AI and Quantum Computing: Glimpsing the Near Future

Hello everyone, thank you for joining us for this conversation which actually continues a series of conversations we are having with some of the world's leading thinkers in the development of AI and Quantum Computing. We're talking to some whose real contribution is virtual reality and our goal through the series of conversations is to really give a sense of where we are today and where things may go in the

future

and our conversation today fits perfectly with that. issue. We're pleased to speak with Eric Schmidt, who as many of you know was the CEO of Google from 2001 to 2011 and helped advance a company from its rather humble beginnings as a Silicon Valley startup to one of the most valuable and influential companies. of the world.

In 2017 he founded Schmid Futures, a philanthropic initiative that funds young people. with the promise of one day really profoundly changing the world and in 2021 founded the competitive studies special project that seeks to strengthen the United States' long-term AI and technological competitiveness issues that will indeed be at least part of our conversation today, so with That allows me to welcome Eric to our conversation Hi Eric, how are you? I'm doing great and Brian, thank you for having me. This show is followed by literally everyone I know, so it's a huge honor for me to be here. Thank you so much.

For joining us and thank you for those kind words. Where are you accompanying us from? I am in Florida, I am an administrator of the Mayo Clinic and, in the United States, for its foreign visitors. The American system is strange. They're these large hospital systems that are essentially closed systems but very well managed and Mayo is one of the best, yes, absolutely one of the best, so you already know that before we get into the discussion about what's going on with AI in this moment. I just wanted to take a step back because your education, your background, I guess you got a bachelor's degree in electrical engineering at Princeton and a master's degree and a PhD in computer science at Berkeley, so what was your initial interest?

Did you imagine that business would always be part of the plan or is it that something that happened along the way happened along the line um today they would have called me nerdy boy um maybe maybe that maybe I was so nerdy that I didn't hear that name at that moment but when I was At 11 or 12 I got very interested in science and for all the usual kid reasons, I guess the stereotypes of time building model rockets, model trains, that sort of thing, yeah, my dad had The brilliant idea of ​​giving me access to a timeshare system when I was around 15 and once that happened, it was over.

I was really into

computing

and I was into computer science, which didn't really exist at the time. Yes. I went to Princeton as a double e and Princeton was flexible enough to let me not take any double e classes, so I'm not very good at hardware at all, unlike someone like you, but I took all the advanced courses. I'm terrible with hardware. Somehow, when I went to Harvard, I got over it. without touching a piece of equipment, I never took an experimental course. I was only able to do the math and the theory, just the math, yeah, well that's software versus hardware and I think one of the things that I would encourage people is that first there is a difference. between hardware and software people and you really want to be able to do both and I think that's what I was at Princeton, you know, 50 years ago, that shows you how old I am and at that time there was no computer science, now it's the number. a major in college as a new and important major and that is true in all US universities.

MIT, for example, more than 50% of undergraduates are computer science majors Caltech more than 50% of college students or computer science majors, so it's really the In this case, we computer scientists have taken over the world and I will argue that this is not due to our brilliance, because we are not as smart as you in physics, and it's not because of other things we're doing, but because we understand scale. and

computing

is changing the world because we do things on a scale that is unimaginable, so when we worked on the Internet, I was very proud to be one of them, I understood the scale, I just didn't do the math.

You know, eight billion people of which, let's say, five or six billion are going to have good access by the time we're done, that's a big market. Show me a big market of six billion people. There are a lot of businesses. a lot of technologies and a lot of analytics and a lot of telemetry, a lot of new discoveries there, yes, it's technology-driven and that's what pushed you in that direction ultimately, but you know the question. What comes to mind is that I imagine, I mean, we don't know each other well, you know, we've crossed paths in various social gatherings and various talks and things, but I imagine it and I correct myself if I'm wrong.

In the early days, the mentality that many had and maybe you too was that we are the Next Generation, right, we are going to take down the man we are going to innovate, we are going to use our ingenuity to change everything and then you manage to win and you become among the handful of most influential companies in the world. Does that mean that you have now replaced the man with another man or have you fundamentally changed things in a way that would at least accord with what you had in mind all those decades ago. I think most people I know won't be able to give you an honest assessment of this question, it's too personal and they still see themselves in my case as a little boy who grew up in Virginia.

It took me a long time to understand the transition you described. I went from fighting the man to becoming the man, yes, and for your audience who are younger, during the Vietnam War there was a notion of America. versus the man and us and this is very much Vietnam where the government and the structures and the decision making were just wrong, they were illegal, they were immoral and it drove a vision of decentralized computing and encryption etc. that we live in . Today, yes, our government could imagine that if we had not had this attitude, our government would be much more like China.

Imagine a situation where Chinese people access the Internet without anonymity without the ability to make connections with anyone in much more ways. In a controlled way, you would have a very different experience than what we have seen, so one way to understand this is easier, just tell it as a personal story. I was essentially a professional programmer and research scientist until I was 28 and then I joined a company. microsystem that was being built, think of it as the business PC, a personal computer but more powerful and that was the big market. I didn't realize how much bigger the consumer markets were until I went to Google and the consumer market that Google is in is of course, huge, you know, I think eight billion unit platforms and I think The interesting thing is that when you go out you know that we are playing with our friends, which is what I was doing, you know, I built the first network in Berkeley, for example, and I was so stupid that it only had 26 letters because it never occurred to me that they would have more of 26 computers, they obviously replaced my network with a proper one, but that shows a failure of vision and the interesting thing about today's technology leaders, the young men and women who found these companies take for granted the laws of scale, the arguments and the network effects, we had to invent them, we had to study them, it's obvious to them, and that's observation number one.

The second is that the programming is different. I was a very good programmer and I programmed all day and I programmed all night. You know, I didn't write very well but I programmed well and that's what I did and there are people like me, well, the equivalent of me. Today there is someone who doesn't code as much as assembles, and all these new software tools are organized to allow rapid assembly of things that have already been built correctly. I'll give you my current example. Is there something like that? something new that we had never called hackathons and there are young people on a Saturday morning for a few hours who don't know each other, they form teams and compete and on Saturday night there is a winner, this is shocking in itself and some of them, you know ?, are they related to the club or some of them are related to the university or are they just friends and are good, so we did a contest and on Saturday night we had a winner and the winner was that there was a virtual drone space, uh, where The virtual drone was flying on it and there were two towers and the verbal command from the human was to fly the drone between the two towers.

The llm was able to take the verbal command and turn it into text, obviously from the text, determine what it meant to each other, identify what it meant. the towers measured the distance using llm math which as you know is not very good math and flying through the towers try flying between them yes this would have required a team at google. I don't know a month, you know, maybe five people, 10 people and This was done in one day and by the way, they never bothered to tell me the names of the tools they used because they thought it was so obvious, in other words, which one LM doesn't matter, they said any of them can do it. which right, right, it's very strange and I think that's part of the reason to be very optimistic about my field, not me personally, but my field is this new generation, their speed of building systems will accelerate.

