Intelligence Is How Efficiently You Acquire Skills You Were Not Prepared For

Lex Fridman opens his second conversation with François Chollet — Google AI researcher, creator of Keras, and author of the paper On the Measure of Intelligence — by asking which thinkers shaped him. Chollet names Jean Piaget, the Swiss psychologist often called the father of developmental psychology. He read Piaget as a teenager. The work itself is old, mostly from the 1930s and 1940s, and later developmental science has superseded many of the specific claims. What struck him was not the dated conclusions. It was the notion that intelligence and the mind are things you construct throughout a life, and that children construct them in stages.

That idea, he tells Fridman, is directly relevant to AI. Building an artificial mind is, in essence, trying to walk a synthetic system through a similar process of construction. Around the same time he read Jeff Hawkins's book On Intelligence, which pictures the mind as a multi-scale hierarchy of temporal prediction modules. The claim that cognition is prediction was new to him then, and he loved it. Hierarchies in AI were not new — they had been around since the 1980s — but Chollet says they found their practical implementation only later, in deep learning.

Fridman asks about memory, the other half of Hawkins's picture. Neural networks memorize, Chollet agrees, yet that is not the kind of memory brains seem to have. The brain is more like a sparsely accessed store: you can retrieve a precise bit of experience, introspect, and ask yourself questions. You can, in that sense, program your own memory. Language is the tool you use to do it.

Chollet calls language "a kind of operating system for the mind." One of its uses is as a query you run over your own memory: words become keys that retrieve specific experiences, concepts, and thoughts. Language is how you store thoughts, not only in writing but inside the mind, and how you reach them again. Imagine you had no language. You would lack an internally triggered way to pull past thoughts on demand. You would wait on an external cue — a familiar sight, a smell — to bring memory back.

Fridman puts a Chomsky-style question: is language at the bottom of cognition, the last turtle? Chollet says no. He disagrees with Noam Chomsky. Language is a layer on top of cognition. In the computing metaphor, the computer exists before the operating system; the OS is how you make the machine truly useful. And this OS, he jokes, is most likely Windows, not Linux, because language is messy and hard to introspect.

Before we think in words, he argues, we think in emotion, space, and physical action. Babies probably express thoughts in terms of actions they have seen or can perform, and in terms of the motions of objects around them, long before they think in words. Visual analogy and motion are fundamental building blocks of the mind, and ordinary language is full of spatial metaphors that still show it. Chollet considers himself a visual thinker who navigates a two-dimensional concept space. Fridman says he has to put ideas down on paper and can only travel on the page. Chollet calls him a Flatlander, then asks whether, even when writing a paper, Fridman does not hold a subway map of where ideas sit topologically. Fridman admits he does. Writing a paper feels like bidirectional path search: A* from both ends until the paths meet.

A mind map, Chollet says, is a way to take the connected mess inside your head and put it on paper so you gain control over it. You start with a key concept — intelligence, say — then add associative links: language, motion, whatever comes. It is more graph than tree, more subway map than org chart, and you can draw as well as write. Fridman, who describes himself as OCD about categorization, worries that an implied hierarchy will freeze him. Chollet's reply is casual: it is your map. Draw any connection you want. If the central node is wrong, draw another map.

The deeper point is about syntax. If you organize ideas by writing them in paragraphs, the prose imposes a heavy syntactic structure that the thoughts may not yet deserve. A mind map is a lower-level, more freehand way of laying them out; once they are on the page you can start voicing them as paragraphs. Chollet puts the challenge cleanly: how do you know what you think about something if you do not write it down? Fridman, sitting on thousands of Google Docs of bullet lists, notes that a bullet list is a tree without the visual layout. Chollet agrees it is not the same object. The mind map forces you to be semantically specific in a way sparse poetic notes do not.

