Gary Marcus: Deep Learning Imitates Intelligence. It Does Not Understand.

Lex Fridman introduces Gary Marcus as professor emeritus at NYU, founder of Robust.AI and of Geometric Intelligence, the machine-learning company Uber acquired in 2016, and co-author of *Rebooting AI: Building Machines We Can Trust*. Marcus has spent years as a critical voice on the limits of deep learning. Fridman asks whether civilization will one day face a singularity that revises our place in the food chain. Marcus's first correction is that the food chain has already changed: machines already do by hand what people used to do. If singularity means one single moment, he doubts it will look like that.

The trend, he says, is still toward human-level intelligence, or whatever one wants to call AGI. Ray Kurzweil sees an exponential; Marcus often sees linear, and in some cases close to zero right now. Tedious and dangerous work will move to machines. Meaning may shift from jobs toward creative expression. The transformation will sit somewhere between sudden and gradual, and it cannot be one day. The deeper mistake is treating intelligence as one number, as if the day a machine hits IQ 262, humans are displaced.

Intelligence is a multidimensional variable: verbal, motor, mathematical, and more. Machines already far exceed most people at mathematical intelligence and at games. At understanding language they lag far behind his five-year-old. Some facets machines have grasped; others they have not. Vital prerequisites are still missing. Machines do not have common sense. They do not understand that bottles contain water, that people drink to quench thirst, that they do not want to dehydrate. That gap, he argues, is a rate-limiting step for reading a story or watching a movie, because plots depend on background knowledge no model currently holds.

Even a machine-interpretable store of common sense would not be a single point. Physical reasoning — how keys fit locks, how a microphone stand works — might come long before good psychological reasoning, because labeled data and direct experiment are easier on objects than on people. Fridman presses the contrast. Marcus treats common sense as including both. A robot can lift a bottle, try a cap, watch it fall off, turn the bottle upside down. Humans do some psychological reasoning by looking inward. Robots have no such intuition, and they cannot run the equivalent experiments on people. A robot that pinches you in different ways to map pain will not pass a human-subjects board, and companies will get sued.

Fridman plays devil's advocate: human behavior is already digital. Facebook algorithms observe and even manipulate people in ways robotics cannot. Zuckerberg, at what was then NIPS, talked about Facebook having a gigantic theory of mind. Marcus grants the point and dislikes how it is used. Facebook, he says, has a really good idea of how to addict people and what draws them back. Even so, only some slices of human experience are accessible through that interface. Love, fear, and insecurity might someday be explored in digital form, especially as VR expands the data. His new company with Brad Brooks is interested in both physical and psychological reasoning.

Some pieces of each are already accessible. Whether a robot can fit under a table is relatively easy if you know two heights. Jenga needs finer data. Knowing that people have goals is not that hard; knowing exactly which goals, or the sources of frustration, is. It is not inconceivable that machines will soon understand that people do not like losing a game — the preference is consistent and easy to code. Expand the scope to sexual frustration, a missed promotion, a fight with family, and existing techniques cannot hold it. Chess, by contrast, is a closed world. Marcus has been crushed by Deep Thought and by Garry Kasparov playing thirty boards at once. He is not sad when a computer takes a cube root faster than he can. AlphaGo does not change the diagnosis: AI is a grab bag of techniques, each with strengths and weaknesses, and the tech that drives AlphaGo does not transfer to reading.

Fridman groups three problems: moving through the world and making an omelet; understanding writing and holding a dialogue; and Go or chess. He argues language is much closer to Go than to physics — huge, but still constrained. Marcus agrees in part and disagrees in part. Language abstracts. One word, *on*, covers an analog infinity of bottle positions on a table. Grammar is a real constraint. People follow it, mostly. Those are ways language filters the world.

The other side is vastness that is, by reasonable measures, infinite. You can stack embeddings: you think that I think that your mother thinks the water is on the table. You can stack adjectives without bound. Almost anything in the physical world can be talked about in language, and many sentences are only understandable if the listener already has a rich model of that world. Marcus will not fully adjudicate which is harder. He will say language is really, really hard for machines and for linguists — and not that hard for children. That gap drove his career. He studied with Steven Pinker, trying to figure out why kids could learn language when machines could not.