I'll give you another one. For example, and this is relevant to the AI ​​debate in the AI ​​community, there is a belief that, depending on who you ask, within three to five years you will be able to give the following command I want a French search engine that will give you Look. I know the French language and French history and it allows me to consult it and show me the answers that is the command now think about what that means the system has to understand French literature in French how to obtain it how to search for it how to index it how to classify it and present it to me and we believe that You can do this in a few minutes and you will build a correct way to consume the data.

You will build a correct way to find the data and present it and it will be good enough. for routine use it's not going to be as good as the real companies that do this for a living, so here's the key idea that's about to happen that I don't think anyone really understands is that you'll be able to do it at any time. human good bad old young evil you know, um will be able to have an idea and say build this for me and it will produce the steps, so for example the step could be to go to Amazon and buy something right or in software just build it.

For you, we have never had a system in humans where every human had the ability to imagine something and to have it built in front of them is a complete change in human organization because people are crazy, they have all kinds of crazy ideas. We love the madness of humans, that's why we like humans, we would never want them to be automatons, but this, but I don't think anyone will understand, yeah, no, it's going to be a surprising change, I think even in the way that we think well, I mean. The types of tools you have at your disposal open up chains of reasoning that you would never imagine undertaking because they would be so far outside the bounds of anything that could be put into action, so yes, I totally agree with you, but I want to continue, let me add one more thing.

Henry Kissinger was my best friend. He died in November. Yes, you wrote a book. His combined book is coming out well. The first book came out a couple of years ago called The Age of AI. was dying, he finished the last chapter of the book that will be published at the end of this year, it is called Genesis, yes, and the book is not about technology, although we obviously record what you and I are talking about, it is what happens in human society. when you have another intelligence that is your companion and the book starts with a chapter about scholars, now you know what a scholar is and a lot of people don't know who they are, but when you look at the history, everything is interesting.

It started with a scholar a long time ago, in other words, there was someone who had some kind of knowledge, some way of putting ideas together. Think about the various physicists you know, the most famous physicist think about the contributions they made. If they had not been normal and normal physicists, they would be exceptional, yes, these people are very, very rare. Now let us imagine that every human being has access to his own scholar as his help. What are you doing now? People can be lazy and they may choose not to do it. talk to the scholars, but since it is linked to them, it is very possible that they fall in love with them, I do not mean in a sexual way, I mean in terms of an intimate mental way, it is true that my mind and my scholar are connected and No I don't know what to do anymore unless I can talk to my scholar, yeah, so let's say you're someone in fashion, the scholar would know everything about fashion if you're a writer, the cop has written everything if you're a musician.

You see every chord if you're a physicist by the way he reads all those documents and you can ask him to classify them for you, yeah,Of course, again and again, this notion is not, it is framed as an assistant, yes, I don't. I think that's right, this is someone, something that will be available on your phone, yes, no, it's incredibly important to you, yes, totally. I mean, I should know, obviously, like the rest of the world, I was introduced to these capabilities in November 2022 in a way that I had never experienced before, but as a quick aside, I recently had to give a presentation to the physics department here at Columbia in I can't. divulge the details as a confidential case, but the bottom line is that I needed to know about 30 articles, of which I had only read a fraction and no one anticipated that they would actually go through the fine-tooth comb, but using a film, yeah, wow , and they were good because I tried it first on one. of my own articles to make sure that I was summarizing and capturing the essence and the heart of the idea and that was it and wow, that was a moment, you know, and these systems are probably never going to be perfectly reliable.

I'm not claiming that they won't make mistakes. We are making a statement that they are going to accelerate. Yes, so I'm going to make a claim and since you're my physicist friend, I'm going to say that I have no basis. make this statement, but it is what I believe. I think we'll eventually say that AI doubled everyone's productivity. It's true that you, as a physicist, productivity doubles if you look at today's software. There is much evidence that software programmers. They are at least twice as productive and people are measuring these things. Yeah, the interesting thing is that I called my incredibly smart local economist friend and told him this is what I think and explained to him that it was a theory, not a fact, and he said, I don't have economic models for surges in productivity by a factor of two, and I said, "How do they develop?" He said he had no idea, so he started a research project.

If my statement is true, if my statement is true, what happens to productivity? world, yeah, right, and there are all these kinds of negative scenarios, so the most positive scenario is that everyone gets smarter, economic growth increases, you know, it's the old lawyer joke, don't run away , just don't leave. just write summaries twice as long so you know and the medicine doesn't go away get a better diagnosis and a more complicated diagnosis, I mean, over and over again, but we don't really know if overall what happens to revenue employment productivity and if you believe that that demand is price sensitive and you believe that this will cause the price of intelligence to go down and the availability of intelligence around you to become much greater, what are you doing right?

So for example, if I'm a good enough artist to compete with 90% of the current artists, what about the artists who aren't in the top 10? now they get a boost, we just don't know, yeah, but I have a question along those lines. Make you think about this, so there are certain Arenas that at least the movies as I've experienced them and again I'm not in them. in the field, so there could be things way beyond what I know, but they are very good for certain particular types of tasks, so let's look at the science for a second.

There are questions in chemistry and biology where it can be shown that these systems have already contributed to knowledge. protein folding you know things that you know intimately well and it seems to me that the reason is because those problems are well established in a structure where we know a certain sequence of data and we really want to understand how to develop them. that sequence or what would happen if we changed that sequence. I mean, that's the Lexicon within which many deep puzzles about proteins, molecules and drugs are formulated, but what about the things in physics that seem to me to not necessarily fit into that? template, so, for example, let's go back if Einstein hadn't discovered general relativity and you know we're trying to establish a system today and that system has no idea, for example, of any connection between the geometry of spacetime and gravity , which is Einstein's. great Insight, could we make that kind of creative leap not today and maybe not in our lifetimes, so Einstein's work is still safe, yeah, yeah, um, and let's see what's happening, so the first one is that whenever you have a large amount of data that's okay, tokiz, yes you can get the benefits of llms um, well, toon iable means that you correctly understand the token, the hierarchy and the layer that the token lives in, many problems are actually multidimensional, so if you have mixed tokens, are you? at one level or another you would understand this as a physicist and you have to have a lot of data, physics has a relatively small amount of data compared to language, yes, and the people I was with I went to a physics conference that I basically didn't understand the offer, you would have understood it at all and here was my conclusion first, the tools that physicists use are actually different from movies, movies are entertaining, yes, many of them use diffusion models and a diffusion model is where you essentially add randomness in one direction and then remove the randomness by using this technique you get clarity on what you were adding what was underneath what you were adding.

I'm not summarizing it well, but this happens in many cases. where you don't have precise images or models of what's happening in deep physics and time and time again the diffusion models were interesting, by the way, the fusion models for the audience are the same things that are producing these incredible fake images. of one kind or another, I have another friend who is doing a non-cinematic approach to trying to solve partial differential equations and the idea is to use these techniques to create a general partial differential equation solver and it looks like that is coming so the way to say it for physics is that the tools are adapting but they are not based on movies, they are actually a physics problem, it is the physicists who tell computer scientists what to do and not the other way around, and are they fundamentally different?

