Steven Pinker: Conhecimento comum, redes sociais, IA, pensamento crítico

Data: Wed, 06 May 2026 04:45:00 +0000 · Duração: 00:36:58

Clica em qualquer parte da transcrição para saltar para esse momento no vídeo.

Steven Pinker 00:00:00.00People left to their own devices, just the default, is they don't have very high standards of accuracy. 00:00:07.54People are happy with myths and conspiracy theories and legends and stories that make their side look good and the other side look bad. 00:00:18.68When legacy media were new, when there was first cheap newspapers, all the news was fake news. 00:00:30.00It would take the equivalent of 30 million years for a child to process as much information as a GPT is trained on and the child can speak fluently in three years, not 30 million years.

José Maria Pimentel 00:00:43.76Olá, o meu nome é José Maria Pimentel e este é o 45 Graus. 00:00:47.56Duas notas rápidas antes de irmos ao episódio de hoje. 00:00:50.74Como alguns de vocês sabem, começa em meados de maio a segunda edição da pós-graduação em pensamento crítico e tomada de decisão no Iskete Executive Education. 00:00:59.78Se tiverem curiosidade, enfim, procurem online e vão chegar à pós-graduação. 00:01:03.72A segunda nota é que estreou há poucas semanas o Debatível, um projeto que alguns de vocês também conhecerão, que foi um projeto criado por mim juntamente com a Leonor Caldeira e o João Marecos. 00:01:16.50São debates ao vivo, debates competitivos entre duas equipas em torno de uma moção. 00:01:21.64Segue um chamado modelo tipo Oxford e combina competição com entretenimento. 00:01:26.42São debates muito desafiantes de assistir. 00:01:29.36O primeiro foi há duas ou três semanas sobre o impacto da inteligência artificial no mundo do trabalho. 00:01:35.42O segundo vai ser já em maio, sobre o sistema prisional. 00:01:38.58E o terceiro vai ser em junho e é sobre a crise do jornalismo e se a solução para essa situação é o Estado apoiar diretamente o jornalismo. 00:01:47.24Uma proposta que já anda por aí há muitos anos, não vai desaparecer e que gera opiniões muito diferentes. 00:01:52.92Os bilhetes para o segundo debate já estão esgotados, mas conseguem ainda encontrar para este terceiro debate em junho na Ticket Line e, como se costuma dizer, nos locais habituais. 00:02:01.82E agora ao episódio de hoje. 00:02:03.04O convidado é alguém que tenho muito orgulho de conseguir trazer ao Conselho 85 Graus. 00:02:06.40É ele, Steven Pinker, um dos meus autores preferidos e uma pessoa que já estava a tentar trazer ao podcast há muito tempo. 00:02:13.34Para quem não sabe, Steven Pinker é psicólogo cognitivo, linguista e um dos mais influentes intelectuais públicos dos Estados Unidos. 00:02:20.94É professor na Universidade de Harvard e destacou-se pelo trabalho sobre a linguagem, mente humana e natureza humana. 00:02:27.76E, mais recentemente, daí o intelectual público, por tomar partido em vários debates, sobretudo a defesa do que ele chama os valores do iluminismo, como a razão, a ciência e o progresso. 00:02:37.40É autor de vários livros, best-sellers, e lançou recentemente um livro que acaba de ser editado em português com o título Quando todos sabem, que todos sabem, reticências. 00:02:46.72É exatamente com este livro que abrimos a nossa conversa. 00:02:49.08O tema do livro é aquilo que o convidado chama conhecimento comum. 00:02:52.16Ou seja, aquilo que todos nós sabemos que os outros sabem e eles sabem que nós sabemos. 00:02:56.28E isso, como ele explica, é essencial para nós coordenarmos uma série de comportamentos complexos em sociedade, desde conduzir na estrada até confiar nas instituições políticas. 00:03:06.16Isso leva-nos a outra pergunta, que é o que é que acontece quando este conhecimento se começa a fragmentar, como tem acontecido em muitos países, em resultado, não só, mas em grande medida, das redes sociais. 00:03:17.60Perguntei ao convidado a opinião dele sobre isto e o que é que ele acha que pode ser um futuro ecossistema informativo em que nós conseguimos, de certa forma, domar as redes sociais, mas conseguindo continuar também a ter os benefícios que elas nos dão. 00:03:30.38Falámos também, como não poderia deixar de ser, sobre a inteligência artificial. 00:03:33.72Perguntei ao convidado se estes novos modelos de linguagem, sobretudo para alguém que é, não apenas um psicólogo, mas também o linguista, se o surpreenderam e também, ao mesmo tempo, qual é a posição dele sobre este debate que tem se tido cada vez mais, sobre se estamos a usar estes modelos em neumazia e isso está a atrofiar o nosso pensamento crítico. 00:03:51.40Foi um episódio curto, para o que é habitual aqui no 45 Graus, mas como vou ver, é um episódio muito denso em conhecimento. 00:03:57.80Espero que gostem e voltamos a ver-nos no próximo episódio. 00:04:00.46Até lá.