Pushed to be arrogant about a hundred-year forecast, he first defends the humility in the book: people who figured out electricity did not foresee cell phones, and people who made transistors did not foresee social networks. Safe predictions are Bezos-like. AI will be faster, cheaper, more general, more pervasive. There is no good in-principle argument that intelligence is magic inside brains that silicon can never capture. AGI, when it comes, will not look just like humans. Humans have terrible memory and motivated reasoning; there is no reason a machine should copy those bugs. Daniel Kahneman and Marcus, on a panel the night before, converged on a line: humans are a low bar. They may be outside current skill. Eventually AI will exceed them. The goal should be general intelligence without the flaws, not a replica.

Fridman asks whether human flaws — limited memory, mortality, bias — are the source of motivation and meaning. Marcus has never found those arguments convincing. We make lemonade: free association becomes poetry; terrible memory becomes the game Concentration, which a computer would find trivial because it sees every card at once. Confirmation bias and motivated reasoning make society rougher because we cannot interpret data in shared ways. Forgetting where you parked in the same lot, because yesterday's memory blurs into today's, is not optimal. Being able to wipe a traumatic sector would be a feature. Humans are a source of things machines should steal, not a ceiling. He wants systems that combine brute-force computation with a doctor's ability to read — a machine that can ingest the entire medical literature in a day, thousands of papers no physician can cover.

Yann LeCun's claim that human intelligence is not general at all, just very narrow, does not land. Anyone can walk into a Hollywood movie and reason about a bank robbery, IVF, almost any scene. Not everyone understands every scene, but the range for an ordinary adult is huge. LeCun's retort is that the set only seems large because humans cannot imagine the true space of possible experience. Marcus will grant that humans are constrained. That is not the same as a Go system trained on five million games that cannot play on a rectangular board without another five million from scratch. Any decent human could transfer. That is what he means by narrow.

General, for him, means transferring knowledge across domains: bank-robbery chase scenes in movies help you understand the one-person chase in *Breaking Bad*, and why it is like every other chase and also totally different. The fact that the number of domains is finite does not make it less general. Humans are a low bar. Right now the bar is on the floor. Eventually we will get way past it. His 2018 paper "Deep Learning: A Critical Appraisal" listed ten challenges: data efficiency, transfer, hierarchical knowledge, open-ended inference, explainability, prior knowledge, causal reasoning, an unstable world, adversarial examples, reliability. Asked which solved would matter most, he is evasive because they travel together. A real solution starts with cognitive models of what is going on, not statistical approximations. Systems know bottles correlate with caps. They do not understand threads, seals, or why water stays in.

Could those abstractions emerge inside a very large net? Marcus is a skeptic of that kind of emergence. Deep learning might play a role in the systems he wants; it will not do the job by itself. He has never seen a deep net really extract an abstract concept, for principled reasons in how backpropagation and the architectures work. Convolution, the idea Yann LeCun is famous for, is the exhibit. It is an abstraction people do not have their systems learn: an object looks the same in different places. Program it in innately and the system becomes far more efficient. In principle it should be learnable. In practice, without it, a net that recognizes a clock here will not apply the same fact when you slide the clock over. We need more of those innate features, not less.

Fridman notes that inventing convolution won a Turing Award and still took less labor than encoding an expert system. That is why people love deep learning: it does a lot automatically. Marcus quotes Bertrand Russell: it has all the advantages of theft over honest toil. Programming causality, or how a bottle works, or what a container is, is brutally hard. He and Ernie Davis wrote a forty-five-page paper just to understand containers. Almost nobody read it. Davis spent about three months getting the logical statements right. Nobody wants that work, including Davis. Everyone would rather feed in videos of containers and let the machine infer. Marcus understands the impulse. He has never seen a system watch video and robustly predict which containers leak. LeCun's royal road is an unsupervised algorithm that predicts the next frame. Marcus's twenty-year bet is that the hard work is a design specification, more engineering of the convolution kind, encoding complex knowledge current nets do not have.

The 1998 experiment used an auto-associator, a precursor of contemporary deep learning, trained to map binary numbers onto themselves. Train on all the even numbers and the net never generalizes to odd. People thought he was faking it. The nets were not stupid; they see the world differently. They asked for the probability that the rightmost output node is one. In every training example it had been zero, so why turn it on now? A person sees algebra: add one and parity flips; *mutatis mutandis*. F of X equals Y plus two works for three and for a million. Deep learning emulates that and does not do it. What it learns is correlations among nodes, not operations over variables. Differentiable programming is tentative work in the right direction. Classic image nets do not have it. Hinton has called symbol manipulation the gasoline engine, obsolete next to deep learning's electric power. Marcus thinks that analogy is destructive. Most working computers are still built out of the variables and conditionals John McCarthy put at the founding of the field. No rational person would machine-learn a web browser from keystroke logs and screenshots.

