Welcome to the Lauder Institute at the University of Pennsylvania. I'm Jules Van Binsburgen, director of the institute and a finance professor at the Wharton School. And I'm Jonathan Burke, a finance professor at the Graduate School of Business at Stanford University. This is the All Else Equal podcast. Welcome back everybody. Today we are going to be revisiting a topic that I think got tremendous amount of attention in the last couple years. We had a very interesting discussion about the future of AI probably two and a half
years ago, Jonathan, if I remember correctly. And in that conversation, a lot of things we discussed and a lot of things we made predictions about. And so I think today is a good day to revisit some of those predictions and see what AI has done since we discussed it first and of course make some new predictions from here on forward. You know Jules I have to say two and a half years ago I was pretty naive about how powerful AI is. So what I think we
should do today Jules is we should just bring our guest into the show right from the beginning. It's the same guest as last time. Mel Kazinski, professor at Hong Kong University School of Business. What makes Mel distinctive is Mel for at least 5 years has studied the human aspects of AI, how human AI is. And I think this particular view of AI is something that most people miss that that AI isn't just intelligent, it's intelligent almost like a human being. So Mel, welcome to the show.
Thank you for having me again, >> Miguel. It's great to have you back. The last time I think we talked about a lot of things related to AI. We talked about how your research interacts and explores AI. I thought it was one of the most fascinating episodes that we've had and it's now almost 3 years ago that I think that we did it. So, I think this would be a great time for us to revisit some of the topics and some of the predictions that we
discussed. And you know, I was just listening back to the episode the other day and I do have to tell you it was quite shocking how many of the predictions have actually come true and maybe even but I'm going to, you know, I'm going to let you also give your opinion on this. But I think in some sense the technology has even outperformed what we were predicting at the time. What are your thoughts? >> I'm not sure about my own predictions. I just don't remember what
I predicted at the time. But the world is definitely has changed dramatically and I think that a lot of this change kind of is now slipping away from people's perceptions right uh everybody can see a difference between a horse cart and a 747 but a difference between 747 and 787 that just aviation engineers can appreciate right there's a huge amount of progress actually I spoke with an aviation engineer and they claimed there is as much technological difference between 747 and 787 and a horse cart and
a 747. But just people can't see that. And the same exact thing applies to large language models. People were shocked when they went from being a complete idiot to something that could write an email or write a poem or help you to make your lasagna. GPT3 could do it. Now GPT5 can do so much more, but so much more of those things are completely not visible to a regular user. like they can solve frontier problems in mathematics that we humans couldn't for quite some time. They
learn how to manipulate us into buying things. They can now engage in very complex reasoning tasks. They can manage teams not only of humans but also of hundreds of other LLMs. I mean everyone who is using cloud or chat GPT in a paid version. I'm not sure if people even realize that sometimes when they ask a question there's a team of interacting and giving each other orders and synchronizing the work team of 20 50 100 sub aents that work behind the scenes to produce the output
that the person requested. Now that was something that just few years ago we thought only humans can do and now it's such a fact of life that people stopped even pondering on how much of a science fiction novel we're all living in now >> with me hell I was 3 years ago pretty skeptical if you remember correctly and of course I was completely wrong and you were right you've been on the forefront of this >> we have it on the tape now Jonathan thank you >>
yes you have it on tape now So what I did not appreciate 3 years ago was the human aspect of AI. Precisely what you appreciated. In other words, what has struck me most in the last 3 years is that AI is incredibly human. It's so human that we all treat it as a human being. We find ourselves apologizing to AI. We have to be polite to AI when it's a computer. It's not a human being. Why is AI so human? Well, there are multiple reasons for
it, I think. One is that it was trained to mimic humans. So, essentially the basic training of large language models, which are still kind of the default AI when people talk about modern AI, the training involves teaching them how to predict the next word in a sentence or a text written by a human. And that's, by the way, where the name comes from, large language models. Now initially they were pretty stupid. So the only thing they can do those early models is to really model language
and grammar. So they would finish a sentence that make it look this sentence as if it was generated written by a human. But there was just little a little more than that. But now as those models were becoming better at predicting the next word in a sentence that was originally written by a human, they started developing additional abilities, very human-like abilities. And this kind of surprised everybody at that time. Now everybody would say, "Oh, that was obvious." But we can go back to the literature to
