Raj Reddy The Future Of Ai Doomers Vs Abundance

TITLE: Raj Reddy : The Future of AI : Doomers vs. Abundance CHANNEL: CMU Robotics Institute DATE: 2026-04-17 URL: https://www.youtube.com/watch?v=ydnOSMbyyQo

---TRANSCRIPT--- So, the purpose of my talk is mainly to share my thoughts on this huge hype and confusion about AI, right? Uh I’m not sure I’m right, but at least I have some opinions about that. And so, uh you can take that along with everybody else. The hype comes in two forms. There are the people that are doomers, AI doomers, and there are people that are AI abundance, you know, that will have so much wealth created that everything will be free. So, so the question is what who is correct? Because these people that are making these statements are all people I respect. You know, for example, uh Jeff Hinton, famous for his deep learning research, and also got Nobel Prize for it last year. He says AI is entering an era of existential threat for the humanity. That’s very serious words. Exist That means you have to wipe out the entire humanity. Uh that could happen. It could also happen if an asteroid hits the planet. Every time you step out onto the street, you could be hit by a truck. All these kinds of things are very low probability events. And uh I’m not taking the existential threat very seriously. But there are other threats, other problems that we face. But the predictions are that we will be so productive, every person will be able to do a day’s work in an hour. Assuming you’re not just sitting around doing nothing the rest of the 7 hours, the assumption is you would create 10 times more wealth. And that that is everybody will produce a lot more, and then we will have a society in which everything is available for 10% of the current cost. And this leads to all kinds of other problems, deflation and all kinds of things.

Uh let me begin at uh my beginning. I started working in AI in 1963 when I came to Stanford as a graduate student. 1963 is very important year because that’s the first time the US government started funding AI research. Before that, it was kind of you know, faculty doing some research on themselves. DARPA, at that time it used to be called ARPA, uh started funding AI research in a big way. The reason it happened was the person that was in charge of DARPA was J.C.R. Licklider. He was one of the professors at MIT who joined DARPA, and he knew all the players in AI, uh Minsky, McCarthy, Newell and Simon here. And so, he funded those three groups. He funded MIT, funded Carnegie Mellon, and Stanford. And I was a first-year graduate student in 1963 at Stanford.

And when we started, one of the things, you know, we were talking about was what does it mean to do AI research? And for us, it was simply getting computers to do things that human beings do. You walk and see and pick up things and do whatever. But more importantly, we also prove theorems and play chess and do other intellectual tasks. So, when the original Dartmouth meeting happened, and if you read the original paper by Turing on intelligence, they mainly concentrated on what we would call intellectual tasks, tasks that require serious reasoning and thinking. But at the same time, human beings do lots of things effortlessly, speaking and hearing, walking, and these things happen at at the age of 1. When you’re 1 year old, and then you get better continuously. And um what we discovered is easy things for human beings are hard to do for computers. And hard things for for human beings, like proving theorems, are easy to do for computers. This is called Moravec paradox. Hans Moravec was one of our faculty members here. He was a graduate student with me at Stanford, and I I hired him to come and join us when we started the Robotics Institute, and he was uh you know, among other things, very prolific visionary in in the following sense, namely, he wrote a couple of books which I think are very seminal. One of them is called Robots. And and the the the interesting thing about this is Ray Kurzweil picked up on it. What they found was if you look at the computational power of different species, different things, they predicted it’s almost not linear, but it’s an exponential curve. They predicted by year 2030 or 2040, uh Hans Moravec said 2040, Ray Kurzweil said 2029, and he may be right at this point. Uh we will the computers will have enough capacity to exceed human intelligence.

And uh that whole thing has been driving AI research in some way. Most of us said, “Yeah, that’s interesting. What does it mean? You know, 20 And if we reach singularity, that’s the phrase.” Many of us said, “That can’t be true. Even if it is true, um there is another statement that John McCarthy made uh of the following kind, which he said, “Before we can have human-level AI, we need 1.7 Einsteins and three Maxwells, and 0.7 Manhattan Project.” He was just He was joking, but basically, you need a lot more intellectual advances, and you need a huge amount of funding. Today, we have both. We are spending, you know, this year, AI research, the data centers, and so on, are going to spend trillion dollars. That’s Maybe not the this year, but the the funding coming into that is about trillion dollars. That’s mind-boggling. Basically, those kinds of numbers for funding AI research were not there even few years ago. The thing that changed all of that was ChatGPT. Before then, people in the field knew what was happening, like transformer models, large language models, and so on. And we knew what might be possible, but when they started taking in huge amount of data and started building the so-called foundation models, suddenly, it became something a ordinary human being who is not in the field could appreciate. They could ask any question, and they would get an answer. And the answer was almost always reasonably correct.