Steven Pinker 00:04:00.74O café do teu hotel é incrível? 00:04:17.82É Nespresso? 00:04:18.70Já conheces os planos de soluções chave na mão? 00:04:20.56Não, o que inclui? 00:04:21.46Escolhes a máquina e o café. 00:04:22.72O resto tratam eles. 00:04:23.66Entregas, reciclagem e assistência técnica. 00:04:26.42Hum, perfeito para o meu escritório. 00:04:28.44Café sem complicações. 00:04:29.58Nespresso Professional. 00:04:31.28Soluções chave na mão. 00:04:32.50O plano de café ideal para o seu negócio. 00:04:34.24Saiba mais em nespresso.com.br

José Maria Pimentel 00:04:35.74I'm here with Professor Steven Pinker. 00:04:39.80Steven, welcome to 45 Graus. 00:04:41.88Thank you. 00:04:42.70And before we start, let me thank you for making the time. 00:04:46.16I mean, I know I have to turn down some requests for interview. 00:04:51.36You get, I imagine, ten times that, so I really appreciate that you're making the time. 00:04:56.98Let me start with your new book. 00:04:58.60Sure. 00:04:59.42Which just came out in Portuguese. 00:05:01.44Tell me a bit about it. 00:05:03.30It's one of those books which is less self-explanatory than the previous ones, I would say. 00:05:10.34So it seems to convey an apparently mundane aspect about the importance of common knowledge. 00:05:16.34But if I know you, I imagine that there's much more underneath that.