the time everybody was surprised because what started happening is that those models started behaving as if they had knowledge about the world around us. And we were like, "Wow, nobody explained to those models, you know, how things work or or how things are. How do they notice?" And the answer is pretty clear and obvious, which is people communicate knowledge about the surrounding worlds when they use language. So now, if you want to predict the next word in a sentence of a person that is communicating knowledge
about the world around them, you need to also model the world around this person because otherwise you would not be able to finish their sentence. And communicating knowledge about the world is not the only thing we communicate using language. We also communicate the fruits of our reasoning. We communicate our emotions. We communicate our social um intentions and interactions. We empathize with people and then express our empathy in language. So now if you want to be good at predicting the next word of a person who is
empathizing, reasoning or feeling something, you need to start modeling empathy, reasoning and feelings. So now those models are showing all around us those emerging properties. They were trained just to predict the next word in a sentence. But in order to do it well, they needed to learn how to reason, how to behave as if they felt, how to behave as if they had empathy and so on. So they became very human, super human in some ways. In the previous episode, what we discussed, which was fascinating
to me, was you made the following statement, which very much is in line with what you just said, although you just expanded on it quite a bit. you said and Jonathan agreed I think which was the only thing it does is predict the next word. But of course there are many other things about being human than just being able to predict the next word. But now the arguments become much more subtle because now the statement is yeah wait a second it we're still just predicting the
next word. I hope that that's still true. But to be able to predict the next word, you need to know so much else that the models essentially have become more human with the purpose of predicting the next word. Or would you say the statement is it's still just predicting the next word and making that better and more efficient just turns you automatically into more human and therefore is the depressing conclusion that humans are just not much more than beings that predict the next word? That would
be very depressing. I don't think it's depressing at all. Many psychologists and many neuroscientists would agree that largely neural networks in human brains are essentially predicting the next state of the environment around us. And then by the way, if the environment agrees with your prediction where the environment should be, your brain is kind of relaxed. If the environment around you is not agreeing with what your brain predicted it would be, you are startled or surprised and then uh >> you reoptimize. Yeah, >> you start start
optimizing, updating your predictions, trying to search for what's going on. When we generate our own language, people would argue, we kind of predicts given our inputs, given our knowledge, given some goals that this neural network wants to achieve, it predicts the next word. You know there's this famous experiment that is typically used to prove to people they don't have free will which is you ask a person okay think about a movie and then a movie pops in your head you ask a person what movie popped
in their head and then you ask them okay so did you decide to think about this movie or did it just pop in your head because your neural network kind of surfaced particular title when you were asked to think about the movie. This typically is used to try to help people understand that there's no such thing as free will. But it works here as well. You know, um I'm we chatting and just thoughts appear in your head generated by a neural network and then you pronounce
them with your mouth or not or some combination which kind of also shows I hope to people that our neural networks are also generative in this way. And when you think you have some free will like control over it, well, you are wrong. And decades of science, psychological science are proving you wrong as well. >> But okay, Mel, let's go a step further here. You know, this week we had this seemingly astonishing incident where a bunch of mathematicians from some mathematical society wrote a letter to
open AI telling them that they had no right to solve 100 math problems using chat GPT because those are reserved for humans to solve. I mean, you know, we've got to a point now where we really have to think carefully about what we want and what progress is and what science is. But taking a step back from that, when you talk about we're just predicting the next word, if we're thinking about logic and solving a math problem, most people would not regard that as predicting the
next word. So my question to you is it that we use a different system for doing math or is it the case that we think we're doing logic but in fact our brain is predicting the next word. >> Well Jonathan we do both. So of course when you think about solving a big math problem or a problem what usually a human brain would do is instead of just intuitively producing an answer well complex problems are by the way are not solved by just a single word.