So, the question was, “Where is this going? What are we you know, why are we investing so much money?” It is not explicitly stated anywhere, but I think I have one answer. I don’t know if you they’re charging for the premium models, $20 per subscription. If you wanted to use GPT-4 or GPT-5, they charge you $20 a month. GPT-3 is free. So, assuming you want the best possible result, a billion people subscribing $20 a month, that’s like 240 billion a year revenues, and for which you do nothing. Because everything is is already there, computed. The the foundation model is already there. And all that happens is uh you’re getting all this money. Uh that’s essentially what happened with OpenAI. They changed their tune. They when they saw this huge dollars, they said, “No, no, no, this can’t be completely free. We have We’re going to OpenAI was started as a not-for-profit entity, completely open.” And suddenly Sam Altman said, “No, no, there we’re also going to have a for-profit entity which will collect all this $240 billion.” And so when they needed more computing power, it was very obvious they could go to anyone and say, “Look at the revenue stream. Look at the profit on it, 99% profit.” So there a lot of venture capital and and also private equity, all kinds of things have come in. And so the the current estimate is we’ll spend a trillion dollars on AI funding. And I’ll tell you in a minute why I think that’s a bad bad investment. But that’s what is happening. There’s an assumption there which is everybody will continue to pay the $20 a month.

That gets me to why we are at that point, namely and so that I can explain why I think it’s going to be wrong. Why is Why are they wrong? For in order to understand that, you need to understand where the computation is going. If you look at the power of computers that today when we started in 1963, 1967 we had a PDP-10. When I came here, we just got a PDP-10 here. The fastest you could come kind of think of as a computing is one megaflop, 1 million operations floating point operations per second. Not even floating point second. Fast forward 60 years. Today the Nvidia chips the the Rubin R100 is 50 petaflops. 50 times 10 to the 15. Even if you throw away the five, you’re still talking about 10 to the 16 operations per second. Compared to 10 to the six operations per second. So all of a sudden in 60 years, we have 10 to the 10, that’s 10 billion times more computing power. One thing that was not obvious 10, 20, 30 years ago is when you get that much computing power, that would actually lead to new advances. There’s a famous saying, “Grove giveth and Gates takes it away.” The Intel was producing faster and faster chips, but Microsoft software was consuming all of that. And the current situation is people are able to use this 10 to the 16 computation. And in fact, they need more. They’re actually building uh Elon Musk is building in Texas uh computing center which will have exaflops. It’s 10 to the 20. Exaflop Exaflop is a 10 to the 18. It’s 100 exaflops is the current system that is being built. And over the next 2 3 years, that’ll increase by a factor of 100. And all of that computing is being used up to to build this more and more complex models.

This was not at all obvious in when I was we were sitting around looking at these things, I said, “Yeah, Moore’s law prediction might be right, but whatever you singularity won’t come. I’m not sure we’ll have intelligence. But today we have what is being called AGI. AGI is artificial general intelligence. In 1975, Feigenbaum demonstrated the first set of expert systems, AI systems that you might call special intelligences. These expert systems were the first AI summer and winter. And a lot of companies got started in the in the ’70s and ’80s. And uh they went public and people got rich. and then there was an AI winter. There was nothing had happened for 20, 30 years. And and the reason was these expert systems or rule-based systems could would work well on one narrow domain. But they don’t work well across all of things. So AGI by definition is a general intelligence, not special intelligence. So the question is, can you build a general intelligence? What does it mean? There are two definitions of it. One is AGI is a human-level intelligence. 3,000 years ago when when Egyptians didn’t know how to read or write, they still invented the script. That was the first information revolution because you could then communicate your knowledge to the next generation, next generation. If you have AGI, you should demonstrate something of that caliber that you would invent writing, reading and writing. Or invent zero for counting, which was done 2,000 years later. Now, I don’t know how to define it and how to evaluate it. It’s one of these mythical aspirations.

