Google’s James Manyika: Who Actually Gets Paid in the AI Economy

It's such an exciting term that possibilities are extraordinary, but the challenges and the risks that complexities are also as important to make sure we're working on both things. This is James Manika, Google's Senior Vice President of Research, Labs Technology and Society. He spent years studying how technology changes work and how workers can benefit from those changes. What do you think will happen in the next five years? The one thing that will happen is that there will be some jobs that will decline. One of the biggest jobs at risk. People are doing a lot of routine tasks, testing code, for example. But I think it's important to complete the picture. The other parts you'll see are jobs gained. New occupations that didn't exist before. Is there a warning sign you would be looking for that could accelerate the job loss? There's several things I've been looking at that could. James, thank you so much for doing this. Oh, thank you so much for having me. I like how you're plugged into everything that Google does when it comes to technology and AI, so you're the best person to discuss so many topics. But I wanted to start with something that my audience is particularly worried and interested in is what's happening to our jobs. And you were willing to take a bet. I watched them in podcasts that we won't see 10% of jobs automated away within a year. That was your bet. This is a topic I've been spending quite some time over many years. The question of what happens when AI intersects and impacts jobs. And I've always had the view and a lot of the research that I've done and actually many economists have done is basically describe that three things will happen and all of them will happen. We can debate the proportions of them. And that's because some of the constituent tasks involved become easy to automate. So we'll see some jobs lost over time. Well, I think you'll see that. So for example, anytime people are doing a lot of kind of routine tasks whether it's testing code, for example, or looking for bugs or those kinds of things. So there's a whole range of tasks. In fact, by the original estimates that I did and many others did back maybe several years ago, we keep updating these. The proportion of jobs that could be lost that way added up to single digits number is a percentage of jobs saying the US economy. So but you will see some jobs lost in that sense. And the jobs gained will be of two kinds. On the one hand there'll be some jobs whose demand will just go up existing jobs. And we see some of that today with electricians and a whole range of jobs where the demand will actually go up for those jobs even though they exist. And this is the story of what happened with radiologists. What happened with even bank tellers that they're advent of cash machines. So there'll be some jobs gained of that kind. The other jobs that will be gained will be new occupations that didn't exist before. There's a long history of this. David Alter and MIT and others have done some research that basically shows. If you look at the US economy back in the 50s compared to now, something like 68% of the jobs we have today, occupations, just didn't exist before. It's kind of like if you had looked in 1995 and said, is there a job called web designer? Consist creator. Consist creator. Consist creator. Podcasts. We say no. But now in hindsight we're going to say, oh yeah, those are real jobs. So jobs gained will be of two kinds. But the biggest effect that almost most economists have looked at this will point to is the jobs changed, which is the third effect where many jobs, the content of them will change even though we may still keep the same name for them. So the reason I was taking the bet was because, you know, starting maybe about four years ago, there was some AI folks were arguing that in two years time, we're going to have a wide part, 50% of white colored jobs. And that's the bet I was willing to take. I took a bet with a couple of the people who were making those pronouncements. And you were right. And I asked them two years later to say, times up. What's happening? What was their response? And I said, no, no, we'll reset the timeline. So they'll say, okay, no, it's the next two years. But I think this expectation that we're going to see this massive job declines. I just don't think what happened. It is in the kind of 10, 20-year timeframes. Because part of what happens, I think the effect of technology, while technology is moving very quickly. And the capabilities are extraordinary. I mean, we're right in the midst of it ourselves and many of the other frontier labs. The rate at which that translates into economic activity and into the economy is much slower. Look at the pronouncements people were making last year. Look at what the job numbers actually say today. Jobs are actually up. Will the supply of new jobs exceed jobs lost? My expectation, the growth in jobs coming from both new jobs and existing ones, expanding will be bigger than the declines we see. I mean, this is, you may have seen the recent, I think the economists did a good roundup of this, which basically said, "at most, the most jobs we can attribute to having been lost from AI is what, I think that number is 200,000. And the jobs that have been grown as a result of the AI effects were actually more than a million. So that's the data so far. And even in areas that we all are very close to, like software development, those jobs are actually up. Yeah, I feel like the actual data. So I think, you know, this is not to say, Marina, that we should, this will be the forever picture. We should take seriously the possibilities of the impacts of this technology on jobs and work. It's just that it hasn't happened yet, or the worst things that people expect have not happened. They may not happen in my view. You touched upon the economic effect. And this is very interesting, because we have such powerful models. And they're capable of so many things, escaping, breaking rules, etc. But at the same time, when it comes to practical tasks, I think your Atlas study suggests it's about 21% that could be automated with AI. Why is that so? What is