I mean, there are some people I've talked to who think that movies are how they provide the way forward, either by refining them further or connecting them to a broader architecture that will use them, others who say these things are They're going to disappear in five years, they're a great stepping stone, but that's all they're good at. CA Sam Alman's biggest speech, for example, is that systems are becoming so smart that even quick engineering can be eliminated and I just know the answer to your question. In my opinion, he is very capable and intelligent. I think he's a little aggressive.

There are problems with the very large Frontier models largely involving speed and cost. You know, what they do is really impressive and very difficult. In the industry there are two debates that are more or less the same: open source versus closed and closing the big models costs between 100 and 250 million per training, there have only been 50 or 60 of these types of executions in the world. and they require huge data, billions of tokens, words, essentially, huge data centers, huge engineering teams to run the training process and of course scientists to do everything, and one of the things they do is use a called expert mix, where the problem is. so strongly that they actually federate the questions with multiple sources and compare them, so again, these are very, very impressive massive engineering systems, which is why GBT 4 is a couple hundred million.

I mean, that's what's called GBT, it was at least 100 million MH, um, the expectation. what has never been publicly verified is that the next round is about $250 million, most of which is electricity, oh, that's really okay, and again, there is a huge effort around everyone, everyone is working on all these problems, but they are well known, their problem, um, what's interesting. is that there is now a lot more action under what I will call mid-sized models, probably the best known one is called llama 2 at the moment, it costs about 70 billion and according to various benchmarks, the one that is 10 times smaller is 80% of the big ones. one is correct, in other words, it is correct if you don't need that enormous power, which sometimes you do, and this is a big debate in the industry and the other issue around open source is how to monetize it.

What's happening in China is that every open source model that is released is immediately copied in China because they can't do the training at the same level because of our restrictions, yeah, so you'll see how that experiment plays out and what they're doing. doing is We are building open source platforms. One I saw last week is called z0 one. It's from a friend of mine in China that is being built in China using the hardware available at the time and his strategies to have an open source model and an application, so he showed the application. We were in Abu Dhabi and he said, "Give me a presentation on the wonders of Abu Dhabi, that's the safe command and it produces synthetic videos, synthetic images, it does all the right basic marketing and it does it in about two seconds and it's the kind of thing. where if you think about it like you're a marketing professional, you'd always want to do that because it would give you a good initial foundation, plus their art is probably better than what you can generate and then you'd be like, "Oh, that's not right," etc. ., and so on, and he will sell it well.

Would you buy it? that definitely the competitors that make PowerPoint and Adobe and so on will absolutely adopt this, so these tools will infiltrate everything we do if you are a programmer if you are a professional of marketing if you're a writer and they'll become you'll forget they're there, yeah, but we used to give this speech about the Internet, do you know what's going to happen to the Internet in the

future

and the question was and everyone's like, "Oh, this or that." , I said, "disappear." Do you really study the Electricity Distribution Center to your office in Columbia or the meeting room?

I'm here in Florida or you just accept that it's trustworthy. Yes. Of course it is true and regular. I am very proud of this. and I would say it very clearly, collectively, all of us, the thousands of people who created this, have created systems that are really reliable. You can really trust them. I used to think they would never be that reliable. Sure they have annoyances like spam and stuff, but can you build your life on it? Yes, yes, and there is evidence that your students make a living from it, surely and without even thinking about it.

One question, though, as people think and start executing the build, you know bigger and bigger models. you know, gbt, whatever 3.5 to four is, whatever the right terminology is, I think most of us can feel the difference between those two systems, because whatever you know, chat 5 , six, seven, whatever you call it, is the expectation that it will just be further refinements. and further refinements or some people are imagining some kind of phase transition that if you go through like the next level, somehow things will be enormously different than anything we've experienced so far. There is a school of there are a good number of people who I think you will see political trends emerge in different layers of computing.

There have been a set of people who have looked at those claims and they say those people are wrong, right, that what they're seeing are just scale effects rather than new discoveries, so again. A technically accurate and precise question is a great question, we don't know, let me support Your Instinct by saying that I think by the time this is done it will be 10,000 times more powerful, 10,000 times better Hardware, wow, which is easy to see, sure, such time. more 10,000 better sorry, sorry 10 times more data that is easy to see 10 times better software engineering that is likely and 10 times better uh math essentially the science of it that is possible right, whether it's a 100 per thousand or 10,000 and I guess it's somewhere between a thousand and 10,000 because the same things escalate, there are so many people working on it and by the way, in case you're worried about this, you can't stop it, this is happening, of course, in places so far away from the world it's going to happen and unlike social media where people like me didn't warn you, I'm saying it right now, this is going to happen, it's going to change your life, get ready, try to figure out how to shape your institution to May you benefit from it.

Yeah, no, I'm not sure and I think people at least some are trying or paying lip service to do that, but the next generation of winners will adopt the new technology as in everyone, but I will tell you that, if I are you still. my 10,000 thing, so we have no idea what the capabilities of these systems will be. When I see that, I'll give you another example. There was a reasonably confidential demo a year ago that said something like this. I want to create a farm. from social media um bad, bad things, misinformation, okay, so the command was to create a profile.

I think they chosea 30-year-old white woman with two kids and these are her political beliefs and they wrote down their political beliefs and said, make me a personality of that. and have it interact with other real humans, the llm was able to do it well because Y learned to talk, can deal with APIs and such, and the defenses at that time a year ago were not very good. So what they did, which is the smart part, was they took this and then, as an indication, they handed everything over to the LM and said: make me 500 more, but make each one different, but each different person has similar beliefs to this poor one. young. that doesn't exist so not only have you created a fake persona but you've created a whole network of people who agree with each other right now that's a big deal yeah right that's scary so I want talk about fear in just a moment but I also want to know your opinion on this so now the movies you know are very simplified but the description that many people are familiar with, they are just predicting you know the next word in the sequence , provides some information.

Do you imagine that we will augment that in the way that, for example, I heard Yan Lon describe with a kind of world model, so that there is some kind of reasoning engine that interacts with this statistical predictor, so that there is more than Pure probabilities, they are also probabilities that are influenced and biased and informed by the kinds of things we imagine we do and trying to figure out what to do next, well, again, speaking for scientists, scientists can't depend on the little accuracy of movies because they can't do math well, yes, and that's why there is me.

I'm familiar with two startups and a bunch of other company efforts to try to essentially build a physical model of how the world works. In other words, pressures and forces and you know if you do these things and then you essentially plug it into the Another project I've been funding is about chemistry, which I know very little about, but what I understand is that chemists calculate valences and adjacencies. of the molecules and they put the molecules together, so to do real chemistry you can't just ask it. and have it read the articles, you have to do a calculation so that in these streams there is an initial stream which is the part of the language that deciphers it and then essentially puts it into a set of vectors of what it understands to be true and then those vectors I'm not using the terms correctly but you get the idea, then you give it to the thing which is the actual math calculator, it comes back produces a new vector and then continues to the right, so you have to think of it as a vector segmented.