Steven Pinker 00:05:20.80Yes, the title in English is when everyone knows that everyone knows, dot, dot, dot, common knowledge and the mysteries of money, power and everyday life. 00:05:34.28So common knowledge is a technical term from mathematical game theory, also used in linguistics and philosophy and economics. 00:05:47.74So common knowledge is a technical term, but I know that I know that I know that I know that I know it. 00:06:01.20And so I'm adding finagum. 00:06:02.74That's why there are the three dots. 00:06:05.24Now, why is this significant? 00:06:07.20Well, it's because common knowledge is necessary for coordination, for people to be on the same page. 00:06:14.12That is, to each do something that benefits them as long as the other person does the same thing or a complementary thing. 00:06:22.16So if you have a potluck dinner and you want to bring, you want one person to bring the main course and another person to bring the dessert, they need common knowledge. 00:06:32.54It's not enough to know that I'm going to bring meat because the other person might think I'm bringing dessert and so he'll bring meat. 00:06:41.08Even if I know that he's bringing dessert, he might think that I think that he's bringing meat. 00:06:46.10And so he'll bring meat after all. 00:06:48.84Only if each one of us knows that the other one knows that we know can we do both do the thing that benefits us both. 00:06:54.82Now, this applies in many conventions of everyday life, driving on the right versus driving on the left. 00:07:02.06There's no reason to do one or the other, but it's very important that everyone drive on the same side, whether it's the right or the left, which they can only do if they know that everyone else knows that everyone else knows, but that's the side you drive on. 00:07:14.58It's responsible for the fact that we can use paper money. 00:07:19.38Why do I accept a piece of paper in exchange for something of value? 00:07:23.60Well, because I know that someone else will accept it in exchange for something value they'll give me. 00:07:28.36Why do they do that? 00:07:29.26Because they know that someone else will accept it. 00:07:32.30And so money has value only because everyone believes that everyone else believes that it has value. 00:07:38.70And if people stop believing in it, the money can become worthless, as in hyperinflation. 00:07:43.52Yeah, exactly. 00:07:44.28Just like bank runs, where people withdraw their money from a bank, not even because they think the bank is unsound, but if they worry that other people worry that still other people, everyone might withdraw their money because they don't want to be the last one. 00:08:00.22Political power depends on common knowledge. 00:08:03.28No government can intimidate every last one of its citizens. 00:08:07.68Citizens can always overrun a government, but only if they all act at the same time. 00:08:14.40So they all have to know that everyone else is discontent with the government. 00:08:19.44They all have to know when the time is to rise up and oppose the government. 00:08:24.58That's why totalitarian governments prohibit freedom of the press, which is a way of generating common knowledge, and freedom of assembly, where people get together and everyone can see everyone else and they know that they are united in their opposition. 00:08:39.04Finally, a lot of the book is on everyday life, and that's how I came to the concept, because I am a psychologist. 00:08:45.84I'm interested in what people think, how they interact. 00:08:50.96So I suggest that a lot of the rituals and hypocrisies and puzzling aspects of everyday life can be understood as ways that we generate and prevent private knowledge from becoming common knowledge. 00:09:05.98When we use euphemisms and innuendo, instead of just saying exactly what we mean, phenomenon of language, which is my primary interest, I suggest it's that sometimes we want to get a message across. 00:09:19.92We don't want the other person to know that we know that they know what the message is. 00:09:24.98And so that's when we say things indirectly or in euphemism, such as in bribes. 00:09:33.54Gee, officer, is there some way we can settle the traffic ticket here without going to court and doing all that paperwork? 00:09:40.62Veiled bribe, veiled threat. 00:09:44.20Nice story you got there would be a real shame if something happened to it from a mafiosi. 00:09:49.74Veiled sexual come on, would you like to come up for Netflix and chill? 00:09:54.74All of these are ways in which we get our meaning across, but there is deniability as to whether we know that the other person knows that we know that they know the meaning. 00:10:04.84And so we can preserve our social relationships. 00:10:08.14Our social relationships are themselves matters of common knowledge. 00:10:12.56What makes two people friends? 00:10:14.10Each one knows that the other one knows that the first one knows that the second one knows they're friends. 00:10:18.04That's what friendship is. 00:10:19.50Or romance, or a dominant and a subordinate, or two transaction partners. 00:10:27.32The relationships are held by common knowledge.

José Maria Pimentel 00:10:30.20Anyway, that's a quick overview of my new book. 00:10:33.10And I find your thesis relevant to a wide range of aspects, but one of them is coordination, as you mentioned in the beginning. 00:10:41.74And it's interesting to notice that, for instance, let's take the example of traffic. 00:10:45.50We find that in different countries there are different equilibria, and some are better than others, in the sense that you can have one equilibria in which the traffic is kind of chaotic, and it sort of works, but it's much better if everyone follows the rules, right? 00:11:00.42Do you have a view on how these different equilibria emerge? 00:11:05.60Does this inform our view on how different equilibria emerge, be it in the traffic or in social trust, like the trust people have on strangers, for instance, which also varies among countries?

Steven Pinker 00:11:19.02The reason that the branch of math matters called game theory is relevant is game theory is about situations in which the rational choice depends on what the other person does, and vice versa. 00:11:34.68So you can't just look up which of my decisions will give me the greatest benefit, because it depends on the decisions that the other guy makes. 00:11:45.40So if in a situation where you have two people, the outcome for each one depends on the other one, then there's no correct answer often, but there can be an equilibrium. 00:11:57.00An equilibrium is when one choice, regardless of what the other one does, you're best off sticking with that choice, and so on, the same thing for the other guy. 00:12:09.50Sometimes called a Nash equilibrium, after the mathematician John Nash, who's the subject of the Hollywood movie, A Beautiful Mind. 00:12:20.22I'll just call it an equilibrium. 00:12:22.32Now, a coordination game is when there are two equilibria, both of which work. 00:12:28.28That is, if both parties make a certain choice, neither one has an incentive to switch their decision based on as long as the other guy sticks to his. 00:12:41.50So driving on the right or driving on the left is an example. 00:12:45.88If everyone else is driving on the left, my best option is to drive on the left. 00:12:50.78And that's true of every driver. 00:12:53.38But how does it start, right?