But there’s a different another word phrase that we use called polymaths. A polymath is like Ben Franklin, Aristotle, Leonardo da Vinci, and so on. A polymath is someone who knows everything that’s known about everything in the world. They’ve They’ve read every available thing. If there’s one polymath I can think of right now, that would be Elon Musk. For whatever reason, he’s, you know, apparently he’s read all the encyclopedia or whatever. He seems to know a lot of things about a lot of things. so the question is, can you demonstrate an AI a foundation model that can answer questions about any subject at at a level of an expert. You may not be the best in the world, but you should be 90% of the best in the world. So that’s a measurable definition, crisp definition. AGI is something that can answer questions about any topic at 90% level of accuracy of the world’s best expert. And we are there. How do I know? For example, one of the things I know is uh there’s such a thing as AP test. The AP board exams, college board exams. They’re They give 38 different advanced placement tests. somebody in the community from OpenAI or the other actually made their their model take all those exams. 38 exams. it’s very interesting. These models did extremely well in all technical, scientific topics. Like math, physics, chemistry, and so on. Biology. But they didn’t do so well in in English and and history and things. They were only got a B instead of a 90% accuracy. And on other topics like anthropology and archaeology and so on, they got a C. But they passed all exams. They got a passing grade in every exam.

And now, is that a polymath? Maybe. So I think we are there for that kind of AGI. What does that mean? And some people said, “You’re an expert in almost every field. I don’t need to go to college and get a degree. My assistant AI assistant will give me the answer on any subject.” Immediately, they jumped to the conclusion it essentially devalues education. Devalues going to college and getting a degree because you don’t need to do it. There’s a flaw in that argument. If you don’t know what the right question to ask, the prompt, right prompt, you won’t get the right answer. The current predictions are given that we have AI models that can be your assistant which can provide answers to various things, we will be able to do the work that we normally do of work in a day in in 1 hour. And therefore, either we can create 10 times more wealth or go to the beach for 9 hours 7 hours a day and not work at all. Assuming everybody produces wealth. Now, all of a sudden where instead of being a society of scarcity where you have to pay and and they have to work and to survive you’ll have a society of abundance abundance unlimited wealth.

So, those are the two theories that are going around. Immediately, they jump saying since AGI is already here almost here. The only people that don’t agree that is there say it’s there in 2 years or maybe 3 years or something. It all depends on what your definition of AGI is. Is that plausible? Can can we actually produce 10 times more wealth so that we have a society of abundance? Here’s the reverse of the other thing. If if if one person can do 10 people’s work the assumption is the other nine will be fired and they’ll be unemployed. So, either you’re have unemployment side where nine nine people are not employed or you have huge wealth. Everybody is And so, there are phrases like uh UBI and UBI yes, I think. Universal basic income. So, if there’s going to be a huge uh unemployment, what should the society do? we already have an answer to that from COVID days. If lot of people are not working and they’re not earning any money the government steps in and gives them some unemployment payment. It’s not UBI. Not everybody got it, but if you’re unemployed, you got it. UBI says everybody should get it. And then if they spend it, you get get it back in taxes and other things.

The other phrase that has been used is universal basic services. We only need money to pay for your basic services that you need including health care and education. But mostly things like electricity and water and so on. The universal basic services simply says the government collect enough money in taxes or otherwise that these things will be provided for free. We already have examples of that in society. So, for example we build roads all over infrastructure. And we don’t sometimes we pay ask for tolls, but mostly roads are free. Mostly water used to be free. It is no longer. And mostly san- sanitation various other services that you we take for granted are free. The question is uh how would those things work? Assuming we have AGI assuming we have all the expertise that is needed how would each of these services what I call Maslow’s hierarchy of needs. Everybody needs electricity. Everybody needs water. Everybody needs food. Everybody needs shelter. I think those can be done, but I’d like to concentrate on two things education and health care. We spend more money than on health care than anything else. Can you provide health care for free?