That leads him to a distinction he cares about for AI. It is reasonable, he says, to assume some topological processing in the brain, because the brain is highly associative. A topological space encodes thoughts as connections: points are either linked or they are not, like stations on a subway map. Geometry cares about distance. Deep learning works in geometry — concept vectors, word vectors, relationships expressed as a dot product. Everything has to be a vector because everything has to be differentiable. If your space is discrete, it is no longer differentiable, and you cannot do deep learning in it unless you embed that topology inside a larger continuous space. Chollet believes a topological space is a better medium for thoughts than a geometric one.

Fridman zooms out to the November 2019 paper and asks an absurd textbook question: two hundred years from now, if there is a chapter on biological minds, what quote sits next to Chollet's picture? Chollet's definition is clean. Intelligence is the efficiency with which you acquire new skills at tasks you did not previously know about and did not prepare for. It is not skill itself. It is not what you know or what you can do. It is how well and how efficiently you can learn new things. You see it on display whenever a human — or an AI — adapts to an environment its creators did not anticipate. Improvisation and generalization are the signal. A system that cannot deviate from what it was hardcoded or trained to do is not intelligent.

Einstein's line, "the measure of intelligence is the ability to change," captures part of the idea. Chollet wants a more precise version: the efficiency of adaptation, of mastering an environment, of picking up new skills. The distinction he keeps drawing is between intelligence as a process and skill as the output of that process. A smart human programmer looks at chess and writes a static program that plays it. The intelligence is in the developing. The program is the frozen artifact. Ask it to play Go and it fails without further human involvement, because the entity capable of the process was the programmer. Do not confuse a road-building company with one specific road. A road takes you from A to B. A road-building company can make a path from anywhere to anywhere else.

The paper's goal, Chollet says, is to clear up long-standing misunderstandings about how the AI community has conceptualized intelligence and how it has scored progress. Recent machine-learning results have tempted people to extrapolate that general intelligence is about to be solved. To evaluate that claim you need a precise definition and a reliable, actionable measure — not a binary badge, but a feedback signal that points toward building more intelligent systems. Skill is the only thing you can objectively measure, which is why the temptation to tally skills is so strong. In a human, a strong chess player is evidence of intelligence because you know they were not born knowing chess. In a computer, you cannot make that assumption. It may have been born knowing chess, in the sense that a human encoded the output of understanding into a static program.

Chollet sketches two historical pictures of the mind that still split AI. One is the evolutionary-psychology view: a collection of fairly static, special-purpose mechanisms hardcoded over a long species history. Early AI researchers such as Marvin Minsky subscribed to it. They saw the mind through the metaphor of the mainframe — static programs operating over a database-like memory — and treated learning as mere memorization. Learning barely featured in AI textbooks until the 1980s. The other picture is the blank slate, Locke's tabula rasa: an information sponge that starts empty and absorbs the world's complexity. Connectionism, and then deep learning, gave that view a new vitality. Today, Chollet says, many researchers conceptually treat the mind as a randomly initialized network that gets trained on data.

GPT-3, then newly famous, is the exhibit. What interests Chollet is the claim that it can pick up a new task after a few examples. If that is really happening, it is novel and worth investigating. He is not convinced it has been shown. Given the amount of data the model is trained on, what it is probably doing is pattern-matching the prompt onto a task it has already seen — recognizing the task rather than building a model of it. Fridman offers the earlier metaphor: maybe the prompt is a SQL query into a giant associative memory, and if the memory is large enough it starts to look like intelligence. Chollet allows that a significant chunk of intelligence may be giant associative memory. He does not believe intelligence is only that.

The ceiling of GPT-N, in his view, is ever more plausible, context-aware text. Scaling transformers and data will not fix the actual flaws. The text is constrained by plausibility and nothing else — not factualness, not even self-consistency — which is why it is easy to elicit false or contradictory statements, and why the same question asked two adversarial ways yields two answers. Program synthesis looks the same: the model does well on programs it has seen and will not generalize through a non-interpolative program space. The bottleneck is not model size or training time. OpenAI is already training on a crawl of essentially the entire web. You will not get another hundredfold of data. Quality matters more than volume. Chollet recalls an image-classification model at Google trained on some 150 million noisily labeled images; a smaller set with human annotations beat more noisy data, and trained faster besides.