José Maria Pimentel 00:12:55.34How do they start? 00:12:56.44It might be imposed top down from the government, but otherwise, why would it start driving on the left?

Steven Pinker 00:13:01.60Yes, in that case, it's imposed by the government, and sometimes it can switch. 00:13:06.62In 1969, the government of Sweden decided to, which previously had driven on the left, this was inconvenient for cars driving into Norway and Finland or taking the ferry across the bridge to Denmark. 00:13:23.22And to harmonize with the rest of Europe, they decided to make the switch. 00:13:26.84The crucial thing is they couldn't just say, well, whenever it's convenient for you, whenever you are ready, make the switch, that would lead to chaos. 00:13:34.98So they picked a particular day. 00:13:37.38I think it was December 31st, New Year's Eve, midnight, I think 1969, and everyone switched at the same time. 00:13:47.22They were prepared for it. 00:13:48.78But sometimes equilibria arise even when there is no top down control, because of the fact that it's, as soon as there's an asymmetry, as soon as enough people lean in one direction, everyone leans in that direction, because each new person, it's to their advantage to do what everyone else is doing. 00:14:07.00And that's how many of our customs arise. 00:14:09.90That's how words in a language are accepted. 00:14:14.08Dictionaries don't decide what words mean. 00:14:16.64Dictionaries listen to the way people use words, and they write down the way everyone is using a word. 00:14:22.36I know this because I was an editor at an American dictionary. 00:14:26.68Anyway, in a coordination game, there are two equilibria. 00:14:30.36If everyone drives on the right, drive on the right. 00:14:32.78If everyone drives on the left, drive on the left. 00:14:34.56Or let's take a couple in a situation where, let's say she prefers to go to the opera, he prefers to go to a hockey game, but they both enjoy each other's company. 00:14:46.54So they'd rather, whatever it is, do something together than alone. 00:14:52.36Well, if she's already decided to go to the opera, he's better off going to the opera than the hockey game. 00:14:58.66If he's already decided to go to the hockey game, she's better off going to the hockey game than going to the opera. 00:15:03.92There are two equilibria. 00:15:05.20Both of them can work, and both of them are better than when they are mismatched. 00:15:10.36One of them going to the opera, the other one going to the hockey game. 00:15:13.40In that case, either of them can switch and would be better off. 00:15:17.88But the equilibria are where neither one has an interest in switching. 00:15:22.64The question is, how do you get into one equilibrium or the other? 00:15:26.50And a lot of our social life is driven by what attracts us into one of several possible equilibria.

José Maria Pimentel 00:15:36.38You mentioned society working, and common knowledge is also essential on the political realm, as it were. 00:15:43.00So for institutions to work and for us to have trust in institutions. 00:15:46.78I'm curious to know what you think of the role of social media here, because we are living through a time of, it has been called epistemic fragmentation, right? 00:15:54.50So there was a basis of common knowledge, I would call it, that we all kind of shared, and now it's more fragmented than ever. 00:16:03.10How do you see the role of social media here and the impacts of this destruction of common knowledge?

Steven Pinker 00:16:11.16Well, social media, in theory, work against common knowledge, because in contrast to, say, broadcast media, where everyone can listen to the BBC and know that everyone else is listening to the BBC, everyone can read Time magazine and know that everyone else is reading Time magazine. 00:16:28.88The social media algorithms deliver a personal feed that you read, and you don't know who else is reading it. 00:16:37.04However, I've done studies with my graduate students that show that for certain aspects of social media, people treat them as if they're common knowledge. 00:16:46.34If in X, there's a trending column, or sometimes a for you column, people feel everyone can be seeing that. 00:16:54.88And even if, because you can generate common knowledge on social media, not just passively receive it, you never know when a tweet might go viral. 00:17:05.22One that, if you see it and retweet it or repost it, or in TikTok or Facebook or Instagram, then anything could become common knowledge if it increases exponentially, if it goes viral. 00:17:18.88And that's even more true now that even mainstream media will often reproduce a social media post, they'll have a screenshot of a tweet, particularly if it's from some prominent figure, and then it really does become common knowledge.

