Gary Marcus

thumb|Gary Marcus at Web Summit 2022 Gary Marcus (born 1970) is an American psychologist, cognitive scientist, and author, known for his research on the intersection of cognitive psychology, neuroscience, and artificial intelligence (AI). Quotes The Algebraic Mind: Integrating Connectionism and cognitive Science (2001) Preface I am not an anti-connectionist; I am opposed only to a... subset of possible connectionist models. The problem is that... connectionism has become synonymous with... back-propogation. ...[T]hat... type of network... occupies... a small corner in the... space of possible network models. ...[A]dequate models of cognition most likely lie in a less explored part of the space... Connectionism... taken... broadly... could help us answer... what the mind's basic building blocks are and how those... could be implemented in the brain. Cognitive Architecture [S]ymbol manipulating models... use logical functions like AND and OR... [T]hose functions can easily be built in—or, implemented in—connectionist nodes. [C]onnectionism... encompasses models that have sought to explain how symbol-manipulation can be implemented in a neural substrate... I focus on... multilayer perceptrons because these are... discussed when researchers consider the relation between connectionism and symbol-manipulation. [S]ymbol-manipulation consists of three... hypotheses... [M]ind [1] represents abstract relationships between variables... [2] has a system of recursively structured representations... [3] distinguishes between mental representations of individuals and mental representations of kinds. ...[T]hese can stand or fall separately. Any given model... can be consistent with one subset of the... hypotheses... or all... I have often encountered audiences that... want me to prove... mind manipulates symbols. ...I can do no such thing. At most... symbol-manipulation is consistent with the facts and... the alternatives thus far... are inadequate. Innateness, AlphaZero, and Artificial Intelligence (Jan 17, 2018) Source: arXiv:1801.05667 One of the oldest debates in intellectual history revolves around the somewhat nebulous concept of innateness. How much of the human mind is built-in, and how much of it is constructed by experience? Plato... in... Meno... argued that a slave boy had knowledge of geometry, despite the lack of... formal training. Noam Chomsky is perhaps the scholar most identified with nativism, ...children ...born with a “language acquisition device”... or what Steven Pinker... called a "language instinct." [E]mpiricists... as John Locke.., Elizabeth Bates and... Jeff Elman... have taken something close to a "blank slate" or tabula rasa view, arguing... knowledge... from experience... through the senses. [M]odern observers... concede that genes and experience work together... "nature and nurture", not "nature versus nurture". Pinker, Peter Marler... and myself... have frequently argued... [that] a significant part of a creature’s innate armamentarium consists... of [innate]... mechanisms... that enables learning. The guiding question for the current paper is whether artificially intelligent systems ought... to be endowed with significant... innate machinery, or whether, in virtue of the powerful learning systems... recently... developed, it might suffice for... systems to work in a more bottom up, tabula rasa fashion. In... the biology of innateness... I made a foundational distinction between... "prewiring" and "rewiring", and suggested that the biological evidence for both processes was overwhelming. Gary F. Marcus,The Birth of the Mind: How a Tiny Number of Genes Creates the Complexities of Human Thought (2004) [L]earning... takes place against a background of... machinery that precedes learning; evidence for prewiring doesn’t speak against rewiring... nor does evidence for (later) rewiring speak against evidence for prewiring. [T]hink of cognition [biological or artificial] as a function over four variables... cognition = f(a, r, k, e) where a = innate algorithms.., r = innate representational formats.., k = innate knowledge.., and e = experience. ...A true blank slate would set k and r to zero.., a to some extremely minimal value.., and leave the rest to experience. This is essentially the view that Locke... articulated... Yann LeCun... appears to believe that... a, r and k should approach zero. ...[F]ew researchers ...explicitly discuss the question. [N]ativism is hardly a central topic in contemporary AI thought. Perhaps the strongest argument for keeping the values of a, r, and k small, while relying on a high value of e, comes from DeepMind’s groundbreaking work on playing classical board games through reinforcement learning, masterfully presented by Demis Hassabis... Genes and Language (2011) Gary F. Marcus and Simon E. Fisher in The Cambridge Encyclopedia of the Language Sciences (2011) pp. 341-44. A Source from MPG.PuRe, publication repository of the Max Planck Society Digital Libray [T]he question is not whether genes play roles in language but how... [N]eural systems that must be biologically constrained... include mechanisms that... simultaneously coordinate SYNTACTIC, SEMANTIC, PHONOLOGICAL, and PRAGMATIC representations with one another, with motor and sensory systems, and with both speaker's and listener's knowledge of the world. Genes... contribute to online functions of the mature brain... [e.g,] by mediating changes in the strengths of connections between neurons. ...