The semantic web of the 1990s — attach machine-readable meaning to the internet — will never work, he thinks, because the labor is huge and nobody has an incentive to do it. Unsupervised deep learning is a bigger step toward making web knowledge available to machines than the semantic web was. That still is not reasoning. Forms of reasoning you see in GPT-3 are reproductions of patterns in the training data, and they break on a genuinely novel situation. Train such a model on data from 2002 and use it today: it will be missing facts, vocabulary, even the coronavirus. Intelligence requires continuous learning and improvisation. It is not enough to assume that what you will be asked is something you have seen, or a simple interpolation of it.

The interpolation assumption already fails on tasks that look simple from a distance, Chollet says. A Google paper from a couple of years earlier showed that some 30 million distinct road situations were not enough to train a driving model, and it was not even Level 2. That is far more data than the twenty or thirty hours a human needs to learn to drive, given everything the human has already accumulated. Tesla's learning-based Autopilot push does not move him off the point. Level 4 is probably achievable. Level 5 — fall-asleep autonomy — probably is not, if the plan is to anticipate every situation with training data. Simulation can help, but it is a massive effort. You get a more adaptive system by using deep learning for perception and intuition, then pairing it with an explicit model of the car's surroundings. Strong generalization in today's AI, he says, tends to come from abstractions in a human mind encoded as programs, not from the weak abstractions a network learns.

Fridman notes that humans cheat: before we "learn to drive" we have watched cars, ridden in them, seen them in movies. Neural nets do something similar with data. How many leaps of reasoning genius does driving actually take? Chollet answers with a transfer story. Suppose you learned to drive in Silicon Valley and now you rent a car in Tokyo. Traffic is on the other side, signs are different, roads are narrower. An average human should zero-shot it, despite novel complexity that is not an interpolation of American driving. That is the test he cares about. Driving itself is not much of an intelligence test, because no task-specific skill demonstrates intelligence unless the task is the meta-task of acquiring new skills. With infinite training data you could train an end-to-end network to drive. The only problem is covering the long tail. That is a scale problem, and it hits diminishing returns. A more engineered hybrid would start generalizing earlier.

Even a Level 5 driver, he insists, would not demonstrate generality. The only possible test is skill acquisition on unknown tasks: take the L5 stack and ask it to learn to pilot a commercial airplane, then measure human involvement and training data. Fridman objects that driving is a black box that keeps generating edge cases, so it does have newness. Chollet agrees it is a very long tail. That does not mean you cannot solve it with a statistical model and a huge amount of data. If the dataset is so large that new situations almost never appear, that is not intelligence. If picking up a skill requires a training set that covers most possible situations, the system is mostly a lookup table. If it requires an engineer to write rules covering most situations, the system is the output artifact of a process in the engineers' minds. Intelligence would be autonomously producing those abstractions.

Fridman says the boundary feels gray. Chollet agrees that deep learning does form abstractions and does generalize — a little. Generalization is a spectrum, not a binary. Intelligence is how efficiently you generalize far outside the distribution of things you have already seen: both the distance of the novelty and the efficiency. He offers an information-conversion ratio. Imagine a space of possible situations. You start with some coverage, from situations you already know plus prior knowledge embedded in the system. The job is to turn that information into a behavioral program that covers a large area of the space. The ratio of that area to the information you started with is intelligence. A smart agent covers a lot of future situations with very little information and very little prior knowledge, without knowing what those situations will be.

If the goal is human-like AI, Chollet would like a test that can be given to both machines and people so you can compare them. The constraints are not the same, and the test has to account for that. An experimenter can buy arbitrary skill in an artificial system by injecting hardcoded priors from programmers' minds, or by training on more data — infinitely many Go games, for instance. A human Go player developed the skill in a constrained lifetime, from a different set of innate priors. To compare the two you have to control for priors and for experience. Priors are whatever information you have about a task before you start learning it. Experience is acquired: the games you have played, seen, or simulated. If you want to measure a human-like form of intelligence, the AI should start from the same priors humans start with, and those priors should be made explicit.