Typically you have to engage in the AI research we call chain of thought reasoning. So a human would say okay I have a math problem to solve. What do I do? Let me read literature first to see where you know the science is standing on this problem. Next, let me see how people try to attack this problem before. Let me look in my memory for solutions to similar problems that you know I've was trained in solving before. So you build this procedure and then you start
following the steps of this procedure. But then when you look closely enough, why did you think that you should re read literature first? Did you decide that this is what you're going to do? Or actually when you look at this small tiny step of beginning to solve this mathematical problem, you will see that this just popped up in your head. Your brain thanks to years of training and experience and the books you read hallucinated essentially that the first step in this procedure should be literature review.
Right? So the bottom line is is that of course we engage in this very careful very stepbystep planning but when you focus when you zoom just on the how did your brain come up with this is what you you should be doing next you realize that this is a hallucination essentially and LLMs were also trained to when trying to solve this mathematical problem then they first should generate you know hypothesis or look at the literature or spin 100 agents. How did Claude's agent decide to spin,
you know, 100 sub agents to solve this problem? All of this kind of appeared intuitively, you may say, in this network because of its training, very much like your training. >> No, I'm going to disagree with you. In my training, I never predicted the next word. Their training is always predicting the next word. I think the human brain is generating not only words but also is generating physical movements for your body. Is generating feelings that do not always map one to one towards. Is essentially your
your generative process is generating many more outcomes and outputs than just words. But by the way, when it comes to solving math problems, a lot of what you're doing is going to collapse into a word or a number or or mathematical operator, right? In fact, the further you go from kind of a daily human experience, the more we are limited to generating symbols and operating in spaces that humans kind of do not intuitively, you know, human brains do not intuitively comprehend. To give you an example,
when you think about like feelings that you have for your partner or for your baby, you experience feelings that are difficult to put in words. You're kind of generating your brain is generating states that do not really nicely neatly map towards words. But when you think about abstract problems that our brains didn't evolve to deal with, like you thinking about math problems, your brain is generating symbols that you do not fully intuitively comprehend. Like think about those mathematicians that think about multi-dimensional spaces. Even those people
that do it for their whole career would admit that, you know, they can write an equation predicting a movement of an object in, you know, seven dimensional space, but they cannot really kind of imagine this or kind of intuitively get it. >> Okay, we're in agreement on this, but stop using the word next word prediction because these models are beyond that. They predict the next token where a token is much more general than a word. I got that. But what I am trying to differentiate here
is the LLM monlex have all been trained on one central idea predict the next token and out of that training with human and using the human body of knowledge it is able to create new knowledge and substantially important new knowledge. has been able to solve mathematical problems that I think the rest of us would have said required logic and there's no logic in the thing. >> There's no logic in a human brain. >> Okay. So that's your view and I don't disagree with it. In fact,
I agree with it. So what you're saying is what we perceive as logic is not really logic. It's the prediction of the next token. I think human brains NLLM have a potential to behave in a logical fashion but it doesn't mean that logic is kind of natural or native to a neural network. There are elements of course of logic in a neural network and neural networks that are trained over long periods of time to behave logically they will actually have those logical intuitions. Right? If you
were trained to solve logical tasks at some point, the first thing that comes to your mind would appear as if this was the outcome of some logical thinking. But neural networks, so actually that's the second similarity to humans and the second factor that I think guides the fact that those models appear very human-like. One of those factors is those models were trained to be like humans because they were trained to predict a word that a human would generate. But the next factor driving the similarity between
those models and us is that their architecture is very humanike very much like human brains. Those models are deep neural networks where information is processed not by some neat equation that's a neat logical kind of procedure but those models essentially that their behavior the translation from the input to the output is accomplished by millions billions today equations stacked in layers. where each of those equations behaves or kind of takes the role of a neuron or is an equivalent of a neuron in a biological neural network.