[L]ikely ...hundreds or ...thousands ...participate in the development and maintenance of the neural systems that underlie language ... [G]eneticists have... been able to zero in on a... gene, FOXP2... implicated in one particular disorder affecting speech and language... [M]utations have been heterozygous... [i.e.,] a mutation in only one copy... (Humans are diploid organisms carrying two copies of every gene, one from the father, the other from the mother, with... exceptions such as the genes on sex chromosomes.) ...[C]ases of FOXP2-related disorder suggest ...reduced ...FOXP2 protein in the brain circuits... important for speech and language. [M]ost genetic language disorders are likely... multifactorial, the product of multiple genetic risk factors... [M]utation of FOXP2 impairs... also aspects of nonlinguistic orofacial motor control... [I]t is highly unlikely that... a single human-specific gene... endow[s] our species to acquire language. ...The "language gene" shorthand is... misleading because most, if not all... genes... involved in language... likely play other roles... For developmental dyslexia... [∃] several candidate genes (DYX1C1, K1AA0319, DCDC2, and ROBO1). ...[H]ighly similar versions of each are found in other primates and in rodents. Genes, like species, are the product of what Darwin called "descent with modification." Each... current function a modification of earlier functions. ...[C]omparisons of gene sequences and expression patterns in different species can... cast light on language evolution... FOXP2... is present in similar forms in... vertebrates, including mammals, birds, reptiles and fish... expressed in corresponding regions of the brain to... humans... [I]n songbirds, changes in expression of the gene in striatal Area X,—a key nucleus of the brain system involved in song learning—appear to relate to alterations in vocal plasticity... Mathematical analyses of [diverse human] genomic sequences suggest... the version of FOXP2... in modern humans arose within the last 200 years, concordant with... estimates... of proficient spoken language. [G]enes may provide the first molecular entry points into neural pathways involved in human communication, and to understand how... descent and modification led to the... human faculty for complex language. The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence (2020) Source: arXiv:2002.06177 Recent research in artificial intelligence and machine learning has largely emphasized general-purpose learning and ever-larger training sets and more and more compute.In contrast, I propose a hybrid, knowledge-driven, reasoning-based approach, centered around cognitive models, that could provide the substrate for a richer, more robust AI than is currently possible. Abstract [I]f we cannot count on our AI to behave reliably, we should not trust it. [C]ontrast robust AI with... narrow intelligence, systems that perform a single narrow goal extremely well... but often... extremely centered around a single task and not robust and transferable to... modestly different circumstances [P]ointillistic intelligence... works in many cases but in fails in many other[s].., ostensibly quite similar, in somewhat unpredictable fashion. To build a robust, knowledge-driven approach to AI we must have the machinery of symbol-manipulation in our toolkit. Too much of useful knowledge is abstract to make do without tools that represent and manipulate abstraction, and to date, the only machinery that we know of that can manipulate such abstract knowledge reliably is the apparatus of symbol manipulation. We cannot construct rich cognitive models in an adequate, automated way without the triumvirate of hybrid architecture, rich prior knowledge, and sophisticated techniques for reasoning. Gary Marcus - Taming Silicon Valley (Sep 4, 2024) | Starmus VII, Slovakia 2024. A Creative Commons video from the Starmus YouTube channel. [A] question I think about every day... is, "Is AI, on balance, going to be good for society or bad for society?" ...What kind of AI should we want? ...[W]e should want one which is consistent with human rights and human dignity. ...UNESCO ...laid out its first global standard on the ethics of AI, and ...talked about human values; respecting, protecting and promoting human rights ...[etc.] ...