going on with the other 79% tasks? Is that the model is not capable, or we're still adopting the technology? I think a lot of it, because the 21% you're referring to, that was the median percentage, the percentage of a median job, where the tasks were being used, where AI was being used in those tasks. And even then, it wasn't even to complete the task. So what we found in the Atlas study, I maybe I should explain that. Well, that study looked at, was we looked at all the activities through the Gemini app, and Gemini APIs, and also an AI search, an AI mode in search. And that's a total of close to two billion people are using, because Gemini app is now about a billion people using it monthly, and AI mode is also about a billion. And then people were accessing the Gemini app. So we sort of got a huge chunk of activity directly. And what we found was that the adoption is very wide, but very shallow. And the 21% points to the shallowness of it. And we also found that in most of those tasks, these were assistive tasks, as opposed to total automation of the tasks. So the human is still working with the technology. So I think that's what will take much, much, much longer, when, you know, before we get to whole tasks being fully automated, keep in mind that there's lots of bottlenecks along the way to complete a full task. And many occupations consist of many, many tasks. And I think that's part of what you're seeing. So what is the bottleneck? Is that the model? Because sometimes you use AI, and you start feeding it data, and then the answer is completely different. You're like, oh, that was how I was supposed to use the app, right? Right. I think it's both things. I think on the one and the technology is to what many people refer to as jagged. So you often need to check the output, you need to evaluate the output. And then sometimes the tasks you want just takes many, many, many more steps. And the technologies are quite ready to do that yet. So that's a big part of it. And now, of course, the technology will improve. And we already know that the time horizon of what these agentic systems can do is actually extending all the time. So that will get better. But the other part of it is that people interact with technology. We haven't involved yet the workflows enough to make the technology fully assistive or to fully automate them. And in any case, most people seem to find greater value when the technology is being used to complement human activity. The combination, the car, it seems to be much better when it's a compliment as opposed to a full automation. The tasks where full automation is the optimal output is still very small. That's a great point. And for someone who's worried about their job, what should they be doing to identify how risk it is for them to not adopt AI? Well, I think part of it is just as humans and as workers and as people just continue to evolve the value add of what we do versus what the technology is able to do. So in any of our tasks and activities that we do, there are parts where we still bring the judgment, the expertise, the ability to check and the ability to frame the right questions. And the technology will be good at some things. And you as a human will be better at some things. So I think this idea is people to constantly evolve the value add of what we do working alongside these technologies. I think that's the ongoing continual learning of what we're going to need to do. I'll say one thing, though, which I often worry about, which I think people throw around this idea that if you're exposed to AI, I think somebody says if you expose your host, I'm just the wrong way to look at it because you've seen the exposure to this technology go both ways. So some people are exposed, yes, some of the tasks that they do become easy to automate. But in other cases, people exposed are actually more able to capture that in technology where the combination becomes more powerful. So I think that idea of using exposure as a predictor of the worst outcomes, I just think is plain wrong. Because I think both exposure can lead in both directions. Because then if you're exposed, 50% of your job is now done faster, but that means your strategic thinking, whatever your taste now matters more. It matters more. It allows you to be more productive because I mean, this is the case of the wonderful case of the radiologist, radios were clearly exposed, but what actually happened because there's so much more demand for what they do, they became actually more productive and as you growth in radios. The electricity of demand. Went out. Right. It's the same thing that happened with, historically, the example of the bank teller. People had predicted back in the 1970s before ATMs were prevalent that we're going to see bank tells us the most exposed obviously because of the activity of taking money in and out of a cash machine of ATM, but what actually happened, the job of the bank teller actually grew. So the demand elasticity was actually up. So I think often we forget the demand elasticity that happened when the technology complements what people do. Yeah. There are some activities that demand is actually outstrips the complementary effects of the technology, but there will be somewhere it actually goes the other way. So just exposure on its own doesn't predetermine necessarily where the outcomes will be. What happens to your salary? If you become more productive, what are you seeing now, maybe inside your organization? I think this is one of the most important questions that the wage effects dissolve. Often we focus primarily just on, you know, what's happening to jobs in displacement. I think the wage effects are important because what happens is, so imagine if you are somebody who you're complemented by the technology and your output is actually even more productive and more valuable, your wages should go up with that because you do more valuable work complement by the technology and its expense. I think the risk comes when, in fact, you're complemented by the technology, but the truly value added portions of your work are being done by AI and what's left over is less value added. And you could have adverse wage effects. So the wage effect