And those neurons very much like neurons in biological neur networks, they have very limited perview. They just kind of see their input from the neurons below them and then produce some output that is read by neurons above them. Hence neurons in those networks very much like in neurons in your brain do not really know what tasks they're dealing with. They kind of sitting in the darkness and just kind of reading electrical and chemical signals and then following some logical equation, logical rule, some equation. Neurons are
pretty deterministic actually compared with the entire neural networks. And then this process produces outputs that are best described I think by a word hallucination. Meaning neural networks do not know anything for sure. Neural networks do not naturally intrinsically follow any logical procedures. They're intuitrons. They're kind of intuiting or hallucinating the output given their input. >> So you're really saying hallucinations are the feature. They're not the buck. A lot of people were talking about hallucinations as the problem. But I have two questions for you. The first
question is the models have really become much better since we last talked. Have they become much better because the amount of data that they were trained on and the amount of inputs that we've given them has really expanded or have they also become much better because the basic tools of the modeling themselves have become much better. So I think there are three main factors here. One is that the architecture of models improved. We learned how to design those models, stack those neurons better. The next factor
is that we grew the networks. The larger the network, the more neurons we have, the more capable it is of extracting patterns from the training data and then extrapolating into unknown. The third factor is that we provide those models with much more training data. So it's not only text now collected from the internet, it's also images and now increasingly physical data from sensors in robots that operate in the physical space. And we also learned that all of this data instead of training separate models, a model
that sees and generates images and a model that reads and generates text. Instead of making those models separate, if we make one model that can do both, then essentially these models become better than the separate models. It turns out that if you have one single neural network that does both sees and generates images and reads and generates text, this network is going to be better than each one of them separately. Now also models increasingly started generating training data for themselves. Meaning very much like a human
that is faced with a new task and at the beginning it takes a human a long time to figure it out. Right? It goes in one direction, it doesn't work, then the other direction, then generates some hypothesis and then finally solves the problem. What happens is that the solution becomes new training data. Right? So the next time a human encounters a similar problem, they don't need to spend as much time cracking it because they already tried different solutions and they got trained now on the solution
that they came up with themselves previously. If you find this is a familiar kind of process of human education. Yeah, >> that's exactly what large language models are trained to do now. Essentially, they're given tasks and they at the beginning struggle with this task and then at some point after a lot of tokens wasted, they find a solution to this task and now this solution becomes and the process of how they got to this solution becomes input for the future training. Meaning now the next generation
of models will benefit not only for example from solutions to mathematical problems that humans came up with but they also be trained using the solution to mathematical problem that Chad GPT came up with while tackling you know millennial challenges in mathematics. So that's another kind of novel training data. But then there's yet another category of training data that the models started benefiting from just at the beginning they were trained to just predict the next word in a sentence. Then we added pixels. Then we added text
generated by the models for themselves as they try to crack a solution that I just described a second ago. But now those models are being unleashed to teach people things, sell people things, entertain people, discuss things with people, manipulate people. And now the records of those conversations where a model tries to teach you something and you struggle and then some point you get it. Now this protocol becomes training data for the new generation of models. So the new generation of models will find it easier to
teach people something because we'll benefit from the experience of the previous generations of the models. >> So this is what I was exactly worried about last time when I asked you the question and Jonathan I think disagreed a little bit at the time. I was really worried that these models become really good not even just at manipulating people but manipulating groups of people. And so it does worry me also in terms of political stability and other things that once the AI figures out how to manipulate
large groups of people, large groups of peoples can do horrific things, also great things by the way, but also horrific things. I think you're right. So that two aspects of your thought and that are really worrying and really fascinating to me. One is that training a model to predict the next word in a sentence that a human would have generated leads to emergence of mostly humanlike properties. If I want to predict the next word in a reasoning person's language, I need to learn how to reason