José Maria Pimentel 00:17:36.02Well, the words are always polysemic, as you know better than me. 00:17:39.28So here in the sense we can use knowledge in the sense of information, maybe I'm stretching the concept too much, but I was thinking knowledge in the sense of the sharing of a certain understanding of the democratic process, for instance, or a certain understanding of the values. 00:17:56.38There's, I think, to use your words, liberal enlightenment values that support liberal democracies, right? 00:18:02.40We'll see that waning and my view, and I think you share it, is that social media has a big role there, right?

Steven Pinker 00:18:11.46Yeah, there it's not so much whether the knowledge is common or not, but whether it is even accepted. 00:18:17.18The crucial issue with common knowledge is, is there a case where everyone knows something, but they don't know that everyone knows it? 00:18:27.14When you're talking about the fact that half the population thinks that in the United States the 2020 election was stolen and the other half doesn't, there it's not private versus common. 00:18:38.30It's common, kind of common belief in two separate pools, two pools of common belief, which is a whole other set of issues, yes.

José Maria Pimentel 00:18:48.08Let me pick your brain on this, because I'm currently writing a book on social media and its impact in politics. 00:18:54.44And one thing my co-author and I have been thinking about is, what can the future look like? 00:19:02.32Like, imagining this will not go terribly wrong and we'll end up with something akin to religious wars of the 16th century, the 17th century, actually. 00:19:10.50What can the future look like? 00:19:11.66So, in the past we had, as you mentioned, broadcast, we have basically the gatekeepers of the system, be them TV channels, newspapers, whatever, which intermediated information and kept the system sort of working. 00:19:25.68Now we have this more chaotic environment. 00:19:28.82Do you foresee an equilibrium in the future, so with different institutions, a different environment, perhaps, in which we are able to sort of take the, preserve the good parts of social media, which it has, and also correct this chaotic and polarizing element?

Steven Pinker 00:19:45.88Yeah, it's a big challenge there, because basically, the people left to their own devices, just the default, is they don't have very high standards of accuracy. 00:19:59.34People are happy with myths and conspiracy theories and legends and stories that make their side look good and the other side look bad. 00:20:11.04It's only when you have standards of fact-checking and verification and editing and peer review and freedom of speech, epistemic standards, that we have any chance of finding the truth. 00:20:30.08These are rare. 00:20:32.14It happens with science. 00:20:34.18That was what the scientific revolution was about. 00:20:36.38It's happened with historical archives, with statistical data sets that typically are kept by government agencies. 00:20:47.10In the court system, we have the adversarial process of a prosecuting attorney and a defense attorney. 00:20:53.08But each one of these are very special arenas where we explicitly try to approach the truth, knowing that human beings left by themselves will not seek the truth. 00:21:09.04They want good stories. 00:21:11.36Unless you have that, then you're just going to get stories, fake news, conspiracy theories, myths, and so on. 00:21:19.68Now, when legacy media were new, when there was first cheap newspapers, all the news was fake news. 00:21:31.74There were stories of babies born talking, speaking Latin, of Martian civilizations, of two-headed monsters pulled out of a lake. 00:21:43.76They were just in the business of selling propaganda and stories until a bunch of editors got together. 00:21:51.38In various countries, they said, this is ridiculous. 00:21:53.80We have a responsibility to try to tell the truth. 00:21:56.26Let's adopt standards like fact-checking, editing, separation of the opinion section from the news section, printing corrections, and so on. 00:22:08.00Likewise, science had scientific societies and journals and peer review. 00:22:13.30Now, what happens with not just social media, but with the general democratization of media. 00:22:21.74So instead of having a small number of channels, small just because for technological and economic reasons, there were only a small number of mass market presses. 00:22:33.24There's only a small amount of bandwidth for every television station. 00:22:37.92It got, we kind of married the limited number of channels with standards for accuracy and fact-checking. 00:22:48.26Then it all broke open with not just social media, but even before that, the World Wide Web, individual blogs, listservs and discussion groups, cable news in broadcast media, well, not broadcast, but mass media, AM talk radio stations. 00:23:11.28They just proliferated, they just proliferated, and as with the beginning of the press, they were free from the gatekeepers and mechanisms of quality control. 00:23:23.42Now, whether people will get fed up with all of the nonsense and fake news and conspiracy theories, because people don't like being lied to, whether there'll be mechanisms such as, I mean, we see some of them already, like community notes in X, where people could post a correction, or using AI to fact-check a post. 00:23:49.86There may be other mechanisms, like people often falsely think that the country is more polarized than it really is. 00:23:58.86In the United States, Democrats think the Republicans have much more extreme views than a typical Republican has, and vice versa. 00:24:06.94If you even present people with data, this is the percentage of people who think this and that. 00:24:11.54It suddenly changes the way they perceive a post, a tweet, and instead of thinking this is what a typical Republican thinks, they know this is just what an extreme Republican thinks. 00:24:25.06Anyway, those are just some examples of ways in which there may be a pushback, and people, since they don't want to believe nonsense, they don't want to be lied to, that there may be some pressure to have higher standards for epistemic accuracy.