question, again, can also go both ways. I think this is one of the things we're going to need to think a lot about economists typically think about this idea of the marginal contribution of labor to any particular task. And that can go both ways, but this is one of the things we're going to need to worry about and pay attention to because I think we'll see some occupations, particularly in the demand for what they do, complement by technology, goes up and there's greater demand. I think we'll see those wages go up what we should. And though in people are becoming more productive and creating more value added output, I think the wages should go up. So let's make sure that happens. But there will also be some other adverse effects where the opposite could happen. Yeah, I think people who are seeing the increase in productivity should be also advocating for themselves. Oh, I think so. I think so. Yeah, because I don't think businesses will be able to catch up that fast, especially small businesses. You're like, you're just not accounting for that. No, I think you're exactly right, Marina, because one of the things, and this has been a long running trend, even before this air moment. If you look at the history of the last sort of 30, 40 years, the idea of productivity is that as productivity goes up and GDP and output goes up, wages should go up with it. But we've seen for the last 40 years the growth in wages go up, but not as fast as the output has gone up. I think that's a structural challenge in our economy. I don't think AI introduced that, but AI could actually make that worse. So we should think about how do we make sure that as people become more productive, that is reflected in their incomes and their wages as they make all these extraordinary outputs and value added contributions? Well, the advice to workers who's struggling with this should advocate for themselves. And I think, and I think, you know, the more they make their particular contributions unique and distinctive, particularly in a way that complements what the technology does. I think that should lead to high output and wages. You could argue with seeing some of that already in some occupations where wages are going up. The questions are they going up, commensurate with the productivity contributions that are coming out of there? Well, I think that's the question. Yeah, and I think because of that, we're also seeing transition to more independent workers like contractors who do a particular function, but for multiple companies, and AI allows them to do that at an increased speed and then you're charging for the output. So that is actually one of the things I'm most excited about. In fact, one of the things we're looking for, and you mentioned our Atlas study, that's an ongoing study. So hopefully you'll see, we're going to keep kind of producing output of that. So one of the things we're watching for is the emergence of new activities and new kinds of firms and new kinds of work emerging because that's what's happened historically, where any time a new technology comes up, you start to have new kinds of firms. New kinds of businesses and new kinds of work start to enter the picture. So we're constantly trying to watch, what is that? Have we seen that before? What's that activity? And is it growing? What do we call it? What do we name it? I think that's going to be one of the most exciting things to watch. What is your advice to entrepreneurs, what markets should they be looking at? I think we're going to see a range of new services emerge. I think the difficult thing, by the way, is to predict what those are. I think one of the things that's been interesting for me is somebody who's interested in the history of these things is that most new occupations and new firm activities, we've only been able to name them in hindsight. At the time, there's a lot going on, there's a lot going on, and it's hard to actually name it and say make those predictions. So I'd be very wary of our making predictions, but I have no doubt we're going to see new kinds of firms and new kinds of activities and new kinds of work emerge. I think the great opportunity for entrepreneurs is to experiment, build new kinds of firms, go after the market opportunities, what are new work being emerging. So are there people now overseeing agents, are there people now watching agent orchestration, are there people now doing new kinds of evaluations, are there new services that large and smaller firms need, are there new, so those are the things I'll be trying to experiment, if I was an entrepreneur. Yeah. And also, I think I see as someone who started as an employee and then became an entrepreneur, you're seeing the process inside your company and that's your idea, that's your business idea. Exactly. So much more possible to execute on it. Okay. To kind of wrap up this AI jobless section, is there a warning sign you would be looking for that could accelerate the job loss? There's several things I've been looking at that could change the dynamics I just described. I think on the one hand, I'll be looking to see, are we extending the capabilities of the technology to do much longer time horizon tasks? Have we now solved and addressed much of the jaggedness that we've seen the AI capabilities themselves? What's the time horizon? Is it now, what is it now, 16 hours? For some tasks, it's up to 16 hours, but it's very specific. I mean, I always think that sometimes the emails and the benchmarks don't tell the full story of what actually happens in the real world. Yeah. So in some simulations, we've seen these things run for much, much, much longer. There's actually a wonderful study that economists at MIT have done and look at the time horizons and that is a predictor of at least an early warning sign or at least an early sign of extending capabilities. So I think we should watch that in the real world, not just in the e-vows and the simulations that we run. That's one thing I'll be looking at. The other thing I'll be looking at also is just what I mentioned is the jaggedness. I mean, it's still extraordinary to me that there's still some tasks of these models of terrible at visual reasoning, for example, to have to do all