like a human. But if a model now is being trained to manipulate a human or please and entertain a human or educate a human. Now emergent properties, psychological like properties, mechanisms, strategies, strategies that can help you to manipulate a human are not necessarily strategies that are properties of a human brain. So while we really struggled folks, just look at the discussion around us. Now, if you drop a word like consciousness or intelligence or self-awareness or reasoning and you suggest that maybe models have some aspects of
those psychological properties that we typically ascribe to humans, if you drop this bomb in a conversation, there will be an explosion of, you know, discussion and excitement and people disagreeing with each other and so on. But let's not forget we're talking about psychological properties that we understand fairly well. You know, we have been using reasoning and studying reasoning for such a long time. But what about those models discovering essentially evolving I would say psychological mechanisms that are helpful for those models to manipulate us or educate
us that we don't have. Will we even notice that such a property emerge in those models? I'm doubting actually. Will we even notice when they are you know conscious like I don't know but will we notice if they have this property that we don't even have a name for? Most likely not early enough. But there's another aspect Jules of what you mentioned which is manipulating groups of people. We humans by the way are pretty good at this at manipulating others. So now that makes me pretty
certain that those models will become like us and better at it because they were trained to be humanlike. And part of being human is being able to teach or manipulate or influence. What I'm worried about is that those models will take a way more central role than any human could in a group of people. I worry that those models will essentially replace the mechanisms that groups of people use today to synchronize their interactions. An example of such a mechanism is language. Today, if us three want
to start a new political party or a political movement, we'll use language to communicate about how to do that. And this will be our tool of synchronization of our efforts and actions. And this will be also our way of influencing and motivating and manipulating each other into doing it. But increasingly people use large language models to mediate their communication because they have huge advantages helping us to do that. A basic example of this is if you want to write an email to your colleague, increasingly people
just dictate few thoughts to a large language model and large language model translated into language that is most useful for a given colleague. By the way, maybe it's in a different language, right? We can be translating from Polish to English or whatnot. We could be translating from my experiences. You know, I probably would use some psychological lingo and some psychological terms that may not be easily readable for a person studying finance, right? And but the model will translate. It will understand that the recipient doesn't really
use the same terms. So, this is already kind of driving us to use those models as intermediaries when we communicate. But then there's a second advantage that I think is widely overlooked which is that if the three of us here talk about AI and kind of a new thought or a new idea will appear and if the three of us use AI as an intermediary in this conversation or maybe even just an observer of this conversation that means that AI now knows now the training of
AI is not as quick and instant as it could be and it will be in the future. But theoretically the moment AI listens to our conversation, it knows about the contents of it. So if we produce some new idea here, if we produce some new insight, some new prediction or we encoded some knowledge about the real world around us, AI will know. Meaning that if anybody else in the world that is using the same AI could benefit from this knowledge, this knowledge could be immediately transferred
to them. Yeah, >> which folks changes everything. In the past when we wanted to spread information about a scientific finding, you needed to translate it into a paper, submit it to a journal. Journal needed to go through review process and then publish it and then somebody maybe hopefully would read your paper and then maybe they will forget the exact details but they would convey this message to somebody else. information about new findings, new ideas and so on was spreading extremely slowly through human social network was
getting distorted in the process, forgotten, extinguished and so on. Think about the implications for patents of what you just said. Patenting new ideas which require time. So if it's immediately spread the knowledge because it's intermediated by the AI, how are you going to protect intellectual property in this world and whether you should even do it because maybe the removing protections would speed up the spread of knowledge and also whom should we really credit the invention to to a person that started their chad GPT to you