On the human side, the relevant field is psychometrics: the branch of psychology that tries to quantify cognitive abilities, intelligence, and personality traits. Psychology often has a bad reputation for irreproducible results. Psychometrics, Chollet says, actually has some fairly solid, reproducible ones. A good test should be reliable, valid (it measures what it claims, correlating with things like school and workplace success), standardized, and free from bias — an English-language test is biased against people who do not speak English. Not every test meets those ideals, but the field knows its weaknesses. Run many different tests — language, spatial vision, mental rotation, numbers — at scale and you see clusters of correlations. People who do well at math tend to do well at physics, and also, statistically, at writing an English essay. The latent variable that explains the fact that every intelligence test you can invent ends up correlated is the g-factor. It is a statistical construct. You cannot measure it directly in a person the way you measure height. At scale, it is there.

Chollet is not interested in psychometrics as a way to characterize one individual. An IQ score or a personality assessment does not tell him much about you as a person. The tools are useful at scale, when you cross-correlate results for many people and read off the structure of human cognitive abilities — which is what is relevant to AI. The mainstream map of that structure is CHC theory, named for Cattell, Horn, and Carroll: a three-level hierarchy with g at the top, broad abilities such as fluid intelligence in the middle, and narrow, more task-specific abilities at the bottom. Different statistical analyses grow slightly different theories. They all describe a hierarchy with a g-factor at the top.

That does not mean human intelligence is general in a strong sense, or that a high-IQ person can solve any problem. Chollet's analogy is physical fitness. Fitness is a useful, intuitive concept. None of us can fly. If you were a scientist trying to measure it, you would run a battery of tests — a hundred-meter dash, soccer, table tennis, swimming — and you would see correlations, plus physiological correlates such as lung volume, and at the top a physical g-factor. High fitness still would not mean you can fly, live on the ocean floor, or exist on Venus. Human morphology is highly specialized. What is striking is that it generalizes far beyond the environment we evolved for. We evolved, roughly, to run after prey on the savanna, and we can climb mountains, swim lakes, and play table tennis. Cognition is similar. We can write novels and go to Mars, far beyond what the mind was "for," and we still have strong cognitive biases that make us excellent at some problems and useless at others. Generality itself was one of the evolutionary constraints, because the environment was variable and unpredictable. The result is a body and a mind that apply much more broadly than they were evolved for — and still nowhere near universal.

IQ items, Chollet says, typically look like a few demonstration input-output pairs plus a test prompt you must complete. In that narrow frame, a difficult item is one whose prompt is very surprising given the training examples. Give two rotation examples and then reuse one of them as the prompt, and generalization difficulty is zero: you recognize the training case and copy the answer. A harder item requires test-time adaptation, some improvisation. The same logic applies to a quantum-physics exam. If you want to test understanding, you write something unlike the mock exams students crammed. If you want to make it easy, you interpolate the mocks. Even an exam that forces improvisation is not intrinsically hard. A working physicist would not find it new. Difficulty is always relative to the test-taker's priors and to the information they have about the task. That is what he means by controlling for priors.

People who write hard IQ items, Fridman has learned, religiously protect the answers, because once you know the approach the item is ruined. Chollet's version of that fact is sharper: the approach is implicit in the training examples, so publishing those examples is the leak. That is why ARC keeps a private test set no one has seen. He is not a fan of ambiguity in test questions. You can have difficulty without it, by forcing a lot of extrapolation over the examples. The tests he wants are not necessarily hard for humans — human intelligence is the benchmark — they are supposed to be hard for machines in ways that are easy for people, and they should be actionable: they should show you that progress is needed and in which direction.