Marcus is not endorsing 1980s expert systems. Those systems tried to capture medical knowledge as handwritten rules: if these symptoms, then this disease. They were almost entirely rules, almost no machine learning. Today's mainstream is almost entirely one species of machine learning, no rules. Both have advantages. For recognizing a bottle, deep learning is the best tool we have. For inferring what a bottle does, something closer to expert systems is still the best available alternative, preferably with a better grip on quantitative and statistical information than the old systems had. New technology should draw strength from both and find new ways to synthesize them.

Adding knowledge at the low level is often really hard, partly because people will not bother to tell you the obvious. Pay someone on Mechanical Turk about bottles and they skip the basic facts that are trivial to a human and brutal for a machine. Crowdsourcing has not fixed this. Virtual Home at MIT asked people to describe an exercise routine and got micro-steps: go to the room with the television and the weights, press the remote, lift, put down. That is not what an exercise routine is — fitting a certain number of movements into a time window, targeting muscle groups. Crafting the right knowledge takes expertise in how the AI works. The ocean of common sense has a mathematical property: every piece, if stated, we will recognize as true or false; we are unlikely to retrieve it unprompted. Most of it has never been encoded in machine-interpretable form.

The closest attempt is Cyc. Logicians worked for Doug Lenat for thirty years. Marcus thinks they stuck too close to logic, under-represented probability, and hand-coded their way into various issues; it has not been that successful. People treat that one instance, conceived three decades ago, as the failure of an approach — a standard they do not apply to deep learning. In 2010 the same attitude hit deep nets; GPUs, ImageNet, and a few tricks changed the outcome. No 2010 deep-learning model, he says, has commercial value now; that did not mean there was nothing there. A friend who did deep learning around 1986 failed for lack of compute and data, not because the algorithms were alien. Marcus does not want to resuscitate 1980s symbol systems as-is. He wants to borrow lessons, combine them with what else we have learned, and maybe get an ImageNet moment for common sense. The Allen Institute, he notes, is building benchmarks to spark that interest.

Fridman quotes Rich Sutton's essay: seventy years of AI research teach that general methods which leverage computation are ultimately the most effective. Marcus's first reply is: most effective at what? They have been most effective for perceptual classification and for some reinforcement-learning problems, which is Sutton's home turf. Fridman pushes back that deep learning is the first thing that has been successful at anything in AI. Marcus disagrees. Classical AI does navigation instructions that everyone on the planet uses multiple times a day. Those successes are so pervasive nobody notices them. Deep learning has been more successful in public, but the line he likes is that building a better ladder does not mean you can build a ladder to the moon.

Throwing data at perceptual classification beats anything else. That has not given material progress on natural-language understanding, on the common sense a robot would need to navigate a home, on problems like that. More compute is always better, the way more money is always better, with diminishing returns. The existence proof sits at twenty watts. A human brain does things deep learning and current symbol systems have not figured out, without Google-scale servers. He and his wife built two such intelligences. Watching his five- and six-and-a-half-year-old, a year and a half apart, he is struck that they set their own problems. They invent games, run what-if scenarios, grow their minds. Machines do not really do that.

Fridman asks nature versus nurture. Marcus thanks him for phrasing it as both. People dichotomize when it has to be nature and nurture together. You cannot learn along the way without some innate stuff; having innate stuff does not mean you learn nothing. In *The Birth of the Mind* he asked how genes even build innate knowledge. A baby ibex climbing down a mountain hours after birth knows, not consciously, something about its body, physics, and 3D geometry. Developmental psychology, including work he has done, suggests children arrive with frameworks: space, time, other agents, places, a kind of mental algebra, causation. Those systems cannot be totally disjoint, because bottle-cap outcomes depend on relations among space, time, and cause. Evolution is terribly inefficient until it gets a good idea, then it runs: a vertebrate brain plan after roughly a billion years, then fish, dogs, humans; color vision spreading fast. Libraries get reused — the multi-digit plan for a hand also builds a foot, with different parameters. Evolutionary computation, he thinks, starts with libraries that are almost nothing, progress is slow, and students give up. Richer initial structure might speed it up. A good engineer can sometimes invent what evolution found, often by looking at biology. AI should do the analog of biomimicry, looking to psychology, neuroscience, and linguistics, not slavishly copying, but taking shortcuts from how biology solved the problem.