And and my answer is there may be models in which that can be made made to happen. But society will not go for it. Not at least if you if you went and told every all the doctors and hospitals there’s an AI hospital next door they’re going to provide everything for We we are a society of rules and regulations and laws and so on. And the government will pass a regulation saying you can’t do that. In Germany, they’re using robots all the time. The labor unions are very strong there. And there in fact, one of our professors was also professor at Karlsruhe, Alex Waibel. And he came to me and said this man came to me as the security guard said you’re not allowed to work on Saturdays and Sundays. Get out and go home. And so so it turns out that they’re very strict about some of these things. An extension of that is not only human beings cannot work, but robots cannot work also. So, they passed a regulation law saying robots cannot work on Sundays. So, these kinds of things will happen. So, even if we’re able to produce a society where AI is able to essentially eliminate a lot of the cost because they’re doing it. We face a problem that that will not happen overnight.

So the all of that is leading up to saying the threats are real. The opportunities and problems are real. But they won’t happen as fast as some of the people at the end of the end of the front end early adopters are saying namely it won’t happen in 5 years. It won’t happen in 10 years. My prediction is it’ll take 50 years. Not because the technology won’t let us happen. It’ll happen. It can happen sooner. But the what we call diffusion of this technology And not because it cannot be done but because the rules and regulations and the society will not let it happen.

In health care I still haven’t yet figured out even if I had you know the the the current statement is it’s not just AI is AI and human-like robots together will be able to do anything that human beings are able to do. AI for reasoning and and and planning and human-like robots which are predicted to be about $20,000 they’ll work four shifts non-stop whereas you you and I have to work only 8 hours a day or less. Predictions are that with human-like robots and AI health care can be provided at 10% of the cost. I’m saying it cannot happen overnight. It’ll take 20 30 50 years. So, if those of you live to be 50 years longer can expect to see what’ll happen and when it happens you’ll say Rafiq said this long ago.

The the second thing is of serious concern to all of us. We’re all professors. Are we all going to be out of a job? Predictions are AI tutors. By the way, those of you who been here for 20 30 years know CMU CMU and faculty members like Newell Simon Ken Kading and so on have been working on AI tutors for as long as you can do I can remember automated tutors. So, the issue is can AI tutors provide all the education that you need make it unnecessary to have classrooms and professors and everything else? It’s doable. uh to for it to succeed they need three things. One one-on-one education. There’s a famous paper by Benjamin Bloom called two sigma where they when they did the study they found if you provide somebody one-on-one education providing immediately answer to anything they might get stuck their performance goes up by two sigma. These students become A students. A students become A double plus students. Very interesting paper. Everybody said, “Wow, that’s a fantastic result, but we can’t afford it. How can we provide one-on-one education for everybody? That means half of the population will be teachers. And now with AI tutors one-on-one education becomes possible.

The second the concept of learning to learn. Learning to learn, in my case, is getting training every person’s brain so that anytime you have a problem you know how to learn the solution. And that’s mainly using ChatGPT and prompt engineering. And so when you go into your class, first grade, second grade, you’re given today’s syllabus problems you have to solve and say, “Go figure it out.” And you can ask your AI tutor and say, “How what do I do with this?” And it’ll kind of walk you through and it’s a one-on-one education. And so but it will make sure that you’re actually doing the things. Learning how to learn. That means you you need two skills. How to ask the right question. And the second is I knowing that you don’t know the answer that you need an one-on-one learning how to learn and just-in-time learning. You need just-in-time learning because with AI in your pocket a petaflop in your pocket, an Einstein in your pocket you don’t need to learn everything that we now teach. You know, I I took four semesters of calculus. You know, most of what I learned, in particular equations and differential equations and so on, I’d never use, you know, have no need to use. And they’re but they’re being taught just in case I need to know it.

We now are entering in a new world in which you learn when you need to learn the subject. But you we all need to have some basic skill. We need to know how to read and write, how to do prompting of questions and a few other But you don’t need to know everything. For example, you know, calculus you can learn enough about calculus in 1 day or 1 week or at most 1 month. And then the rest you can learn when you need to learn. Most of the population doesn’t need to more than 1 day’s worth of a calculus. Some can spend a little bit more time. All of these do not involve a professor or a classroom or anything. You’re on your own. but reality is you need social interaction among people. We need to to reinvent education where certain things that we don’t need to learn immediately can be postponed. Things that need to be learned for example, that all the skills if you want to learn how to play piano or if you want to play basketball or football or baseball or many other learning how to cook. These are skills where you cannot simply say I don’t need it, you know. Maybe you don’t need to know how to cook. If if but but in general if it is a skill-based learning, you have to learn it through some appropriate means. If it is purely informational learning, facts including multiplication and addition and you can you can do it whenever you need to it because you have a tool which will give you just-in-time