[T]hat's a really good place to start. Ref: UNESCO 41st session, Recommendation on the ethics of artificial intelligence (Nov 2021) [T]he place we actually are [at] has much more to do with surveillance capitalism... There's a wonderful document... "The Blueprint for an AI Bill of Rights"... from the Biden White House that talks about safe and effective systems. We should all want that. We shouldn't want algorithmic discrimination... [etc.] Ref: Blueprint for an AI Bill of Rights: Making Automated Systems Work for the American People (Oct 2022) It's... obvious what we want, but... the AI that we have... especially the AI that is popular.., generative AI, is... technically and morally inadequate. [G]enerative AI... is just one possible approach. It's... fun to play with. It's... popular, but... has a lot of problems. I... call it "rough draft AI". You can trust it to make a rough draft of anything.., but you can't really trust it. It has a factuality problem.., [e.g.,] it's... rotten at physics... [I]t's factually challenged.., and it's presented as absolute fact, as absolute truth. There's a saying in the military, "Frequently wrong, never in doubt," and that's... what we've got... Large language models... put together... superficial statistics of the words and sometimes the images that they see. They may seem smart, but... they can't sanity check their... work. They can't.., like... an intelligent person... go to Wikipedia... or Google... to... sanity check. It all sounds very plausible. When I went on 60 Minutes last year I called it all "authoritative bullshit" Ref: The new world of AI chatbots like ChatGPT @CBS News. "Gary Marcus: I actually like to call what it [ChatGPT] creates "authoritative bulls***." (LAUGH) It blends the truth and falsity so finely together that, unless you're a real technical expert in the field they're talking about, you don't know." There's also toxicity problems with these systems... [E]ven with so-called guardrails, it's very easy to get these systems to say things that are truly vile. Generative AI... has a bias problem. ...[Y]ou can type in something like "leadership" and get a bunch or white males leading a bunch of white males... [A] paper... last week showed that if you use these systems to evaluate job resumes, there's lots of discrimination. These things are... arguably worse... than the history of discrimination in our society, because they tend to perpetuate stereotypes from the data, and they can't reason in the way that we can about what would be ethical to do now... Generative AI... has a plagiarism problem. ...The New York Times noticed ...these systems often just repeat verbatim things that they've seen before. ...They're trained on a lot of copyrighted material, and they're very prone to repeating it. I did some work with an artist Reid Southen... and we showed that these systems... plagiarize in the image domain and now we know it's true in the musical domain... [A]ll of these... [images] came out of Midjourney. They're... repeating, spitting back things that they've seen in copyrighted materials. ...There is no embarrassment or shame in these systems, or the people that are making them. Ref: 1) Gary Marcus, Reid Southen, "Generative AI Has a Visual Plagiarism Problem" (Jan 6, 2024) @IEEE Spectrum 2) About Reid Southen @ReidSouthenArt.com Generative AI... has a personality problem. ...We tend to anthropomorphize them too much, but for the sake of argument, if you did a psychoanalysis.., GPT models are not a picture of good mental health... [T]hey... exhibit low self-esteem. Think of all the apologizing they do. ...They have a disconnection from reality.., and... they display severe psychopathy. Like... somebody asked about schools and the Meta AI referred to the experience of its children. ...[O]nly a sociopath tells you about the experience or their children... if they... don't... have children. So these systems are not really what we want. [T]his is from my forthcoming book, Taming Silicon Valley, and I go through a dozen different immediate risks. ...I'm most worried about... disinformation. We have 70 elections around the world. The election here in Slovakia was at least influence by misinformation at the last minute. Deliberate disinformation, deepfakes, text that is generated automatically by these machines. Even though these machines aren't very smart, they can easily mimic human styles, and because they can't fact check, that makes them... perfect tools for disinformation. So there is a real chance that one or more democracies will be disrupted this year. Overhead: The Biggest Immediate Risks of Generative AI: Disinformation, Market Manipulation, Accidental Misinformation, Defamation, Nonconsensual deepfakes, Accelerating Crime, Cybersecurity and Bioweapons, Bias and Discrimination, Privacy and Data Leaks, Intellectual Property Taken Without Consent, Over-reliance on Unreliable Systems, Environmental Costs. That's my biggest concern, but there are many others. There's market manipulation, accidental misinformation, there's all these fake books... [etc.] now... [D]efamation has happened a bunch of times. There's nonconsensual deepfake porn. There's accelerating crime, cybersecurity.., bias and discrimination.., privacy and data leaks.., stealing intellectual property. There's also the problem that people over-rely on unreliable systems because they hear that these systems are great, when they're not... Fake Books Ref: 1) Jane Friedman, I Would Rather See My Books Get Pirated Than This (Or: Why Goodreads and Amazon Are Becoming Dumpster Fires) (Aug 7, 2023, & Update, Feb 18, 2025) @JaneFriedman.com 2) The Authors Guild, AI Is Driving a New Surge of Sham “Books” on Amazon (Mar 15, 2024) @authorsguild.org 3) Kbilz, Buyer Beware: Identifying AI Slop Books (Oct 24, 2025) @kentonlibrary.org 4) Elizabeth Blair, How an AI-generated summer reading list got published in major newspapers (May 20, 2025) @NPR.org [T]here are huge environmental costs... because now there may be serious long-term risks. ...