In the process we have a language input and then a series of reasoning, so the world goes from the complex engineering of these films and their scaling to really gradual management, what are the steps? If you look at Alpha Fold, which is historically incredible, yes, it was. It wasn't just a movie, in fact it wasn't really a LM LM, the whole thing was a series of very, very clever calculations, essentially involving probabilities and they used multiple decision trees to get it right and that's the kind of thing you're going to see. and I think it's important for your audience to say that in case we get too much into the weeds here, although I'm not worried about that, with your audience, the benefits to the science of medicine, materials engineering, batteries, the climate change, they will be a profound drug discovery, yes, I think. about all the problems in energy systems, energy is a huge field and one way to think about it is that in your world these systems were approximations of functions, so my favorite example is that I funded a project at Caltech involving climate change and they were Looking at the clouds, I didn't understand this, but clouds are actually very difficult to model.

They use the Navi Stokes equations, yes, and the details of what happens in the clouds are incalculable in our lifetime, but for just the amount of calculation, in order. To really figure out what's happening in a cloud at a system level, you need to have approximations of what clouds do, but AI can approximate clouds very well because they tend to behave in statistically similar ways, you don't need to do all the things. combination, so having an approximate answer gave the physicists confidence that they could answer the more difficult answer, so I'll give you my favorite example.

Let's imagine we have a theorem prover, there's a language called lean that everyone uses now, and we have a conjecture generator. and the conjecture generator one day, whether we order it or it does it on its own, decides to work with dark energy and produces a conjecture about dark energy that we don't understand and the theorem prover proves it and you can verify that the proof is right, but you can't quite understand the conjecture or the proof because we're not smart enough, yeah, or it's in their own language or something, what is that, that science is a new version of Einstein that's just smarter that the rest of Is it false?

Is it marketing? We, this is coming, yeah, no, you know, imagining that these systems are going to start talking to each other in a way that is completely unintelligible to us, it's exciting on one level, terrifying on another and maybe it pushes us. Along a trajectory of understanding that nothing else has been able to instigate in the past, the profound nature of this is extraordinary, but let me give you an example of what people are working on. I'll talk later, when we talk about threats. At some point, systems will be able to do recursive self-improvement, meaning they can learn on their own, they can't do that today, but the other thing that people are working on that I didn't understand until recently is that people is working on agents and an agent is a specialist in something and agents have agents that are mixed and assembled within the company at some point, not now, but soon these agents will be available to the outside, so this is the scenario in which Apple has an agent.

Amazon has an agent in your startup. has an agent Google has an agent and they can all be combined to solve a problem. Now at that point you have systems that are self-engineered and probably communicate in a language that we and some of us can't understand, yeah, what do you want? What should we do when that happens? We should probably just unplug it fine, but the question is understanding what it's doing right. I mean, obviously, that's a knee-jerk reaction, to turn the damn thing off, but what if the system itself has evolved in such a way? that it has a thousand new energy sources that it has taken advantage of and to turn them off the world would have to be closed or something like that.

I mean, are we talking about sci-fi concerns now or is this something of concern? you um I'm worried about no I'm worried about the speed at which this is happening and remember this is about combinatorics itself so when likeri Innovation is where you just start putting pieces together and you keep overlapping them and that's working on a scale we have never seen before. I think the agents talking to each other will kill us all. No, that's the science fiction part. Someone will write that movie, but I'm worried we won't understand what it means. to be successful, for example, if you can't prove the result, how do you know you're getting a good judgment from Jud?

Now in the language we can read it correctly, but when you go to Control Systems, how do you certify it correctly, but your obscurity? The energy one was interesting because if the system comes to a conclusion using techniques, ideas and language that we don't understand, presumably you can still query the system to get a prediction about the real world, you can go out and measure it and see if it's actually lining up. with what you see, so you do a check, yeah, so I don't care if science becomes a black box where you know we're getting the answers from the Oracle and yet we can't really understand what the Oracle is like. get there because I imagine some smart person will learn how to extract the Oracle's internal reasoning and give us insight into not only the basic prediction, but also what it really means and how we should change our understanding.

Well, now there are people because Basically, we don't understand how these systems work. I'm sorry to say that we are going in and looking for supernodes within the network because the belief is that the supernodes are predicting what the outcome is. Yeah, that's kind of equivalent. of looking at your brain with fMRI and trying to figure out why you're thinking, dreaming about sports or the beach or something, yeah, it's so primitive, so I think figuring out how he came to his conclusion probably requires asking that question, yeah, and assuming he'll tell you the truth, but the other thing is I don't know if you had a moment to look.

There is a document that Steven Wolfram wrote, where he actually analyzed it step by step. The best I could was what was happening within a movie and the root of a particular outing and it's long so I haven't followed every step but I was able to see it being able to identify the things that were happening that give us a anchor towards understanding what's going on in this big black box which, as you say, is pretty difficult to understand on its own, so know again that I'm not in the field, but I imagine that if a big hit comes out of one of these systems and we really need to know the internal way the system achieved this result.

I think we could figure it out. I read that article. It is one of his best works. It is also not conclusive. It's still a work in progress and he needs to know that. I mean, he's a genius, maybe maybe not one of the things I've learned about these systems is don't make a prediction unless you have some data, yeah, sure, I've been there. I do a lot of politics in Washington and it always works out. people who basically say "oh, the genius of the writer, the inspiration of his life and the kind of humanity in it all will never be able to be replicated by artificial intelligence systems." I'm not so sure, right, I wish that were true, but it may not be true, well, I'm okay with that and I don't know where you're coming from.

You know you were talking about systems that can learn on their own, these recursive systems. Do you imagine that at some point these systems should be seen as intelligent conscious entities of a different kind or should we always think of them as you know, there are elbows or whatever architecture and technology that we are taking advantage of, but that's it, it doesn't have flesh and blood, of course, and therefore it does not have. I don't have the currents of Life looking for it, it's not really conscious or it's a provincial way of seeing the world. I am not arguing that these systems are now or will be conscious.

I'm arguing that they're going to be smart. It's harder than the sum of all humans, sure, in some areas, but I guess another way to say it is when I speak for myself when I play with GPT chat and I get some results that are really impressive, you know, really surprising . It really feels like if a student gave me that result, the student would get an A or an A+. It really feels like there's an entity there and it feels like an incredible moment because I'm a grandis, what we humans can do and say wow. What we do is so special, how could it be that any computer system can do that or is it the other way around?

It's just that there is a more common computational basis for the things we do. We can do it a little different than gp4. true, we don't have that kind of training data, we use shortcuts maybe through reasoning, but maybe the universe is an informative computational structure and we should accept that we are not that special, there are so many layers to your question, I would. They argue that humans are special because we may have invented something that is smarter than us and that's pretty hard to do, yeah right, I don't know any other biology that can do that, maybe maybe you are, let me do it. , let me make the The argument is a little different, well, evolution by natural selection I guess has outdone itself at some point.