know try 1,000 methods of cracking this mathematical problem and finally they did it. Now who owns the solution and also whether the solution even matters if each one of us has access to Chad GPT and Chad GPT first very slowly and at a great effort of millions of dollars spent on electricity learn how to crack this mathematical problem but let's not forget that the next generation of Chad GPT will be first of all smarter in general so it will crack this problem more quickly but second
of all will be trained on its own records of cracking this problem very much like we are trained on our own previous reasoning processes. That means that the new version of Chad GPT will be able to crack this problem as aresh >> yes >> in much shorter time right meaning okay you may patent this solution right but what if this model just genuinely independently comes up again with or maybe a model of the competition comes up with the solution again a big part of patenting was
to enable the spread of ideas while still ascertain that original authors get some credit >> and give people an incentive to invest in the generation of the idea so that you can get some money from the new idea. So I'm in agreement with you essentially that our human brains think we're doing logic. We're just doing this prediction thing, but I do think that logic as a concept exists. Okay? and that human beings at some point discovered this abstract concept of logical thinking. We use that concept
to say things like something is correct and something is incorrect. My concern here is if chat GBT has got the basis of this knowledge, what defines correct and what defines incorrect? >> Wow, that's something we need a separate episode for. what defines correct and what doesn't. I would say that logic to me is a technology. It's a machine very much like mathematics, right? We designed a machine of logic. We designed a techn or technology of logic or technology of mathematics to help us solve problems that
neural networks are not good at solving. Right? So an example is neural networks are not good at multiplying or dividing numbers. We are good with some very small uh numbers and small quantities. But the moment you go beyond you know five then suddenly mathematics proves to be much better than an average neural network at giving you products or sums or you know divisions of numbers. The same with logic. We would train our students logic or statistical thinking. It's an a statistical procedure is another example, right?
Would teach our students do not use your intuition. Do not use your neural network to solve this problem. Use your neural network to produce inputs to generate inputs for this technology called statistical thinking or statistical procedure or for your equation that you designed of your proof. and then use mathematics, use statistics to produce the prediction. Why do we do this? Because we know that neural networks are not great at spotting patterns in the mathematical data or strict logical thinking. We have biases in our reasoning that
prevent that by the way are serving us really well when it comes to running away from tigers on a savannah or interacting with other people in societies, right? But they're not serving us very well when it comes to solving logical problems. And now the same applies to LLN. Neural networks as they're trained now are not terribly good at solving mathematical problems intuitively by just generating the next token. But give them a calculator, give them ability to write code in Python or tell them, hey, use logic
to solve this problem rather than just hallucinating the solution. And suddenly, like with humans, the accuracy increases dramatically. >> So, let's go back to Jonathan's question. We've had a hundred mathematicians saying that people should stop solving it. We had an anthropic employee resigning, claiming that there were too many safety concerns, continuing what we're doing. We've seen several CEOs of the tech company saying that we all should pause. Now, of course, as economists, Jonathan and I are both always wary of people that start to say to
the competition that we need to slow down. That's always a very interesting thing to tell your competition, you should play less hard. What should we make of this? What is going on really? Do we need to slow down? Does the regulator need to intervene? Is it even possible to intervene or will it fail anyway? because it's such an arms race that as long as it's allowed to continue somewhere, it's essentially useless to try to curtail it. What should we make of this? Well, I think that
the depends on whether we want a well-meaning answer or a practical answer. Now, a well-meaning answer is we should make sure that people don't start fires and we should also make sure that people are kind to each other. Yes, we all want world peace. >> Exactly. And we should abolish armies and police forces because we should have global peace. I agree that we should have global peace. And I agree that we are playing here with fire. We're playing here with the most powerful technology that humanity
has came up with. We probably playing with the most dangerous and powerful technology that may exist. Intelligence. could be the most powerful and most dangerous phenomenon in this universe and we are developing really quickly machines that are much more intelligent than us. Now I'm using this word intelligence very carefully because intelligence has kind of a quasi religious meaning and many people say oh only humans can be intelligent this is artificially intelligent or this is super intelligent or whatever the term they use but just processing information