José Maria Pimentel 00:24:39.40So, one solution would be, maybe on the part of the platforms, be it through regulation or through their own accord, changing the algorithm. 00:24:50.52Do you see something, you mentioned how the editors got together and coordinated, to use the term again, back in the day. 00:24:58.16Do you see something akin to that occurring today, in that podcasters, online opinion makers, influencers, whatever, create a sort of a guild, to use an historic term, and sort of agree on some standards? 00:25:11.28I see that very far away, but it might be a solution, I don't know.

Steven Pinker 00:25:14.34Yes, I mean, you don't necessarily want government coercion, because no one knows what the truth is, including the government. 00:25:23.28You need a marketplace of ideas, and government, unlike podcasters and social media posters, has the police and the army. 00:25:33.78So, there's a particular danger of letting the government legislate truth. 00:25:37.92It did happen during the pandemic, for example, that some opinions, which we now know are reasonable, such as did the SARS-CoV-2 virus escape from a lab in China? 00:25:51.92We don't know, but it's certainly possible. 00:25:53.84But during the pandemic, that was considered to be disinformation, and it was censored from some of the social media platforms. 00:26:02.00We're better off knowing the truth, and the only way to know the truth is if a variety of opinions can be voiced, as long as they are also evaluated and criticized.

José Maria Pimentel 00:26:13.30Yeah, exactly. 00:26:13.30There you go. 00:26:14.46So, somehow it needs to occur bottom-up, right, and not top-down. 00:26:19.48So, somehow you need to... 00:26:20.24Yes. 00:26:20.96Yeah, we need to do that. 00:26:21.92Let me turn to AI. 00:26:23.66You mentioned AI before, and I'm very curious to know your view on Gen AI, because you are a scholar of the human mind and also a linguist. 00:26:33.12So, AI, I would say, raises your curiosity at least on two levels, right? 00:26:37.36On its comparison with the human mind and on the fact that it's using language and you're coming at it as a linguist. 00:26:43.62How do you view these models? 00:26:45.20Have their success surprised you? 00:26:49.34Yeah, it did surprise me.