Novelty is a requirement. If you can practice the items you will be tested on, you are not exhibiting intelligence. You are retrieving what you were exposed to, the same way a deep-learning model aces a test after training on all the possible answers. A student can pass by memorizing a hundred mock exams and hoping the real one interpolates them, with no understanding of the material. Many students do pass that way. A machine intelligence test, then, has to start by acknowledging the priors it requires. If the goal is to compare machine and human intelligence, assume human cognitive biases. Then test skill-acquisition efficiency, not skill: every task should be novel, unanticipatable, not brute-forceable in advance — not just by the system, but by the system's creators.

Making those priors explicit is, Fridman says, one of the most powerful moves in the paper. The researcher who has done the most on innate human knowledge priors is Elizabeth Spelke at Harvard, whose core-knowledge theory outlines four systems we are either born with or hardwired to acquire so early that the distinction barely matters. The first is objectness and basic physics: something that moves cohesively is an object; objects can bump and occlude. The second is agentness: some objects have goals. Two dots moving in rough synchrony make you infer a chase. Spelke has shown that three-month-old babies identify agents and goal-directedness. The third is basic geometry and topology — distance, navigation — backed by grid cells and place cells in the hippocampus. The fourth is number: we can do basic counting and compare quantities without culture teaching us arithmetic. The list may not be exhaustive. Spelke has considered a system for social relationships, which would matter less for ARC. A 1970s mental-rotation experiment found that the time to decide whether one shape is a rotated version of another is linearly proportional to the angle, as if the brain had a turntable with a fixed speed.

ARC, the Abstraction and Reasoning Corpus, is Chollet's attempt to embody as many of those principles as he could under real constraints. He does not claim it is perfect. The format is close to classic IQ tests, especially Raven's Progressive Matrices. Each task has a few training input-output pairs: colored grids of variable size. You are shown demonstrations, then a new input, and you must produce the correct output. The hard constraint is that every task should require only core-knowledge priors — no language, no English, no trees or dogs or cats drawn from human experience — only reasoning built on those four systems. Some tasks are written to probe specific forms of abstraction. One motive for building ARC, he says, was that a problem as murky as autonomous abstraction has to co-evolve with its tests. Designing the tasks was a way to clarify his own ideas. Humans, including young children, find many of them easy. Machines found them near impossible.

The main difference from IQ tests is that the priors are explicit. ARC is also more diverse, and producing a solution can mean a lot of clicking; grids can be as large as 30 by 30. There are about a thousand tasks, including a private test set. Writing them is hard because every task must be novel, unique, and unpredictable. If you accidentally built a small world simple enough that a human could reverse-engineer an algorithm to generate every ARC task and its solution, the test would be invalidated. Chollet found that a human is not a very good source of unthinkable novelty. There are only so many truly unique tasks you can invent in a day, so you have to pace the work. He treats ARC as a work in progress, with redundancy and imperfections he wants to keep refining. Crowdsourcing, with several layers of filtering, could grow a bigger, more diverse set and wash out some of his personal biases. One part must stay hidden: the test used to benchmark algorithms cannot be available to the people developing those algorithms. Otherwise engineers solve the tasks themselves and encode the solutions. What you would be measuring is intelligence in a human mind, captured as a crystallized output that is not itself intelligent.

An unexpected side effect: people, including children, started playing ARC as a game. Chollet had released it as a benchmark of fluid general intelligence. He finds the game-playing encouraging. Fridman notes that because ARC lives in the context of a machine-intelligence test, solving it as a human forces you to introspect how you found the solution — to watch your own mind generate abstract representations. Chollet thinks that is because the prior set is so small: everything is a recombination of a very small set of assumptions. He ran ARC as a Kaggle competition, plans to keep evolving it, and is starting a collaboration with psychologists at NYU to collect human data, including whether human-perceived difficulty tracks machine-perceived difficulty.

