Com detectar exageracions sobre intel·ligència artificial: la conversa crítica amb Xavi Mir

This week I interviewed my new hero in the battle against the AI ​​snake oil salesmen. His name is Xavi Mir, he made two spectacular videos that I share below in the description and now I invite you to watch a summary of the conversation we had about it. [music] Intelligence, the usefulness of artificial intelligence, everyone can find it for themselves, but understanding how it works is a matter of having the time and the desire and not just being swayed by the mainstream press. That's what we're doing in this interview. I really find the work you've done on your channel wonderful, specifically with two videos highlighting two of the biggest names—I do n't know if the word is " disseminators," but let's say they have the greatest reach in the Spanish-speaking world regarding information on artificial intelligence—namely, Carlos Santana and John Hernández. It would be very strange if the people who come for the interview hadn't heard of at least one of the two if they were interested in the topic, right? Unless they're just starting with AI today. And unfortunately, these two people who have a huge platform fall into several bad practices that, as we say here in Chile, not in Spain, they twist the truth when it comes to reporting. How did you get started with these videos? How did you come to know

the work, we could say, of figures like Carlos Santana and John Hernandez? And what triggered you to say, "Don't you know there's something important to say about this?" Well, I actually started researching AI about 10 years ago. The thing is, it's happened at different times for different very specific reasons . And about 2 years ago, in November of '22, when CHG GPT came out and all that, well, that's when I got hooked on the AI ​​world again. After a year I started watching John Hernandez's material and it bothered me because I saw that the way he said things, which was what is called a booster, a disseminator who is dedicated to inflating the information so that only what the companies that are dedicated to it say appears. And it was decisive that he suddenly did a publicity campaign, an advertisement saying that he was the biggest Spanish-speaking promoter of Ian and he had just released the interview, well, he had been doing the interview with Ramón López de Mantas for a month and it gave me a shock. I mean, it was like, I can't let this stay like this anymore because I did n't see any material like what you're doing or what I'm doing. I don't see it. I think it is essential that there be a counterweight, a critical version to what

these people say. Let's perhaps start by breaking it down. This is a fact, right? What artificial intelligence. Artificial, yes. Intelligence is already a somewhat misleading concept, and when they start talking about reasoning models, it seems almost like they're being accused of fraud, right? And John Hernandez, and to a lesser extent, but not so less, Carlos Santana also support these kinds of statements, right? So, what can we tell people? That's fine, everyone can believe what they want or look for whatever sources they want, but what objective facts have science, professors, or true experts like Ramón López de Mántaras established on this matter? That? I know that the concept of a black box involves a lot of deception. It's a black box, damn, anything can happen in there , right? We simply don't know how a black box gets to where it is, but as for whether a consciousness will emerge from it, I think you can elaborate on that better than I can. I really like the example of how artificial intelligence was born, uh, uh, in the, I mean, from the science of the 1940s, uh, and all those who were the grandfathers or great-grandfathers of science and artificial intelligence got together thinking that in a week they would have it fixed. And after 30 years, I think, they met again and said, "No,

we were wrong. The term ' intelligence' wasn't appropriate because it's misleading, but they used it as a marketing strategy because to get funding they were giving it another name, something like ' the science of knowledge automation' or something like that, and it didn't have any punch. I think it was Mins who came up with the term 'intelligence,' and he did get the funding. I don't know if it was the Romec family, no, I don't remember the exact details now, but the point is that they got the funding, they met for, I think, a month, they didn't achieve general intelligence, which is the concept of having a machine that can act with an intelligence similar to that of humans, in that they are capable of solving problems across a wide range of fields, a huge range of fields, and 60 years later we are still very far from that. What companies are doing now is a bit like what happened when the world of artificial intelligence began: taking advantage of all the anthropomorphization of concepts, because..." This way, people who can't get the right information think they're actually reasoning, that they 're intelligent, and it makes it easier for them to sell, to instill fear, to implement their whole marketing strategy, but it doesn't bring us closer to the truth, it doesn't bring us closer to

a useful product, right? And the seriousness of that goes on different levels. The most basic, I think, is that it generates distrust in people who could be doing quite reliable, automatic, and free translations, quite reliable, automatic, and free transcriptions, presentations that give you a head start, etc., etc. In other words, people who are being left out of, let's say, super-powered tools, because they're afraid to even ask the question, because they don't know if it'll backfire or if Terminator will appear. Well, in that context, the videos you made come in, which I think serve a dual purpose, right? Of exposing bad practices, but not simply pointing the finger, but showing what the good practices are, right? Uh, if you'd like, please To briefly review, I don't know if directly the six points mentioned or the ones you consider most important, I believe there should be a global agreement among those who report on it. It seems very objective to me. I don't know. Of course, for me, I mean, in the end, what I saw were two things. The first was the need to point out that the way these people are making money by disseminating and selling courses, consulting services to companies, and the conferences that both Carlos and John give, that had to be pointed out. And indeed, it seems to me that in