[W]e could have real catastrophe from AI, and we should be worried. ...We already have a lot of AI risks, whether or not we get to anything... existential, or if AGI poses new risk. We already have enough to worry about... with generative AI. ...Generative AI has not proven itself to be a net positive for society. There are some arguments, maybe that it's produced some positives.., [e.g.,] it has some computer programmers program faster, but there are... serious risks. I would not want all the computer programmers... to program faster if we're going to lose some democracies. Because there are many risks, we shouldn't expect a silver bullet... We shouldn't expect any single law is... going to handle all of this. We need to look at each of the risks. Generative AI is almost like a Hydra. New risks show up every day. ...[E]xisting laws were written long before large language models.., [e.g.,] our defamation laws... care about intent. These systems can defame people without... intent... That doesn't mean that we should exonerate them. [A]gility is key. As these models change, we may need to change the laws relatively quickly, which means... [e.g.,] the U.S. Senate is not the ideal system for handling these things. We need transparency. We want a full accounting of what data is used to train the models because the biases... come from how they're trained... [I]f we want to do the scientific work of mitigating the risks, we have to know what data goes into the systems. They're all black boxes.., [i.e.,] we don't really understand them. If we don't understand the data in, we have no idea how to get to the data out, and it's hard to do anything about it. [I]f we're going to deploy something to 100 million people we should do something like what the U.S. FDA does, which is they say, "Tell us what the benefits are.., what are the risks, and why is this worth it?" We... need to look at systems after the fact. We have to audit them and say, "How is this working out?" ...[W]e have to make the job of the scientist... easy. We have to force the companies to give us the data that we need... [e.g.,] "Are these systems discriminating against people in employment?" ...[T]he companies don't want to share that, but we have to make them... We need to have liability laws. ...[I]n the United States we have ...section 230 that allows the social media companies to basically publish anything they want without liability, even if they hype things that are misinformation because that gets them more clicks. We can't do the same thing in AI. Also see: Communications Decency Act. We need layered oversight. ...[W]e make the commercial airlines so astonishingly safe ...because we have rules about how you can design.., manufacture.., maintain a plane. We have systems for investigating accidents... [etc.] We should think about AI like we think about aviation. There's a lot potentially at stake. ...[S]cientist have to be in the loop ... We... need national and global AI agencies. ...[M]any senators from both parties were supportive of many of these things! I was astonished that people were with me... but so far, in a year, not one of these ideas has... been put to a vote in the United States... The EU has made real progress in regulating AI. The United States has not. [T]hat brings me to money and context. ...[M]ost of my career I worked on research. I built AI companies. I didn't really think about politics that much... [T]he interest of... technologists do not necessarily align with the interests of humanity. ...[T]here was this cartoon... in The Economist just after the UK AI Summit... [Y]ou have the different countries, but it might as well be the different companies saying, "We declare that AI poses a potentially catastrophic risk to humankind... and I cannot wait to develop it first." So... none of them are slowing down. We... shouldn't let big... tech companies decide everything for humanity... [e.g.] a very hard question is, "Should we open source AI?" Nobody's quite doing that... They're open sourcing the weights of the model and not the... source code. ...Some ...like Yann LeCun at Meta think this is a fabulous idea. We'll move faster if everything is open... [O]thers like Jeff Hinton are saying... maybe that would make it easier for bad actors to produce bioweapons, for example. ...It's one of the hardest questions.., but it shouldn't be up to Mark Zuckerberg to decide... [W]hat's happening is exactly that. ...The rest of us should have a vote on that. There is a huge gap between what big tech companies say about responsible AI, and what they... do. ...