Machine performance started at zero while humans found the tasks easy — a flashing light that something was missing. It did not stay at zero. Within two weeks of the Kaggle competition there was a non-zero score. At the time of the conversation the state of the art was around 20 percent of the test set, against human performance closer to 100 percent. That is the profile he wanted: hard enough to show the need for progress, not so impossible that you cannot start, and still far from solved. The competition also showed there was no obvious shortcut. People who scored were, in a sense, exploiting flaws by anticipating what the test set might contain — human hacking of the benchmark, which future versions will need to address. Chollet believes it will take a while to reach human parity, and that by the time we do, we will have systems with a degree of generalization that matches human fluid intelligence. They would not necessarily look like AGI. They would not necessarily look like us.

In machine learning, a system "generalizes" if it can make sense of an input it has not seen. Chollet calls that system-centric generalization: novelty relative to that system. A good intelligence test, he argues, should demand something stronger — developer-aware generalization. The system has to handle novelty or uncertainty that not only it has never seen, but that its developer could not have seen either. That is the edge-case problem in autonomous vehicles, restated as a definition: neither the designer of the training data nor the contents of the training set know about the case. The system still has to cope.

He puts degrees of generalization on a spectrum. The lowest level is what machine learning usually tries to do. New situations are assumed to be sampled from a static distribution you already have a representative sample of — your training data. You generalize to a new draw from a known distribution, along dimensions of novelty the developers already understand. That is robustness: holding up to noise and small variation on one fixed, known task. A higher degree is flexibility: an L5 car, or a robot that passes the coffee-cup test — drop it in a random kitchen somewhere in the country and have it make coffee. Flexibility is dealing with unknown unknowns, dimensions of variation the creators could not have foreseen, still within one task. The long tail of driving is flexibility.

Extreme generalization is flexibility without the single-domain fence. A robot designed and trained for cooking that, after you buy it, can teach itself gardening in a couple of weeks would be showing it. The ultimate goal is a system that reaches human skill parity on arbitrary tasks and domains, with human-like improvisation when it hits something new, over roughly the same range of work humans do, using roughly the same amount of practice. That would be human-level extreme generalization. Chollet does not think humans sit anywhere near an optimal intelligence bound, if such a bound exists. There is probably a hard limit on how intelligent any system can be. Humans are not near it. We are bounded by our environments and by the problems we try to solve.

Fridman asks about Neuralink and outsourcing brainpower. Chollet is skeptical of neural interfaces because they target a bandwidth bottleneck he does not think is currently binding. We already have sensors that take in more information than we can process. Wikipedia made humans more intelligent, he agrees, but that is externalized cognition through information systems, which is not the same as a brain-computer port. Your brain already has fairly direct access to Wikipedia via phone, hands, eyes, and ears. Speed-reading does not work; the faster you read, the less you understand. The brain is slow — on the order of tens of milliseconds for the fastest events, seconds to form a conscious thought — and a trained typist already emits thoughts about as fast as they form. Augmenting human intelligence is a real avenue. That is what computers are, and what culture and civilization are. Culture is externalized cognition. Language and books are forms of it. Civilization itself has capabilities far beyond any individual brain, and it is not bound by one. Search engines are already part of that organism.

Fridman asks about the Turing Test and the Loebner Prize: open-ended dialogue scored by human judges. Chollet is not a fan of the test or its variants, for two reasons. First, it cops out of defining and measuring intelligence by handing the whole problem to a panel. Those judges may have no methodology, no definition of intelligence, and no reliability. That already fails a core psychometric principle. Biased human judges also violate standardization and freedom from bias. You are outsourcing everything that matters — a precise description of intelligence and a standalone test for it — to people. Turing, Chollet reminds Fridman, did not propose the imitation game as a goal for the field or as an actual intelligence test. In the 1950 paper it was a thought experiment in a philosophical argument: that something indistinguishable from a human mind could, in theory, be encoded in a Turing machine. That was a daring claim then. It is much more widely accepted now that the mind is an information-processing system.