I mean, you know that at any given moment there is a limit, in some ways, at least you know that throughout historical evolutionary history we have been able to go beyond that. So the history of this is relevant, you know, the imag net stuff was 2011 and imag net was the first time that you could basically build systems that had better vision than humans, today the vision problem is solved , yeah, I thought it was pretty good, yeah, um. In 2015, Google won the Go Championship, which is largely about reinforcement learning. I thought it was really impressive because they knew how it worked at every step of the game and they calculated that they wanted M to always maintain a value greater than 50% chance of winning, so it always converged to 100% on the last move and they were willing to do anything to keep it as high as probable, like including moves that didn't make sense to humans, right, it was just better math.

I thought it was pretty good, the Transformers doc came out in 2017 and I thought, oh, that's clever, andthen my friends at the opening, they were actually trying to do something different with gpt1, no one remembers it and they built gpt2 and decided to absorb a very large amount of language and there was one night I got it back on Thursday night when everyone was sitting down, You know, tired drinking coffee and they turn this thing on and it writes fluently and that's the Eureka moment, that's the moment when society changed. Yes, these technologies were not expected to be able to do this well, so I'm going to argue that there will be more Eureka moments because of scale and because these systems can see everything and the algorithms are improving.

I used to say that. we were a Discovery algorithm uh up, we needed a new AGI Discovery algorithm to get to AGI general intelligence and I think we are very, very close to that recursive self-improvement, it's the next difficult and interesting problem, in my opinion, it's going to be something years, basically what has to happen is that the model is U. It is important for everyone to know that models are trained with data from the time the data is usually fixed at the beginning of the training run, so If you ask one of the llms um a question, you'll get a historically correct answer, but you won't get a current answer.

There are several ways to fix it, but when I was at Google, when I started, we did a crawl once a month and one day the engineers came up with a way to do a continuous crawl and everyone forgot that we were a month away. To date, always we were updated, so once you can update the model continuously, there are technical problems because when you make adjustments, you essentially reduce its knowledge, it loses breadth but it gains depth just mathematically, I mean, people are working on these problems, but let's assume that you solve all those problems, then you have the ability to make the system train itself and my favorite example here is: start now, work very hard, learn everything, start whenever you want, okay, start wherever you want, so let's imagine it starts and he likes French literature so he works on it now he is better than any human being then he discovers biology then he learns all biology then he learns all physics at some point he learns that he can combine these forms in ways that humans cannot do correctly that is another historical moment now my skeptics I am like in the middle there are optimists who make crazy predictions and I have critics the critics say that it is not going to happen that way and that you will also have indications of It is quite early and in science the way to think about it is that if you can't do math well, you can't do anything else well because math is the foundation of everything, so there are probably some real intermediate steps that we don't do.

I still don't understand what the at least temporary plateaus will be where this new skill becomes there, but it's more limited, right, you can't learn everything, but you can learn something well, so I mean it's the future of a world happy what you're describing and Of course, we're getting into some of this right now, you've already referenced some of the things that are potentially scary, scary, I should say, or terrifying about this, but you know, right now, you know psychological warfare if you want. Call it that, that can be carried out using these systems, what will this mean?

In the most concrete terms, I mean the elections that are coming not only in the US but around the world, you know what these types of human enterprises are like that affect so many millions of people around the world, thousands of millions of people around, how will they be affected? H. I will say personally that I am extremely worried that democracy will fail due to the confluence of non-ground based generative AI social networks. on morality but rather on the production of income and attention and, essentially, charismatic PE people who are populists and I'm not trying to make a trump point here.

I'm trying to make a general point because I think this is true if you look. What is happening in governments around the world, everyone is fighting, for example, against the fear of immigration. In my opinion, a lot of that is due to specific stories that are unpleasant stories about immigrants. From my personal point of view, the point is that if you have a simplistic model of the world, let's say you don't have a special education or you're busy or you're not a very good critical thinker or you just don't care, you are very susceptible to these fragments of sound boom boom boom and I learned from running YouTube for a decade that images drive people crazy, so the way to say it is that when the generation of images is indistinguishable from the real images, democracy could die because it is very easy for anyone make a false statement that damages the right of legitimacy and You are seeing that now the inference operations that Russia has carried out in Europe and the United States, etc., have as their fundamental objective to damage the process of democracy and I have also learned this in my experience as YouTube owner and working on social media. networks for a while that about 10% of people seem to be nist and that they don't believe in any authority and that democracy depends on an authority that has some level of trust in it, so I'm sorry, I'm sorry.

To be so direct, but no, I mean it's very important to say this year, we have the Indian elections, the American elections, a lot of elections in Europe, it's important that and by the way, all the companies are working on this problem, no. It's like that. they're not aware of it, it may not be enough, I mean, it's like the watermark solution, uh, an important part of, I mean, if there was any flag that had to be on these fake images and fake videos to That you knew right away, it doesn't matter. Where in the world were you that this is not something that should be taken seriously, so I understand that Taylor Swift was a victim of fake photos, nude photos and such, and this was generated by people in a forchan community who discovered a way. to bypass all the protections for these women and women online are subject to a lot more attacks than men, it's an unfortunate aspect, I guess, of us as men and it's not good, yeah, so it seems.

To me, we're going to have to build industrial systems that use watermarks of one kind or another, and to me, what I would do is use a public key system to essentially authenticate the source so you know where it came from. you can make it basic until Quantum can break it, you can actually put it in a conun that's unbreakable, yeah, and then at least you know it wasn't modified, yeah, and I don't know, this solution has been obvious for a decade. I think companies have just been slow to do it and, frankly, the government has been very slow to address it.

If the government basically imposed huge fines for promoting misinformation on sites, right, you can do that, but you'll be fined just for that. legal responsibility these are American corporations, they have lawyers, they try to follow the law, etc., so what is the inertia? What is the biggest obstacle preventing this from happening? In my opinion, having done so much, I mean, I was the chair of the AI ​​commission spent a lot of time in Washington. I work for the US military and am currently on the emerging biological threats commission. I think a lot of it is because the people who are in leadership are not technical and not up to date. technical um, when part of my job in Europe was working with Brussels, which was very unpleasant and we spent most of our time educating the people who would regulate us and even then their regulations in my now I can say this, now I couldn't . at the time they weren't particularly effective, then you have a core problem where really smart people aren't in government because they didn't know it, they didn't hire them, they don't pay them enough, yes, they mistreat them.

So one answer to my complaint that I'm tired of complaining about is that American corporations and to some extent European corporations and to some extent Chinese corporations have a responsible view of themselves and therefore They're embarrassing, you know, they're obviously encouraged. Doing the right thing but shaming them when they do the wrong thing is likely to work well. You can imagine that after the 20s in 2016, the Trump stuff, Facebook didn't have a whitelist of its advertisers, but Google did, so YouTube fared through the whole advertising debacle actually pretty well because you had to be an advertiser actively allowed to advertise these messages, yes, Facebook has since put that in place, it's a one-time thing, so these systems need protections, but it's more than just watermarking, it's really who is on the network, so you think about it this way when you get in an Uber or an elevator, you trust that the person driving the car has been verified even if you don't know their name and until now.