There is no single good test of intelligence, because intelligence is not a single thing. Marcus tried to organize a Turing Olympics — many tests, not one — and hopes François Chollet will take it over. Howard Gardner's multiple intelligences, kinesthetic and verbal and the rest, are the crude starting point. What would impress him personally is comprehension. After Eugene Goostman won a version of the Turing test by pretending to be a sarcastic thirteen-year-old from Odessa who barely spoke English and ducked hard questions, Marcus wrote in *The New Yorker* that the test itself is not that good. Turing wanted to know how to make a machine intelligent. Evasion is a way to win the imitation game without doing that work.

The alternative he proposed, later written up with Praveen Paritosh for a special issue on the Turing Olympics, is a comprehension challenge. Watch an episode of *Breaking Bad* — maybe the whole series — and answer: if Walter White took a hit out on Jesse, why? Arbitrary questions about character motivation, not multiple choice, which is too easy to game. Open-ended. Why did this person pick up the stone? In a well-constructed movie, every viewer knows. The *Spartacus* scene is the book's example: a crowd stands up and shouts "I am Spartacus." Viewers immediately know they cannot all have that name, that all but one are lying, and that they are lying to protect someone and something they believe in. A machine that can say why, on arbitrary films, not a Spartacus-only special, would mean AI had really achieved a lot. Even all Westerns would impress him. Building in seven classic plots is fine. Cheating on one movie is not.

Could he be wrong, and a single neural net unlock this next year? He does not stay up worrying about that. He worries about rebranding. Deep learning will start to encompass symbol manipulation. Hinton says we do not want hybrids; Marcus thinks people will work toward hybrids and relabel them as deep learning. AlphaGo is already described as a deep-learning system when it is more correctly deep learning plus Monte Carlo tree search, a classical technique. IBM did the same with Watson: first a particular system, then anything the cognitive division built. A DeepMind researcher tried to outfox him with GPT-2 on "a rose is a rose, a tulip is a tulip, a lily is a ___" and got lily. Marcus asked whether it generalized, including to novel words. Twenty minutes later the answer was that those cases had problems. Predictions from 1998 still hold despite a million times the memory and compute. In a *Wired* excerpt they ran a children's story about a lost wallet through Talk to Transformer; the model emitted fluent text that was conceptually inconsistent with the story. Chomsky and Miller made a similar prediction in 1963. Architectures that are just layers of correlational stuff, with no built-in knowledge, are not going to solve these problems. The emperor is not naked. He is in his briefs. Speech recognition is genuinely great. The hard human problems from twenty years ago are still hard.

Fridman notes that the cynicism is entertaining and that Marcus is also a dreamer. People oversell technology in the short run and undersell it in the long run. *Rebooting AI*, written with Ernie Davis, ends with an optimistic chapter that Marcus says nearly killed Davis, who calls himself a pessimist and Marcus a contrarian. The chapter asks what would happen if AI actually incorporated common-sense reasoning and the nativism they counsel. Suitably reconstructed so we could trust it — which we cannot, now — it could be world-changing. Then the practical question: how do we build systems that align with our values?

The first thing, Marcus says, is to replace deep learning with deep understanding. You cannot have alignment with a system that traffics only in correlations and does not understand concepts like bottles or harm. Asimov's first law was, roughly, do no harm. You can quibble with the details; if we build real robots we need something like it, which means programming a notion of harm, which means abstract ideas deep learning is not good at have to be in the mix. Committees of wise people, ethicists included, should figure out what the rules ought to be. It should not be left only to software engineers or to people who own large tech corporations. Those rules then have to be translated into code, or neural networks, or something machines can work with. Right now there is no way to do that translation.

Suppose a broad, representative committee decided Asimov's first law was pretty good, with some exceptions. Someone still has to turn "do no harm" into a program. One way of taking the whole book is that we cannot do that yet, and we are fooling ourselves if we think we can build trustworthy AI when we cannot specify harm in Python or TensorFlow. If we are going to use AI for job interviews, surveillance, or medicine — not that he personally wants the surveillance — the system has to understand stuff like that. Part of why they wrote the book was to inform those future committees and to point students at the hard problems. Another part is to train non-experts who talk about AI to know what is realistic. His favorite tool in the book is six questions for any media account. Look for the demo you can try. If you cannot try it at home, maybe it does not work yet. When Google Duplex was announced, it only worked on hairdressers, restaurants, and opening hours. That is not general. Sophia has very good material written for her and does not understand the things she is saying. If a system only looks good on the distribution it was trained on, it is interpolating, not understanding. The conversation ends on guitar: Hendrix's "All Along the Watchtower," a weekend with John Lennon, and *Imagine*.