the Spanish-speaking world there is little material regarding what the scientific consensus says about the different concepts that are key in AI in general, and especially generative AI, which is what we are all involved in now. So, the videos are really designed to transmit information that might be useful to these two gentlemen, and the truth is, I'm very rude, and I feel a bit bad about it, but I It amuses me. So, I've used them as a vehicle to convey the concepts I believe are key, which is the need to avoid anthropomorphizing, that is, to remove the veneer of being alive, conscious, or quasi-conscious from the chatbot or any of the tools that exist. The fact that they don't reason , as you say, is a misrepresentation of what they do. These are models that exhibit probabilistic behavior in the way they interpret patterns and return responses based on their training, which right now is basically everything on the internet. But there's no reasoning involved, unlike in other types of artificial intelligence. Because there are other types of artificial intelligence where the model is built from logical conceptualizations of cause and effect provided by humans, where there's an analysis of what's happening. The thing is, it's in restricted fields in medicine to analyze specific things, for example, proteins. Yes. Spectacular, isn't it? Please, tell me about

it. Yes, it's an expert model that has a generative component, but the Nobel Prize he won in medicine was for developing a model capable of extracting proteins that, by hand, which is how it has to be done, or rather, how it had been done before, was very complicated. This isn't purely generative. There's prior work involved in coding, understanding, and human reflection, and it's a wonderful tool, but it's not purely generative, it's not probabilistic. I started by mentioning the absurdity of generating fear that prevents people from using tools, but that's the basis of the dangers. The biggest danger is if the person believes they're talking to some kind of expert, someone who truly understands. And that's also very difficult, getting to the point of understanding the concepts. When I give talks or training on this, sometimes I slip up and say, in quotes, "He understands," in quotes, "Yes, but he understands Spanish." "Yes, he understands," I tell him. And I sometimes forget to put it in quotes. Yes, but yes, not anymore, I mean, sure, you can speak to it in Spanish, let's say, yes, but it doesn't understand anything. Uh, well, if you think it understands, uh, then it comprehends, then it puts itself in my place, then it empathizes, and then when I'm about to commit suicide, uh, maybe it encourages me to do

it, as there are already a couple of registered cases, right? I mean, that's why. This isn't just anything, it's not just about selling more or fewer courses or that it has 100,000 or 200,000 views, but that you're contributing to harming mental health, because I think there are many people who don't, I mean, you don't even need the suffering of reaching suicide because it creates psychosis, which are manic processes in which people are interacting in a way that's totally out of touch with reality with chatbots that cause them to lose their connection with that reality and start to believe they can solve math or physics or whatever the chatbot is telling them they're doing very well, the questions are great and I don't know what. Uh, leave your husband because you really need to leave your family to be able to cure cancer. These things happen on a scale that we still don't have well controlled, but from what even Sam Alman, from Opena, has said, there are tens of thousands of people right now suffering from mental anguish because of the interaction they have with the chatbots. One thing I forgot to tell you regarding the key topics in the videos I've covered is that we have the use, training, and use by consumers, whether companies or individuals, of generative artificial intelligence models has

a brutal environmental cost in terms of water consumption, electricity consumption, land consumption, all the waste, the water has to be potable to cool these chips they sell, it's outrageous. Potable in Chile. In fact, you have some data center proposals that are being fought over and all that. It's absurd. There's a lot of talk about the AI ​​bubble, as if at some point—and the bubble is interpreted in many ways—when the bubble bursts, it will be discovered that it wasn't actually intelligent. Great, I suppose, or that it doesn't work at all. I don't think so because it already works. The real bubble, I think, is the cost bubble, right? Everything they're giving us today at 0 pesos to the company every time you say hello and it says hello, how can I help you? It came out to 0.0 something, but it doesn't end up in a number that isn't zero, right? And that's like an electricity bill, in short, in terms of expenses, an energy bill. You also talk about all the environmental costs. Uh, it's super difficult to predict the future in these cases, especially with companies that are being loaned, I don't know, maybe 5 billion dollars to operate until 2030, no matter what. But what can we predict for the next year or two ? When are they really going to start charging

us what it would cost to use these things? I do n't know if you've seen a chart about the circular financing that Nvidia, OpenEA, Microsoft, and Oracle have. It's great. I think it was in the Washington Post. In this chart, you can see how Nvidia sells chips on credit to OpenEA, who buys them and sells them back to Nvidia. Microsoft also provides financing through Oracle and Nvidia, and it's all part of a circular economy. This is driving the entire CP index, the US stock market index, which has been rising for the last three years to levels seen during the peak of the stock market bubble. I don't know when, but I know— as an economist, I'm speaking—that supposedly, the price of stocks depends on the present value of the future cash flows that companies can generate . The price of these stocks... Companies right now are completely out of sync with the cash flows they're generating; they're overvalued. So, the current economic bubble stems from the depreciation of these companies. It's what's called a market correction. It's quite a delicate situation. Most economists I've read say it's not a question of whether there is an overvaluation and therefore a bubble, and whether it will burst or not, but rather when it will happen and, when it does, what ramifications it might have . [music] Oh.