[T]here was a news item that [OpenAI and Microsoft] had given $2 million to warn people of the potential harms of deep fakes... [Their] collective valuation... is $3.1 trillion and they gave $2 million, so we should all give them a very slow clap... [T]he amount that most individual artists have received for... their work... [that the AI's] have been trained on... is zero. [T]here's been zero disclosure of what copyrighted materials have... been used to train these systems. It's all just been seized. [T]his cartoon... applies not only to climate change, but possibly to AI... "Yes, the planet got destroyed but for a beautiful moment in time, we created a lot of value for shareholders." Sam Altman... said almost... the same thing except he wasn't kidding. He said, "AI will most likely lead to the end of the world... But in the meantime there will be... great companies created." This is not a lot of solace... [W]e shouldn't trust Washington... Maybe the worst thing would be no regulation, but the second worse thing would be regulatory capture where the companies write the rules. That's where we're headed, at least in the United States... [e.g.,] there was a meeting... Chuck Schumer... held... [I]t was all the CEOs of the big companies... invited, and Marietje Schaake... wrote about it... "Imagine convening about how to legislate for CO2 reductions with the CEOs of Chevron, Aramco, Shell, Exxon, BMW, Ford, Tata, BP, oh and a Greenpeace activist" ...[T]hat would be regulatory capture, and that's what happened. The Department of Homeland Security just released a new board on [AI] safety and security... They put the CEOs of OpenAI, Anthropic, Google, Microsoft, Nvidia... [etc.] ...[T]hat sure does look like regulatory capture... We really have to worry about that, especially in the United States where... Citizens United, a Supreme Court decision.., allows essentially unlimited amounts of money from the corporation to flow to the senators. Ref: Notice by the Homeland Security Department, Establishment of the Artificial Intelligence Safety and Security Board (Apr 29, 2024) @federalregister.gov [T]he only way we're going to get to good AI policy in the U.S. is if people organize and demand it. That's why I wrote... Taming Silcon Valley. [It] is to try and get people to realize how much is at stake here. ...[W]e face ...the same challenges ...apathy and so forth, that climate change faced for a long time. We need to move fast, and we should learn from that situation. As we develop AI policy we need to be realistic about the limits of current AI... [i.e.,] we have to know what's going on... [S]ometimes I feel like the people making the rules are seduced by the companies... [e.g.,] the Chip War... the U.S. is withholding chips from China. ...This could lead to military conflict ...and it's all predicated on the idea that if China got GPT-5 first it would be a really big deal. ...[W]hat would they do with it? Write more boilerplate rough draft text faster than we can..? This is not going to war over! Building a foreign policy, or building our AI policy around fantasy is not a good idea, but what Silicon Valley has learned is if they hype stuff, the stock goes up... [T]hey've learned this over and over and over. ...Elon Musk ...keeps playing the hype card with driverless cars. It... goes back to Google even earlier, but Musk said... we would have driverless Teslas that would drive from LA to New York City by 2017. ...[I] still hasn't happened. He promised a fleet of a million robotaxis by 2020. It hasn't happened. He promised that... he would mass produce them by 2024. It hasn't happened. But it always drives the stock price up, so we get a misleading representation of what's actually possible. Sam Altman was fired and everybody started talking about this algorithm Q... Ilya Sutskever had seen it and it was going to change the world. I don't that's actually gone anywhere yet. Ref: Q-star (Q*): What it is and what it could mean for the future of AI @superannotate.com Driverless cars are hard... [e.g.,] the company Cruise had more people in teleoperations centers than... cars on the road, which should remind you of the Wizard of Oz, "Pay no attention to the man behind the curtain!"... [I]t might remind you of some other stories where people tried to fake it till they make it... [W]e've reached a point of diminishing returns for the current paradigm. Large language models... are... fun to play with but there are these myths... that they're getting better every few months. ...Ethan Mollick has a book on the bestseller list. He claimed that the current best estimates of the rate of improvement in large language models show capabilities doubling every... 5 to 14 months. ...[H]e misread the study that he was citing ...[I]t actually looks like things are slowing down. Everybody's talking about an exponential curve... [i.e.], that the trend... on the left part of the graph would continue... doubling and doubling and doubling, but they're not anymore. They haven't been doubling for the last 14 months. ...