Steven Pinker 00:26:50.90I would not have guessed that just increasing the sheer scale of these models, both the amount of input, which is basically the entirety of the world wide web, and the scale of the compute, that is the number of hidden layers between the input and the output. 00:27:11.00That that increase beyond what we can really imagine, the number of parameters, the number of layers, that in combination with the gargantuan amount of input could lead them to something that is in many ways quite impressive. 00:27:32.74You can create a limerick explaining Kurt Goodell's famous theorem of incompleteness, and it comes up with a limerick, not perfect, but not bad. 00:27:44.58I mean, that's smarter than anything a human could do, at least in a couple of minutes, a couple of seconds, really. 00:27:50.30On the other hand, there's no one really understands how they work with a trillion parameters. 00:27:57.76A lot of the knowledge is smeared in diffuse patterns through the trillion connections, and it's only beginning to see how they come up with the answers they do come up with. 00:28:09.90And we know that there may be some overlap with how the human mind becomes intelligent. 00:28:17.36We also process a lot of information from the time of children. 00:28:21.26On the other hand, it would take the equivalent of 30 million years for a child to process as much information as GPT is trained on, and the child can speak fluently in three years, not 30 million years. 00:28:36.84We also know that when you train a child as a child learns language, their utterances aren't like a half-trained large language model. 00:28:46.94Also, the large language models have a funny combination of brilliance and stupidity. 00:28:53.02They are subject to hallucinations and blends. 00:28:58.26They cannot be updated by single facts in real time, but have to be trained on a whole corpus all over again. 00:29:06.84They make bizarre blends. 00:29:10.58They have failures sometimes of simple spatial reasoning. 00:29:13.84If A is above B and B is above C, A must be above C. 00:29:17.74They don't really have an internal model of the world, of objects and places and people and events, but rather hundreds of billions of statistical associations. 00:29:31.94I mean, it's amazing how well they do with that. 00:29:34.82A lot of it thanks to the reinforcement learning from human feedback and the post-training, where they are kind of hammered into standards of relevance and truthfulness and harmlessness. 00:29:49.32And there now are calls within the AI community by some of the experts for a new kind of architecture that would have a model of the real world and not just associations between words and other words.

José Maria Pimentel 00:30:04.74Yeah, exactly. 00:30:05.34That's exactly where I wanted to get, because that's a common criticism on these models, that they lack a model of the world, as you mentioned. 00:30:13.98So while they are very powerful pattern recognizers, they don't have a model of the world. 00:30:18.60Maybe this is too rough an analogy, but that reminds me a bit of the distinction between System 1 and System 2 of our brains from Daniel Kahneman and Amos Traversky. 00:30:31.34And it almost seems to me that, in a sense, these models lack a System 2, so they don't have a model of the world, but they have the most powerful System 1 we've ever seen. 00:30:42.30Because our System 1 of our brain is, in itself, also a pattern recognizer, right? 00:30:49.28It's sort of how it works. 00:30:51.12And these models, they do that on steroids, basically, because they are much more powerful. 00:30:55.34And they have, as you mentioned, was it 30 million years of data? 00:30:59.68Yes, it depends on the model.

Steven Pinker 00:31:01.80Often we don't know because it's proprietary. 00:31:04.08Yeah.

José Maria Pimentel 00:31:04.50Anyway, it exceeds our intuition, right? 00:31:06.28So it's way beyond. 00:31:07.90But do you agree with this, or is this too rough an analogy?

Steven Pinker 00:31:11.70There's something to that. 00:31:13.72You know, one way of thinking about it is, in any computation, there's an input, and then there's an output. 00:31:20.60And in the first neural network models, every input was connected to every output. 00:31:26.28Turns out those models were pretty stupid. 00:31:29.24They just couldn't really learn anything of interest. 00:31:31.78Then, in the 1980s, there was the brilliant idea of a hidden layer between the input and the output. 00:31:38.80So it could have abstract concepts, not just stimulus response. 00:31:44.56The breakthrough, and Jeffrey Hinton won a Nobel Prize for this, was figure out how to train the hidden layer, because the hidden layer did not actually interface with the real world. 00:31:57.26So there was no easy way of training it with the correct connections. 00:32:03.82So Hinton and his collaborators came up with an algorithm for training layers that had abstract concepts, one layer of abstract concepts, called error backpropagation. 00:32:14.56Now, that was smarter, but it was still often there's a limit as to what they could do. 00:32:19.96Then they said, well, let's try two hidden layers. 00:32:23.46Let's try five hidden layers. 00:32:25.24Let's try 100 hidden layers. 00:32:27.38Now, you might think 100 hidden layers. 00:32:29.22You'd have to backpropagate the error signal, getting more and more diluted with every step. 00:32:35.68That can't possibly work. 00:32:37.46But what we now know is with hundreds of billions or trillions of examples, it, I mean, not exactly could work all the time, but can be very impressive. 00:32:49.62So it is kind of a system one on steroids. 00:32:52.02It doesn't do, it isn't at least designed to do step-by-step reasoning, to do logic, to do algorithms, but it's got so many abstractions at such a high level that it is coming very close and in some ways can deliver answers that step-by-step reasoning could not deliver.

