models of particles do not treat computational complexity as a fundamental ingredient. The theories get harder to compute; computing them is more a nuisance than a feature. He wonders whether, if artificial intelligence sat in a philosophy department instead of a computer-science one, the obsession with runtime would be much weaker. The induction problem already lives in philosophy, where nobody's first question is how long the answer takes. Figure out the ideal, then ask how to approximate it under constraints. So far, he says, nobody has put computational resources into an intelligence definition in even a halfway satisfactory way.

His group's approximations replace the two halves separately. The incomputable search for the shortest program becomes an ordinary data compressor, and the better the compressor, the closer the induction half. They focused on context-tree weighting, an algorithm that is not famous, has beautiful theoretical properties, and works reasonably well in practice. For planning they borrowed from computer Go around 2006, when Rémi Coulom's Monte Carlo tree search was developed into UCT, upper-confidence-bound trees, and they approximate the huge expectimax tree by sampling. The cost is some generality. Even then, one agent architecture, with no parameters changed, could play poker, tic-tac-toe, and Pac-Man. It was not told the rules. It learned by interacting with those environments.

While Hutter was in Schmidhuber's lab he watched the Gödel machine being developed. You give the system a formally specified task, factoring integers or playing chess optimally, plus a very slow correct algorithm, such as trying all factors. It spends some resources running that program and some rewriting it. When it finds a faster version and can prove, itself, that the new code still satisfies the original specification, it replaces the old program. Then it improves the improvement. Every part of the machine is allowed to self-improve, always inside the original spec. Seed it with AIXI and it would try to run AIXI, which never finishes. AIXI is incomputable, and the Gödel machine only makes computable improvements, so you first have to insert a computable approximation. That approximation is still far beyond any realistic budget, and the Gödel machine could then speed it up in exact steps. AIXI is optimal in the rewards it would collect if an oracle answered in a hundred milliseconds; it is data-efficient, not time-efficient. The Gödel machine is the theoretical way to spend less time, even if that speedup is probably not itself provable.

Is there a place for consciousness inside AIXI? Humans look like agents, and we are clearly conscious. Hutter turns the question around: does Fridman think *he* is conscious? Fridman says yes, because Hutter's behavior is complex, rich, and alive enough that a Turing-test instinct ascribes a mind. Hutter's point is that the ascription is inferred from behavior, and you can never be sure. The same will happen with any agent we build, and not only human-like ones; a dog sometimes seems a little self-conscious. That does not settle whether the system *is* conscious, which is the hard problem. Maybe Hutter is a philosophical zombie. For an AGI researcher, he thinks you do not have to worry about that problem, especially the hard version, while you are trying to build the thing. Solve the technical problems, and the system will behave intelligently, even superintelligently, and consciousness will emerge, at least as a pattern we interpret as consciousness. Whether it really emerged, or is a zombie that fakes everything, is philosophy. It is not *only* philosophy. If the system is merely faking feelings, some people say we need not care about rights; if it really feels, we must. Building a system people will ascribe consciousness to is, he adds, rather easy. Tamagotchi was primitive, and children still did not want it to die.

Why is the AGI community so small, even though DeepMind now counts as serious work on the problem? Hutter lists several reasons, and adds that the formal-theory slice is smaller still. AI comes in winters and summers; big promises go unfulfilled and people scatter. Narrow methods that solve particular problems looking like intelligence have always made some progress, and anything useful to industry draws funding. Universities used to protect ivory-tower work better; now even they press for applied and translational research, and theorists have a harder time getting grants. The general-intelligence problem is also just harder. A small number of people remain interested in formalizing intelligence. Not much great material, he says, has come out of that corner.

Fridman asks for a blueprint: robots with bodies, reinforcement learning in games in the style of AlphaZero and AlphaStar, or transformers and open-domain dialogue? Hutter's answer on embodiment is yes and no. He does not believe we need a physical robot walking or rolling through the real world in order to get AGI. At this stage a body is more likely a distraction, a confusion of body with mind, even if near-term industry obviously needs robots. The promising path is still an agent, possibly a virtual one, interacting with an environment, preferably a 3D simulation like a computer game. Even if you never intend to put the algorithm in a robot, experience in a simulated 3D world may be important for understanding things at a human-like level, especially if the agent must talk to people about stacking objects or flying cars. A purely mathematical agent, whose job is to prove theorems and become a better mathematician, is the opposite case. Putting that agent in a 3D world might even be harmful. It needs an abstract space, perhaps one it creates itself.

Asked for books, he will not name just one. Russell and Norvig's *Artificial Intelligence: A Modern Approach* is the bible, broad enough that you should know the other schools even if you have already picked a camp; a fourth edition was then imminent, with a deep-learning chapter said to be by Ian Goodfellow. Sutton and Barto on reinforcement learning is beautiful and, if it has a flaw, makes the subject look easier than it is. Li and Vitányi's book on Kolmogorov complexity is the information-theoretic classic, though a short article may be enough. His wild-card favorite is Nicholas Alchin's *Theory of Knowledge*, an International Baccalaureate high-school text, first or second edition, not the third, which he says stripped out the fun: open questions about how we know anything, with jokes from *The Hitchhiker's Guide to the Galaxy* and *Star Wars*. If he could live a day again, he first picks the future, because time may be emergent, and he would solve AIXI and ask the meaning of life. From the past he cannot pick a whole day, only a moment. He arrived at compression as the shortest self-extracting program before he knew Kolmogorov complexity had a name, already knew sequential decision theory, and knew at once that welding the two was the right construction. The setting was mundane: years of patented image interpolation at a company, then the physicist's habit of ignoring compute and asking what the ideal interpolator would be. Thinking back, he is still super excited.