[A] big part of the hype is this idea that if we just keep doing what we're doing.., just put in more data, and if you give us $7 trillion... we will magically create artificial general intelligence... We're not really in the regime of exponential improvement anymore. More likely we're in a regime of diminishing returns... probably because for a while you can get more and more high quality data by harnessing the internet. But now all of these companies are basically using all of the internet. They're doing all kinds of dubious things to get every last morsel of data, and they just don't have that much interesting data left. Ref: Ethan Mollick, Co-Intelligence: Living and Working with AI (2024) They're talking about things like synthesizing data. Well I heard that with driverless cars for years. It still hasn't made driverless cars that we can trust. It could... be that even though OpenAI looks... invincible... that they aren't really. ...Apple was the last to smartphones.., Google was later in the search engine space. I wouldn't assume either that OpenAI... or that large language models are going to win this race. [B]etter AI is... possible. ...I am skeptical of things that aren't true. ...I'm a scientist. I spent a quarter century being a scientist before I moved full time into AI. ...I ...want AI to succeed. That's why I built a couple of AI companies. That's why I've studied it since I was 8 years old. To get there... we're going to need to take some clues... from the human mind. We have a field... where people who are good at... matrix math have dominated, and... ignored everything else. thumb|Gary Marcus: Generative vs. Symbolic AI analog to Kahneman's Two Systems of Cognition, Starmus VII (2024) [T]here's a wonderful quote from... Brian Scholl and Chaz Firestone, "The human mind is not one thing, but many." ...[W]e do a lot of different things, and the AI we have... is good for some.., but not all of them. It's not a well-rounded cognitive system. ...[T]he late Daniel Kahneman's two systems of cognition..: System 1... was reflexive, intuitive, data driven ...[etc.] System 2...is deliberative and reasoning based... [etc.] ...That's in humans. In AI we have analogs to those systems. A generative AI comes pretty close to... System 1. It sucks down a lot of data and it forms something... like intuition.., [C]lassic symbolic AI... looks a lot more like computer programming.., is more deliberative and reasoning based and grows out of a tradition of formal logic. ...[T]hey ...each have their strengths and weaknesses. Ref: Chaz Firestone, Brian Scholl, "Cognition does not affect perception: Evaluating the evidence for 'top-down' effects" (2016) where it is stated, "There is no one way the mind works, because the mind is not one thing. Instead, the mind has parts, and the different parts of the mind operate in different ways" Generative AI is really good at learning. It's pretty poor at reasoning. You can't really count on it. It's data hungry. It needs... the entire internet. It's uninterpretable... [i.e.,] it's a black box. We don't know what it does. It's prone to hallucinations... On the positive side it's scalable. You can just make it better and better to some point... by adding more data without doing a lot of hard work, and that's very appealing. ...[I]t's grounded in the use of language rather than reality, which... is problematic. Classical AI is almost the opposite. It's good at reasoning, and not very good at learning. Nobody's been really able to learn the rules that classical AI depends on at any kind of scale. There's some toy models, but nobody can really do it well. It's data efficient on at least some kind of tasks, maybe not all. (Maybe that was too charitable...) It's interpretable. You can understand what it is doing. It doesn't hallucinate in the same way. It doesn't just make up things... [e.g.,] if it consults a database, it just returns what's in the database. (Imagine if you had a new database system... and it made a mistake 1 in 10 times like a large language model... People would laugh in its face. They would never get any commercial traction...) On the other hand it's mostly handcrafted, and that's a pain, and on the positive side it's grounded in facts. They're almost complimentary, but neither approach has ever worked on its own. ...People used to complain that symbolic AI was brittle. You would give it a slight variation on a question and it would break. That's true of the new neural network approach, the generative AI approach. We just don't know how to make large scale AI. We don't know how to make driverless cars that work. We don't know how to systems that really read with comprehension. There are things we can do, like we make navigation systems work pretty well... [etc.] But AI is basically an unsolved problem, no matter what the Sam Altmans of the world might want you to think. [H]ere's the trick. The human combines both. That's what Daniel Kahneman's Thinking Fast, Thinking Slow was about. We use both systems. We don't always use them perfectly, but we... balance those... It's kind of a checks and balance that allows us to get through life reasonably well. ...I wrote a book ...Kluge which is about all the times