José Maria Pimentel 00:33:12.82Yeah, yeah. 00:33:13.18And we've all have that experience. 00:33:15.10I'll ask you one final question, because I know you're pressed on time and take your time to respond. 00:33:20.30I teach critical thinking, so the connection between AI and critical thinking is very important to me. 00:33:26.16And as you know, there's been a whole lot of talk about reliance on these models leading to so-called cognitive offloading, right? 00:33:34.18So that we tend to offload our thought process to these models and thereby shrink, I would say, or atrify our critical thinking. 00:33:44.08Do you agree with this and do you have any recommendations for people to use these models without this side effect?

Steven Pinker 00:33:51.94You know, in some ways, this reminds me of a debate that took place when I was a college student more than 50 years ago, because then the innovation was handheld calculators. 00:34:02.78The question was, when you had a calculator and you could put in an arithmetic problem and it would come up with the answer, will our arithmetic abilities atrophy so no one would be able to multiply, you know, seven times nine anymore? 00:34:18.46So, you know, a couple of things, one of them is, you know, calculators were really good things. 00:34:23.26And then, of course, spreadsheets and then a lot of software libraries and so on. 00:34:29.16So, I was taught, for example, as an undergraduate, how to calculate square roots with paper and pencil. 00:34:37.34This was, trust me, very tedious. 00:34:40.34I bet very few people in the world now know how to compute a square root with paper and pencil. 00:34:46.00But, you know, we're not missing much. 00:34:48.50One button on the calculator, square root pops up and that's good. 00:34:53.06You know, on the other hand, it still is good if people could do mental arithmetic. 00:34:57.30One reason is to, because you don't always have a calculator or a spreadsheet or ChatGPT. 00:35:03.88And another is you have a deeper level of understanding of many things, both how calculators work. 00:35:10.30You can fact check to make sure that you haven't entered the wrong number into a calculator. 00:35:15.60And also just to understand how math works so that you might be the engineer who develops the next algorithm and you have to understand this. 00:35:23.98So, we're facing a parallel set of questions with generative AI, namely, it can look up facts, it can often reason pretty well about facts, but with possible hallucination errors. 00:35:37.98And so, it's essential both in education of kids and in keeping our own minds powerful enough to use AI that we still know what fact leads to what fact, how claims are established, how peer review works, how logic works, and so on, so that we can use AI more effectively. 00:36:03.32If we retain those skills, then we can perhaps multiply the power of our own thinking by offloading some tasks to AI in the same way that we offload computing a square root to a calculator.

José Maria Pimentel 00:36:18.44Yeah, yeah, I completely agree. 00:36:18.44Well, Stephen, it has been awfully interesting. 00:36:18.44Thanks a lot for making up the time. 00:36:18.44I hope your publisher here manages to get you at some point to Portugal to present your book.

Steven Pinker 00:36:30.42I hope so. 00:36:30.42I love Portugal. 00:36:30.42I would love to return. 00:36:33.86Have you been here? 00:36:34.20I have, yeah. 00:36:34.82Ah, I didn't know that. 00:36:34.82Okay, cool.

José Maria Pimentel 00:36:36.40I've been to Lisbon. 00:36:36.40I've been to Porto. 00:36:36.40Yes. 00:36:38.92Okay, great. 00:36:38.92So, maybe then we'll meet you in person. 00:36:38.92Thanks a lot. 00:36:38.92Thank you very much. 00:36:43.22Thank you. 00:36:43.76Este episódio teve a sonoplastia de Hugo Oliveira. 00:36:48.66Ouça e veja outros podcasts em expresso.pt e acompanhe-me em josemariapimentel.pt ou nas redes sociais. 00:36:56.02Até ao próximo episódio.