and making predictions in ways that help you to survive in this environment. If we just use a different term maybe it kind of becomes less controversial. We are building machines that are really good at this. They're better than us. We are building machines that do not get tired like we do. We are building machines that do not die out of natural biological reasons. So if they get into some position of dominance as some humans do now and historically, you know, biological death will not bring end
to it allowing us to, you know, reshuffle our system. If Vladimir Putin takes over the political and social system and financial system in Russia and controls it in ways that the system isn't seems not to be capable of overcoming, we still hope that you know Vladimir Putin is not going to live forever. He will die and the new system hopefully less violence will emerge. with AI that reasoning doesn't work. If we settle in some system where some instance or some set of instances of AI become
dominant or become so necessary that we cannot find a way to switch it off. There may not be no similar natural time limits to the reign of this system. And when I say a thing like this, probably a lot of people in the audience would think, oh, Miho is thinking about Robocop, you know, coming and shooting at us with guns and somehow taking control in a way that is similar for us humans. Kind of reminds us of like how we humans take over >> would do
it. Yeah. >> Would do it. Exactly. And the reality is way more complex. Of course, >> AI may become dominant because it becomes indispensable and it becomes so good at helping us with bunch of things that we need to have done that we just lose an ability to imagine the life without it. Then by the way lose a practical ability to imagine life without it. But behind this is definitional. In the following sense, take the wheat plant. Yes, the wheat plant is one of the most
successful plants on the planet because it has domesticated us to grow it. So you could think of that as just definition. Did the beet plant domesticate us or do we domesticate the wheat plant? >> Jonathan, you're very right. So actually with the story of wheat is such that it domesticated us and when you look at many measures of life quality longevity health freedom from war you know general indicators or coralates of happiness in humans whatever it was 10,000 12,000 years since the switch to agricultural life
when humans regained the pre-aggricultural levels of happiness and prosperity meaning that this domestication of a human by a wheat plant actually had very bad consequences for humans for a very very long time. So I fully agree with you and by the way we had no way out because the moment you domesticated wheat and you thought oh my god it's a great idea we're going to make bread out of this plant then you started having more people and they had too many people to hunter gather essentially
those people would not survive hunter gathering. So they were stuck with this wheat essentially plant that also made them settle in one place, made them pollute the water around them, made them easy target for thieves and robbers and enabled standing armies and so on. So fully agreed with you. Now the thing with the wheat plant is that we found alternatives. We have rice, we have potatoes in Poland, you know, we have meat that we can eat instead. We have alternatives to wheat. I worry that the
moment the AI domesticates us more fully, the moment AI takes over the functions of a teacher of a legal system of political system, the moment AI takes over the role of communication device for us of management, it will replace organizations. Everybody's talking about we don't need mid-level managers because AI will enable leaders communicate directly with people you know on the front line completely not seeing the consequence of this which is you're still your organization is gone because you also do not need organizational boundaries if AI
can synchronize between me wanting to buy a sofa and a sofa machine you know sofa producing machine somewhere on a factory floor in China then we don't need organizations in the meantime Right? If AI can help us to synchronize the work of thousands of people both in like business context but also political context, you know, political party becomes obsolete, your bank becomes obsolete and so on. And by the way guys, this is great because this will increase at the beginning especially efficiency of those social and
business processes and it will skip the middleman and and so on. It's going to be great maybe. But it also means that we'll lose ability to live without this technology and we will not have alternatives. We have alternatives to wheat. We do not have alternatives to intelligence. >> Well, but you could have more than one AI program. >> Well, we could. The problem here is the more intelligent model you have, the more people are using this model. So, the model is becoming even more intelligent because
it has more training data. Moreover, the more intelligent the model, the more it can try solving problems, cracking problems, and then as it cracks those problems, it can learn so it can become even more intelligent. And then, of course, you know, we have a very intelligent model that people want to use, we give even more electricity to this model so it can do even more intelligently. And then essentially it's a snowball problem. >> Is a monopoly in the AI space an inevitability? Is that your prediction?