The future of AI and robotics humanoid robotics is here. Will happen. The only saving grace is it will not happen as fast as everybody is saying. I think it’ll take 20-50 years. That’ll give society enough time to find other things to do. And the the doomsday scenarios like extinction and so on won’t happen. You if if you have another intelligence AI and it’s coexisting with you it’s not in my view it won’t coexist. It’ll be your assistant. It’ll be a tool rather than anything more. We will have a a society in which potentially in 50 years from now we don’t have to work. If you work, you will work. And as Elon Musk says work will be optional. That might sound funny, you know, or I’m amazing. But where I grew up in my village half the population didn’t work. It was off it was optional. For example even if I wanted to if I was walking from here to the bus or something was carrying something somebody else will come and take my bag. I I didn’t even have to carry anything, right? That’s in a little village. But in general, that what I call Maharaja model. If you’re an aristocracy, you don’t work. You you sit and kind of propose things. I don’t know if you have a good idea of what king kings or kings and senior people do, but it is not they don’t work.

[Q&A]

Q: What do you think people should do right now to develop these new skills, and how should curriculum be reshaped to prepare students?

A: Every morning when I go to the class my teacher will give me saying, “Here is your assignment for the day. Here are all the things you should learn.” And that’s it. And then you’re supposed to learn from the thing. I’m reminded of Herb Simon, who was a professor here did the same experiment in 1980. And his his research topic was learning from examples. He had two classes. uh learning algebra through conventional class classroom. And the other one learning algebra from examples. So for the class which was learning from examples the teacher would come in and hand each of them a set of worked out examples and say, “Now you have understood those, you should be able to solve this 10 problems.” That’s the work for day for day. And he did nothing else. He would be sitting around. And then if somebody got stuck and didn’t understand they would raise their hand and he’ll go and answer the questions. And it was they did a very careful study of this Chinese Academy of Sciences. And Simon was the principal investigator. And what they found is after 3 years the students that learned from learning from examples actually remembered 50% more than the people that learned from lectures in a classroom. The same classes. My my assumption is in the future with an AI tutor the professor will come and say, “Here is the work for today. Figure it out.” You’re doing by yourself. Learning by doing.

Q: Will there be two kinds of people — those who spend waking time being really good at a field, and those who aren’t?

A: If you had the best neurosurgeon, you are the only one that can do that surgery, you may be working longer hours. But in general, the assumption is a humanoid robot with all the knowledge about medicine will be able to do most routine surgeries that a human being is needed. On the other hand, if you come to violin, if you want to play, you know, good you have to practice. You have there’s a skill. And there work is optional. You don’t you’re not being forced to do it. You do it because you enjoy it. And you may be that neurosurgeon might do the same thing. He’s doing the work because it’s something that he can do and he’s the only one that can do and he will save some lives and therefore he enjoys working on it. But it’s not required.

I’m kind of excited to be at this point in time in in our society because it will be going to be very interesting next 10 years of what will happen. And so we’ll come up with some some solutions. And we may see a few things like what happened with the Luddites in the Industrial Revolution. We may have some job displacement. There will be, for example, programming. I’m told, you know, the current models can do large-scale programming very easily like Linux operating system. So, if that is the case, then the current lead trained coders, programmers will be out of a job. But I don’t think that will be the case. What happens is even if they can do the programming, it still requires someone to prompt them, give them the right request and then right give them a right evaluation function, all kinds of things. There are certain skills human programmers have that will be become more important. And they will play a role of human in the loop who will essentially fix bugs or kind of provide guidance and so on. And so we need to figure out how to train our current programmers not only to see the program and see it’s okay and how to write an evaluation function for it, but also validate it and then say the there are four things that we usually require. What I we call RADS, reliability, availability, dependability, and security. Every time we write a program, when none of us worry about those attributes. But if it is a industry strength scale program that is being used all over the place, all of those attributes become important.

Q: How do people learn from mistakes? We’ve had so many inventions like rubber — dropping sulfur into it. Will that go away because AI will always be right?

A: That’s the interesting thing about human beings. They have multiple learning modalities. We learn from a teacher. We learn from mistakes, from examples. We learn by discovery. And not all of them are currently reflected in the AI models. But they’ll come. There’s nothing inherently impossible about it.