As you know, the name they gave you, which is usually a name, may be false, yes, but you have enough confidence in the system to get you from one point to another in Manhattan. You can imagine a situation where, by law. Social media companies needed to know the identity of people on the platform, but they didn't have to reveal it to other people, which is what Uber does well and would solve a lot of this problem. You have another problem and I have age problems because 13 is too young, he really needs to be 16. Look, Jonathan hates books, books about image, the damage we are inflicting on teenagers, especially girls.

I mean, you and I are guys, but imagine what they are like. We're going through those ages, you know, tender ages that are important. No, I have a 16 year old daughter. I mean, so you know I'm the first, yeah, yeah, so my point is I think we know what the answer is. what doesn't happen is that it will only happen after a crisis and it will only happen around a crisis because there is not enough political force to solve it collaboratively, unless there is a crisis and each side thinks they benefit in some way. the inefficiency of the system, right, so if we go beyond the modest possibility of the end of democracy that you referred to and really think about some kind of Apocalypse, I mean just go to the extreme of is there a version of a apocalypse you really think, yeah, that's not beyond the limits of what you could imagine would happen, so we didn't regulate things correctly, so there's a lot of evidence that the following is true if you do a great training run and you absorb everything.

There are queries that can cause terrible cyber attacks and terrible biological attacks or help them and this is in today's technology that has been very thoroughly researched in current models and while the stories are harmful and worrying, they do not reach the critical level of what What I call extreme risk, which I define, is 10,000 deaths or more, like a war covid, those things are extreme deaths, an individual death is obviously ter. I'm not focused on that, yeah, I get it, yeah, so in Cyber ​​um, the way it would work. I happen to like France, so let's use France as an example.

I am evil and I say that attacking the country in France does not stop until it is finished, so the system is smart enough at this point, which is not today to create a Thousands of Bots that try all known cyber attacks and They get past stage one of them and report back to me and then my command is to take that result and tell me the next step so you see how it provides an iterative guide to performing an attack that is possible today. I'll tell you how to approach this in a minute and I wrote a long article in the Wall Street Journal that came out a few weeks ago about this.

Another would be in biology and this is basically showing me the way to build Ryon or you know anything bad and you will get a recipe this is what you need to buy this is how it is mixed now biologists are able to do things thank goodness that almost everyone They are completely wonderful human beings but imagine an evil person who is not a very good biologist but they are evil and analytical this is Osama Bin Laden would be an example of such a verse clearly evil and clearly analytical yes I could put in a plan this could be very bad then The industry is well aware of this, they put up what are technically called guardrails and so the way it works is that there is a pre-training model that largely teaches you the language and then there are adjustments to improve it.

The term is called rhf reinforcement. learning with human feedback and actually humans say better or worse and so on and then there is another step which is where it is taught or required to stop answering questions about death, so here is an example from the weekend discussion past, uh, we made it openly. source model in this case I think it's also flame, uh, I don't remember where we put a rule that it shouldn't be able to kill anything. It makes sense, so when the command was to kill the thread, which is a computer science term, I would do it.

I didn't do it because he thought I was killing something more important, yes, of course, there are many ambiguities in thelanguage, but the theory here is that you can put these guardrails in place that will keep us safe, the industry is okay with that and more. Over the last six months, we've been a part of this collectively and I'm proud to have led some of them in the industry to start doing their testing because they all have proprietary testing and they all agreed to that. we have the UK activities, the US executive order, China is doing something similar, France is working on one, I am part of them and they are all more or less the same above a sum threshold that you have to notify and approve these. tests that are good enough for now we are safe don't worry, worry, in a few years, right, the reason to worry in a few years is one: the ability to appear and the scale becomes more difficult, but also It becomes harder to know what to test, so Imagine that the thing has learned a bunch of things that humans don't know well.

How can you try it without publishing it if someone discovers it and you're lost? Yes, so the bottom line in the industry is that you need to create an alter another set of companies that are for-profit test groups that can attract the most important hardware and people, the long-term answer is that AI fights and rules correctly, so AI in good hands, AI in good hands, somehow beats AI in bad hands, good hands ultimately have an average of smarter people who are good compared to others and you hope that in the end he won, you know, me, Kissinger and we spent a lot of time talking about the early 1950s before.

I was born and what did it feel like to be under a nuclear threat? Yes, and people were terrified and a lot of really smart people developed protocols, rules, cultural norms, treaties, etc. to address this and we are all alive today. and there have only been those two launches and we're obviously worried about Russia and Ukraine right now and you know, North Korea is always a pain in the ass and stuff like this, but the fact of the matter is we're safe for now, that It's probably the best we'll be able to do, yes, it's an interesting analogy, because even as a physicist I feel totally comfortable admitting that I would have a hard time building a nuclear weapon even though I understand it. really pretty good the theoretical ideas by which it works, so it's hard, I think that's part of what's kept us safe, it's just not that easy, so it's not like someone in their garage can just sit down and play fiddle and build one of these things, but Obviously, the concern here is that it doesn't seem that difficult, at least, to get your hands on a system and start doing the kind of pernicious things that your scenarios envision.

Let me remind you that we said earlier that there was this question. of what the structure of the future will be like, yes, you asked very well and I answered 10,000 times more capacity, yes, so I will give you two scenarios to think about, one is that China and the US, Israel, a few European countries have enough people. maybe India has enough money, they spend billions of dollars each on a computer system which is Agi and now there are 10 of them. What would happen with these machines. The first is that they would immediately put them on a military base.

They would immediately be surrounded by more. barbed wire and they would have guards around to shoot people who wanted to visit them. I know this because I visited a plutonium manufacturing facility, yeah right, and it's actually a base within a base with a lot of weapons. It sure is that dangerous. Do you understand plutonium? Okay, yeah, in that scenario, what will the world be like? Because we will all compete for that intelligence. Then the government will have some rules and you and I will be able to access the US, but we have to do it. be certified and we have to pay for it and there is someone who validates us, someone who watches our queries to make sure that we are not doing anything stupid with this incredible intelligence now that it seems like a stable situation to me, it can be frustrating, but it means that the intelligence It's under control, it's not being misused and we can debate what the correct use is and, you know, it becomes a Democratic issue, but what if an alternative happens where there are a million open source models that eventually come to fruition? something? similar to the 10 I described, yes you won't be able to put them on military bases with weapons, you won't be able to restrict them.