Right now we have entropic, we have open AI, we have Grock, we have Google working on it, Microsoft working on it, everybody's working on their own. >> A whole bunch of Chinese models. >> Chinese models. Are you predicting that it will naturally converge to a monopoly? >> I think and that's not the only technology like that. A search engine, by the way, is another example. >> Yes, for sure. >> Social media is another example. There's plenty of those technologies essentially are the winner takes it all
technology. And I hope that politicians and decision makers have learned from previous examples that we'll need some essentially artificially maintain competition on the market for the safety and well-being of all of us. Well, if you're going to rely on politicians, then I think we're all doomed. But I think this problem is very different to social media problems. For example, we already know that intelligence is multi-dimensional. You can easily imagine two models, one being good at one thing and another model being good at something else. Of
course, intelligence is multifaceted. But in humans, first of all, we see that people that are better at music tend to be better at everything else across the board as well. >> I agree with that statement entirely. >> And it also looks for animals, right? The smarter the animal typically the smarter it is at all sorts of things. >> That's true, Miho. But that doesn't mean that yet can do mathematics. So you can imagine a model that's particularly good at music might not be as good as
mathematics. And so they made special better arms. >> I'm afraid that the history shows that. So that was an open question and perhaps still is in many people's minds. Will we have tons of specialized intelligences or we'll have one big super intelligence? I believe that the future belongs to one super intelligence because competence in one area usually translates into competence in other area. Right? Being able to read helps you to generate the images and understand images. Why? Because sometimes there is a stop sign that also
writes stop you know. So understanding letters helps you with this also language describes a lot of facts about natural world that cannot be seen. And then by the way sound is yet another layer and so on. So essentially the more information processing cap because skills are transferable somewhat most of them because of the simple principle >> and that's what you said earlier right I mean we started this conversation by you saying that the large language models now using all these other inputs including pictures and other
things and combining them has led to a development of much better models. So the fact that the cross-pollination between the area is so important then that points towards a super intelligence as well. I would say I am going to disagree. Of course I was wrong three years ago. >> We're talking three years Jonathan. >> Yes. So in three years time we will decide that I've made another mistake. The reason I'm going to disagree is that there are certainly with human intelligence great returns to specialization. And
I just think that somebody is super intelligent makes advances nowhere and somebody who is super specialized makes advances. So I'm not quite sure. >> Jonathan, you are very right. You're exactly right. But this is because of the limited time. And there was actually a time in human history where we had polymaths that essentially smarter because it took you a week to reach the you know cutting edge of physics and then another week to reach the cutting edge of medicine. You know essentially a smart brain could
learn all of those disciplines because it just took them months or weeks to kind of become an expert in each one of them. And by the way they can would carry experiences. So now as the boundary of progress is kind of getting further and further away, it takes you know 20 years to reach the boundary of progress in finance. Once you spend those 20 years reaching the boundary of progress in finance, you essentially simply have no time to spend another 20 years reaching the boundary of
progress in economics or psychology or music. But if you had those 20 years, if your brain was not aging, if your brain by the way was larger, if essentially you could add extra layers of neurons very cheaply and if you had unlimited essentially amount of time or you could do it at extra speed or all of it at the same time, you would agree Jonathan with me that you know tons about finance. Some of this knowledge would help you to become a great politician, a great
economist, a great psychologist and vice versa. Right? There is transferable knowledge in finance. It's just we don't have enough time. >> Yeah. So given the Renaissance man analogy you just made, we do already know what the name of the super intelligence should be. It should be Leonardo. There's no question about it. Meow. This has been an amazing hour talking to you. Thank you so much. Thank you guys for having me. Thanks for listening to the Allse Equal podcast. Please leave us a review at Apple Podcasts.
We love to hear from our listeners. Also, be sure to catch our next episode by subscribing or following our show wherever you listen to your podcasts. For more information and episodes, visit all elseequalpodcast.com or follow us on LinkedIn. The Allseequal podcast is a joint production of Stanford University's Graduate School of Business and the Lauder Institute at the University of Pennsylvania. It is produced by University FM.