it fails. We're not perfect. We're kind of a low bar, but could do better... [W]e do combine both, and that allows us to get along in the world, to solve new problems we haven't seen before. Ref: G. F. Marcus, Kluge: The Haphazard Construction of the Human Mind (2008) AI has been trying for a long time to make do with one or the other with poor results, and that's a fact about sociology and money. There have been two traditions for... 60 some years, and people in the academy... fight over money, and graduate students, and prestige... [T]hese two fields have been fighting for all of those things for so long that they've forgotten that maybe that's not the right thing to do. [I]f we can figure out how to build an AI that combines the best of both worlds, we can create an AI that's learnable, data efficient, interpretable, reliable, verifiable and grounded in facts, and one that's capable of reasoning about human values, like don't discriminate against people, or don't harm people, which current AI just can't do. A few companies own System 1.., they own the technology and the GPUs to run it... [etc.,] but it's not too late to develop a new approach to AI that can be a public good, working in the public interest. AI Oligopoly Ref: 1) Tejas N. Narechania & Ganesh Sitaraman, An Antimonopoly Approach to Governing Artificial Intelligence (Fall 2024) Yale Law and Policy Review, Vol. 43, Issue 1 @YaleLawAndPolicy.org 2) Shaleen Khanal, Hongzhou Zhang, Araz Taeihagh, Why and how is the power of Big Tech increasing in the policy process? The case of generative AI (Jan, 2025) Policy and Society, Vol. 44, Issue 1, pp. 52–69 @academic.oup.com/ Oxford Academic. [W]hat... we need... is something like a CERN for AI, where we have a global project where people work together to make a public good, a form of AI that everybody can use, that perhaps combines the two systems... It is not driven by commercial imperatives like "Let's get all the data and surveil people," but instead is motivated by trying to make AI that is good for medicine.., for neuroscience... [etc.] My mission is to try to tame Silicon Valley. ...The point is to try to ensure that AI works for all of us, to get the general public involved... [T]he issues with AI parallel those with climate change to an incredible (and maybe somewhat depressing) degree. In both cases we need to act with urgency. With climate change we can... plot how much time we have, with AI we can't because we don't really know how fast we're going to get to more dangerous forms of AI. We don't know how fast democracy is going to be undermined. We don't have the same kind of mathematical models, but we... don't have a lot of time. In both cases there's a limited time window for action. ...[E]ven if we figured out what to do 10 years from now, that might be too late! We can't in either case count on the government to do the right thing because of all the exigencies of money... [etc.,] unless people speak up and say, "These things matter to us! AI policy matters." I wrote this book in a frenzy to get it out... before the election because I want people in the U.S. to talk about it and realize how important it is. ...[A]nother thing that's really important in both cases is that grassroots action is essential. If anyone wants to teach me about grassroots action, I'm very much all ears right now. The choices that we make now about AI will have lasting effects, possibly for decades, even for centuries. We've got to get this right now and build an AI that works for us! Interview: Cognitive Scientist Gary Marcus’ Lifelong Disillusionment with A.I. (May 31, 2025) by Alexandra Tremayne-Pengelly @observer.com Source It’s a very specific moment. I learned how to program on basically a simulation of a computer that was on paper. It was part of my after-school program for gifted kids, and that night I explained to the media how it worked. So, my media career and my A.I. career began that day. I didn’t know too much about, but they thought it’d be fun to have a human interest story. They nominated me to do the explanation, you know, cute kid explaining what a computer is. And I saw it on TV that night. There’s no archival footage that we can find, but I did actually see it on my father’s little black and white TV later that night. I’ve never been fully optimistic about it. I don’t think we’ve made as much progress as I’d hoped that we would. I would have guessed that 40 years after I was looking at this, that a lot of A.I. would have been solved problems. We have a lot of tools now that are very useful. But I think the bigger questions of, like, how do you represent and acquire knowledge? You still don’t have really good answers to those. So I’ve never reached a point of being really satisfied. I always want it to go better. I started writing about it right away in my Substack and I said immediately that it’s going to excite people but it has limits. External links Gary Marcus articles @Medium.com Marcus, G articles @arXiv.org Category:Cognitive scientists Category:Academics from the United States Category:1970 births Category:Living people