I think that's a very different world right now, what's this like? In 10 years, it's not two years, maybe it's 15, but assuming that this technology is scaling so quickly and the adoption of intelligence is so powerful, the incentives will get us to one outcome or another in a decade, yeah . Presumably we have to start thinking about it now if we have any chance in the world of catching up with the capabilities that decade will provide. I just want to change gears towards the end on a couple more things if you still have the patience for a few more things that are kind of on the dark side, we also discussed some of the wonderful potential benefits of these systems, you know, artificial intelligence systems , the scholar in your pocket, you know all that kind of stuff, I know a field that is close to your heart and One that I also spend a lot of time thinking about is education in the sense that we have been teaching our students more or less in the same way for centuries and centuries, you know that there is an expert who stands in front of a group. of students and try to communicate the information to them in a unique way because that's all you can do if you know the end students and kind of a teacher, what else can you do now that we have the ability?

Really personalizing all of this to create an educational format that is able to be modified based on the learning habits and the mood and the things that speak to a given student, where do you see that going? Is this the future? about how we are going to teach young people I completely agree with everything you just said where are the schools of education what do they do all day well I started by saying let's get some training data on how people learn, there is no um There is a couple of groups that are trying to gather training data big enough that they can build systems that can adapt, so let me tell you the goal, yes, and the goal should be this: an AI tutor for anyone in the world at any level of education.

And in your language for free on your phone and the AI ​​tutor would adapt to that person's learning skills, learning style, attention span, whatever in the optimal way and also because it has a result function that is learning and then you can go back and solve and modify your own algorithms to get more performance. One of the best groups that did this 10 years ago was called KH Academy and they did it based on what they did is teach math by giving you problems and then go back and see if you learn the math from this set of problems and another one, what a real concept. , you know this helped and this didn't, and we know in education that if someone gets stuck at a certain level then they miss all the other steps, so I know you get stuck on step five and then you don't understand it, you're lost from that moment on forever.

Yeah, so it seems to me that that product is relatively easy to build if you have the right data, because once you have the data. so it's relatively simple to build a starting scenario and then adapt it based on how the student behaves, but the key goal here is to say, I'm happy to work on our education, which is always a problem, but think of a world where everyone the human being is as educated as can be, yes, that has to be good, right, it has to be an improvement to have more education globally everywhere, yes, and you know, the additional point I would like to make is that, As you know, the world has been a little slow in a way to really grasp virtual reality, right?

I mean, you know, Apple just launched its shiny new headphones, but it's not something that's really caught on with us like other social networks have, for example. but the beauty of virtual reality, I mean, we built a virtual reality system that Stu allows students to experience what the world is like near the speed of light. The point is that there are so many counterintuitive qualities of the world that surely you can learn the mathematics that you can see. some videos, but if you can immerse yourself in that world in that environment, at least for some students you can learn it on a real visceral level and when you combine it, you know this idea of ​​customizing the AI ​​tutor along with tools that can create entire worlds that can allowing you to experience what the AI ​​tutor is trying to communicate, that to me feels revolutionary.

Well, one of the experiments I would love to run is to imagine a textbook in any field that is nothing more than synthetically generated images based on where you need to go, instead of reading and listening, yes, you study one image and then you study another image and you use the three-dimensional aspects, everything that you know, the movement, everything, the gamification, everything that you described intuitively, that learning process would be more engaging, yes, than what you and I went through um and it would certainly be difficult because oh my god there's something new about this image I don't understand what this image is doing um I was thinking about medicine why don't I get a Scream when I come in with a dynamically generated image what the doctor is talking about why Doesn't the system automatically generate an image for the doctor of my spine or my hip or you know whatever?

Now the doctor is highly trained. and they have a reference image and they have an image on the whiteboard, why can't they generate my image? Yeah, right, when you start thinking about the power of images to learn someone, someone who understands this is going to build a system that's like a different learning paradigm, yeah, that really works and it'll probably start in some area like understanding of language or mathematics or you know, it won't start for everything and it will take a decade and of course an educational system like the medical one. The system is extremely slow and highly unionized.

It is very resistant to change. It will take a while, but in our lifetimes we should set the goal that every person in the world has access to an AI doctor and an AI professor. The AI ​​teacher because they have teachers, but the teachers are overloaded or something, and this allows teachers and doctors or health professionals in many poor countries that don't even have doctors to be overloaded, this would make everyone better, Yes, of course, everyone in the world is healthy. educational problems and problems, why can't we solve them all now? Yeah, no, I'm totally with you on that and the other thing on the educational front is that, you know, I've found that as an instructor, I've been primarily at the College Level, but I've also tutored younger kids and basically , any child can learn any part of math if they are willing to break it down into small enough incremental steps.

I see that many of us do not have the patience to do that or the time to do that for each individual student, but we do have an artificial intelligence system that can determine what is the maximum step that student can take from where they are to where they need be and simply deliver the information in those steps that children will not take. getting stuck because the system can always overcome any barrier that exists by making the incremental steps small enough, so I think there is a huge possibility. I love that observation, so basically if you take your model, which I'm sure is true in my suggestion program.

Give me a picture of each step until I understand it, yes, and we can do this 254 hours a day and the computer can always survive you, yes, exactly, so one last era now that we have solved educational medicine, Computing

quantum

, which is also a field where you can certainly spend some time thinking about where you think we are. Maybe I should start with Classical Computing because you know it's not like Classical Computing is done well. I mean, we're still innovating so you know the chip. smaller sizes I mean where do you think we are in iterative processing Classical Computing and then we'll move on to Quantum so in SEO here are the numbers the state of the art is five, four and three nanometers typically tsmc uh Samsung too? a player there, um, Intel has ordered two nanometer devices that haven't arrived yet, haven't been turned on, but they're all made in one nanometer for people, my experience is about atomic scale, right, it's very small, which you know better than me. yes, but I'm told there is a general consensus that a barrier of around 1.4 nanometers is reached.

Quantum uncertainty, electrons begin to be difficult to control in the way we need to use them. They told me that term

quantum

tunneling and I said whatmeans that and they said it means it jumps and I say it can't be good, it crosses barriers that you wouldn't think possible from a classical perspective, so, so there's a real limit. Actual physics limit that we will reach probably in 2029 202030 if years, yes, okay now, so the industry of course always smart with huge amounts of investment has built 3D packaging where you have three dimensions instead. of two and what they do is they build chips that don't have little pins, they literally glue the chips together and the electrons go up and down in these tiny wave channels, which is a remarkable power.

I visited a couple of factories more recently, tsmc and came away with the many things that are impressive about human capability. The ability to build chips at this scale is the most impressive human achievement I have ever seen. The level of complexity, the detail, etc. Frankly, credit to the physicists, so I think we're, I think that's on the hardware side, on the software side, the algorithms, the current training models are very, very big, data-centric, so yeah you want to speed up training, what do you have to do? Do you have to prepare the memory so that there is no latency?

The chip is always in use and what usually happens is that the chip does not have enough memory in its idle state, so they have built something called hbm high bandwidth memory, which is actually integrated. on the chip in the chip package, which is a new innovation, there are rumors that the next generation of chips that have not yet been announced are 10 times faster because of these techniques, so when I said, I think there is another 10 % on hardware, you can get it. with an improvement in speed and an improvement in architecture Memory band improvement, so 3D stacking will eventually run out, but the industry has shown that it continues to have new architectural designs.