The second reason is incentives. If the goal is to convince judges they are talking to a human, you have an incentive to use tricks. Chollet's analogy: suppose the test for teleportation in physics were to convince a seated panel that teleportation had occurred. David Copperfield could do that in Vegas. The act is elaborate. It is not physics, and it does not advance understanding of the universe. Fridman pushes back: maybe it is easier to solve the problem generally than to keep fooling a large number of judges. Chollet disagrees. In practice it is easy to deceive people. We have theory of mind. We constantly project emotions, intentions, and agentness — one of our innate priors — onto everything around us. Paint a smiley on a rock and the rock becomes happy. That anthropomorphic bias is why tricks work.

He likes the Alexa Prize better because it is more pragmatic: it incentivizes a useful human-machine interface rather than pure imitation. Being indistinguishable from a human, he argues, is the last step in creating machine intelligence, not the first. The first systems that show strong generalization and implement human-like broad cognitive abilities will not look like us. A good test points you at the first rung of the ladder, not the top. Fridman defends interactivity: conversation forces adaptation in a game-theoretic environment the ARC grid, as a silent observer's puzzle, never changes. Chollet grants the point. Interactivity is a good setting for forcing generalization. It is also not scalable or reliable, because it relies on human judges. He originally wanted an ARC variant scored by number of attempts until the right solution, so you could apply something like the scientific method — form a hypothesis, probe, update — and even tolerate more ambiguity. Measuring the quality of those attempts would be better still, and even more subjective. He would love to build it. In practice it would be very difficult.

Marcus Hutter's Hutter Prize asks contestants to compress English Wikipedia, on the idea that cognition is compression and the brain is a compression algorithm. Chollet calls it a fun, striking, old idea. He used to believe it. He no longer does. It is easy to believe because we compress constantly, we have a bias toward simplicity, and Hawkins's "cognition is prediction" is compression with a time axis. Compression is a tool in the human cognitive toolkit. It is not cognition itself. Cognition is about operating in future situations that include fundamental uncertainty and novelty. Take a ten-year-old with ten years of rewards and punishments. The shortest optimal behavioral program for those ten years — what you would get if the child's mind were a compression algorithm — would be utterly inappropriate for the next seventy. The models we build of the world are not perfectly compressed, because they have to include things that have seemed useless so far and may matter later. You do not know the future.

The investing analogy is the same one he used for local versus extreme generalization. Compress twenty years of market data and the "best" strategy is to have bought Apple, then Tesla. That is probably not the strategy for the next twenty years, which is why a smart investor holds a balanced portfolio. Children, driven by play and curiosity, are not efficiently compressing. They are exploring, retaining things that seem useless because they might turn out useful. That hedging against future uncertainty is antagonistic to compression. Cognition leverages compression to promote efficiency, in the spirit of the line Chollet attributes to Einstein: make it simpler, but not too simple. A good world model includes apparently useless diversity, the way a portfolio includes stocks that have not performed well yet, because fundamentally you do not know what you are doing. The future is not going to be like the past. Compressing the past is not a policy for it.

Fridman closes with the question Chollet did not get last time: what is the meaning of life? Chollet's answer starts from culture. Everything that makes us up, even our most personal thoughts, is not actually our own. Those thoughts are expressed in words we did not invent, built on concepts and images we did not invent. We are cultural beings; that is a large part of what makes us different from other animals. Everything about us is an echo of people who lived before. If we contribute something to the collective edifice — a new idea, a piece of music, a theory, even a new word — it becomes part of the minds of future humans, essentially forever. Everything we do creates ripples that propagate forward. That, he says, is a path to immortality: culture becomes future humans, and we keep influencing people thousands of years from now. We are the sum of ripples from the past, and we create ripples into the future. That is why we should be kind. Every act of kindness creates ripples. So does every act of violence. You want to choose which kind you propagate. Fridman notes that Chollet's own ripples will run through future humans and future AGI systems. That, both agree, is a fair place to stop.