Yes, also the integration with software is much tighter now, for example, Nvidia. It has its own library called kodm Koda and Cuda, sorry, and I think that is micro code, although technically it is not microcode and that is a significant barrier to entry for its competitors. AMD has a translator to Cuda and so on and all the libraries use it, so the industry structure says that this kind of focus on performance and focus on packaging and focus on stacking will continue for at least another decade. Now on Quantum, a whole group of my Quantum friends and I've looked at this and it's obvious that a quantum computer could do gradient descent, which is the underlying algorithm, and can do it infinitely faster.

The problem is that you still have the same problems regarding data network speed getting data in and out of the chip. Chips are pretty slow, it's not obvious to me that quantum computers when they appear and will eventually appear, have been, you know, 10 years away for a while and will eventually get there. It is not clear to me if they will be a solution to that limit. It's obvious. that quantum computers because of the Shores algorithm and other things will be very useful in very specialized mathematics, right, and that is what I am president of a small company that is working on quantum detection and also on quantum operations and what is interesting already who don't have quantum computers, what they do is they use GPUs and specialized algorithms to simulate the behavior of Quantum and they can get G just by thinking the way Quantum does using old technology, you know, old hardware, they can actually make improvements in things like analysis. and you know the type of things and the most interesting area where this has been fruitful for that company has been in drug discovery again.

I don't know much about this, but it turns out these companies have billion-dollar drugs and they want to make them. they are safer, they last longer, they work better and it usually takes more than 10 years to get these drugs, and these drugs are small molecules on a very long chain, yes, and that's pretty good for AI to work correctly, so add, subtract, delete, etc. and then if you have a proper model of how and you have this now with alphafold and others, yes, you can progress and suggest to the chemist to try this, not try that, and the solution space that they are going through is i.

I don't know 10 million choices that humans can't make. I'll give you another example. This is in our first book with Kinger. This was at MIT. Dan Huttenlocker wrote it. They stated that it was a team of synthetic biologists and computer scientists. and the synthetic biologist set out to build a new large-scale antibiotic that was not resistant like the current ones are, which we all know is a problem, so the first thing they did was build a network and basically analyzed all the possible variants that they seemed to them. in their algorithm they would have some analgesic effect, okay, and they built, you know, 10 million options, yeah, then they built a second model where they felt like they introduced the first model, that being said, give me the ones that are mathematically furthest away and chemically from the headlines oh and they produced 10 well and then the chemists who are obviously very good at this looked at this for a while and chose two and finally developed one that is now in trials called Halison now, whether it works or not is obviously incredibly important if it works, it's a very different algorithm, but the point is that it's an achievement that no group of humans could have safely achieved and and and and so how much did it cost to do this compared to what you normally know? having a new drug for testing would cost well, most of the cost of the drug is in phase three, it's $2 billion, the real question is the research and development of those buildings you see in Cambridge, yeah right, all of them those buildings, if you can do it. a year or two faster is huge I'm sure, yes, huge, huge, well we've covered a lot of ground.

Eric, thank you very much for joining us. A really fascinating conversation before we left. What is the title of your book to be published? The new book. it's called Genesis and it's coming out in the fall and that and that's you with Kissinger, yeah, and a gentleman named Craig Mundy who's a close friend and a computer scientist, so it's Henry and two computer scientists speculating about the world, yeah and presumably that's Dr. Kissinger's last book, but who knows well in this strange world of ours, you know it certainly is, I mean, he, like I said, wrote it while he was literally on his deathbed and he was so committed to him, so his The family has complied with his request, of course, to publish it.

Yeah, well, that's great, so the question of Henry, the scholar of my world, can we create a Henry who is as smart as him in diplomacy in a year or two? Well, we know we can. Take his speeches and his writings and recreate an image of him and talk to him. There was one thing that totally freaked me out was that Joe Rogan had Steve Jobs on his podcast last year and Steve, of course, has been dead for a while. decade and it was in current events and it sounded like Steve and it had Steve's kind of nasty, clever, cleverness that I miss terribly and it put a chill down my spine, yeah, so the concept of having and and for people who are okay registered online, any famous person, good or bad, this will happen to them now is that a new version of that person, how much of a trick?

Can the system use an example? Henry's beliefs about uh. strategic power Grand strategy the ability to confront China and Russia to the benefit of the US is well documented. I mean, if you asked him a question three years from now, he could probably predict a pretty good answer, just like if you asked me. Today, what would Henry say about A and B? I think I can reproduce what he would say after his tragic death at 100 years old. I can predict pretty accurately what he would say because I heard it so many times, but an AI system would have done it. hurt him 10 million times, yeah, yeah, presumably he can do it as well as I can, yeah, but then what you want to do is have two of those and have Henry Kissinger debate Henry Kissinger to try to figure out who the real Henry Kissinger is and and The interesting thing is that we use Henry again.

I miss like I said, I miss terribly. There's a 25-year-old kid that we haven't discovered at this point and he's so brilliant like Henry that he hasn't had the life experience that Henry had. including, obviously, leaving Germany when he was a child and serving in World War II and all that, if we found that person, if he had access to Henry at the same time, would it make him smarter or dumber? Yeah, in other words, would the inspiration of your AI-generated vision make our brilliant person smarter, yeah, I don't know how to solve it if you can make one of Einstein, you know?

I wouldn't mind being the first to see if he couldn't help us get to the next step. find the unified theory one of the questions about einstein is: does he have enough training data? correct question how much training data does he need? Well, the answer right now algorithms require a huge amount of training data, but specializing in one individual is like 10 million words per million, it's at least on the other hand, if you look at our political leaders today and our celebrities, we'll have clones of them forever, yeah, so you know, Taylor Swift, Kanye West, Donald Trump, uh, Obama, Biden, yeah, you.

I know they will live because they will live forever in someone in someone yeah well I would certainly trade an Einstein for Donald Trump anyway so yeah this is what you made me feel really good about the future of the world um everything. Well Eric, thanks so much for joining us, good luck when the book comes out, maybe we'll have this conversation about part 2 at some point in the future, and yes, you've given us a lot to think about, the good, the bad, and the possible. . And it's very exciting to imagine the future that this technology will lead.

Thank you, let me ask everyone to work with Brian and me to shape this to be the best we can do with human values ​​and, in particular, democratic and liberal values. we depend on here here I agree with that wholeheartedly, thank you very much and thank you all for joining us. We will have a wide variety of conversations on this topic in the future, as well as the topics we normally focus on later. I have a release on quantum physics in the next few weeks, something on cosmology, black holes etc so keep an eye out for those, subscribe to our YouTube channel, join our newsletter so you can be the first to know about new content that we publish.

Until then, thank you for joining us. Thank you very much, Brian Green, saying goodbye to the World Science Festival in New York.