kinds of clueduaries to solve for counting, for example, because these systems are not very good at that. So once you start to see more consistency across different kinds of activities and tasks, that's something to pay attention to. The final thing I'll just also mention, which I think will be a big factor we should watch for in the labor markets, is whether robotics comes along as quickly or not. The AI systems are making a lot of progress in terms of the cognitive activities and thinking activities. But robotics, even though it's making progress, or you'll see all these demos, is still lagging quite a bit. The only real advance we've seen in robotics is in driverless cars, Waymo. In some ways, it's an AI on wheels, it's something like that. But robotics still lags. But if robotics was to fully catch up, that would change it down. And I'm quite a bit because something like, if you look around the world, I mean, I think the number's roughly 60% of people in work in the world do work where it's dominated by physical tasks. So if that catches up, that's an important scenario shift that we'd have to think deeply about it, what the implications and opportunities of that would be. I'm very excited to talk to you about science, because from what I'm getting by talking to scientists and entrepreneurs who work with scientists, that we're seeing an incredible progress where we can do more research. But at the same time, the scientist's job is now running multiple experiments and verifying the AI's output. Is that what do you see as well? The intersection of AI in a sense, like one of the things I'm very excited about, very excited. Because I think it's one of the most underappreciated big advances that are happening today from AI. So that cuts across everything from biology to physics, chemistry, a whole range of the natural sciences, but also recently in some of the formal sciences, some mathematics and algorithms and computer science. So the progress is extraordinary, but I also think it changes often the nature of the work and where the bottlenecks are. So one of the things that we just put out here, you know, it's study where we're looking at. So how is scientists actually using these tools? And what's interesting is you see two patents for now. We'll see whether that persists or not. There's a lot of use of the general models. So I think Gemini, people using Gemini, the frontier models, you do a lot of the, you know, the operational tasks, writing papers, researching papers and those kinds of things. So that's kind of roughly half by the study we just published recently with work on some folks at MIT. The other half of scientists still using very specialized models, by specialized models think alpha-fold, a Nobel Prize winning model, or any of these specialized models that are either working on proteins or cell structures or connectomics or genomics or any number of areas. So most scientists are using both. And often in tandem. So they may use Gemini for the general task, but then come back to these specialized models. So that's what's driving a lot of the progress and lots of great results. I'd like to come back to them. But to your question is to where are the new bottlenecks? It's actually quite interesting. This has led to a huge generation of hypotheses, ideas, simulations, things done in silicone. For example, one of our models, Genome, was able to predict crystal structures in chemistry that were like over two million structures that were theoretically possible in simulation. But then the task of verifying those is now the bottleneck. In fact, after the 2.2 million crystals that were predicted, I think only a few hundred thousand were actually looked at as viable for synthesis. And even then, when people then try to synthesize, I mean, the number dropped to like a few thousand. So on the one hand, generating ideas and possibilities is almost like easier now. But the bottleneck is lab verification and then often in vitro or in an unnatural cell structure. We had the same experience where one of our models, C2 scale model, was able to predict these pathways and possibilities in cancer treatments, for example. But of the possibilities that came up with, we then worked with researchers at Yale to verify them in the lab. And only a handful were things that were actually reasonable in the lab. And so now we have bottlenecks at the backhand of verification and validation. But also the process of generating possibilities in hypothesis seems to be much easier. The possibilities is much more expanded. I wonder if overall, then, is AI speeding up science? Or because now we have this new bottleneck of verifying 2.2 million different crystals? Well, I think when it comes to science, I think AI is actually doing two different things. There is a part of where it's speeding up the process of research where you can now look at all of it, all at once. And like in the protein structure predictions, you can look at all of them. Recently, you may have seen, we actually looked, for example, at all the single letter misspellings in the genome. This is the alpha genome atlas and we looked at all 9 billion of them. So this idea that you can look at all of them quickly is clearly one effect from science. But the other effect is actually being able to do things that we couldn't have done at all before. It's not really a speed up thing. So for example, in the case of some work with you in connectomics, so connectomics is the area where you're trying to send, if you like, the wiring diagram of the brain structures. And actually, we hit a milestone recent a couple of weeks ago that we published working with the Howard Hughes Medical Institute, where we did the first full brain mapping of a fruit fly. Drosophilia was the first times I've been done. But the reason that's interesting is also much a speed thing as much as think of the tools we used in that case as more of a new kind of microscope, which allows you to look at things you could not have looked at before. We were talking about jobs earlier. I think what's interesting about AI is a technology. It actually has two properties that are interesting. One, which we often think about in the kinds of economics, which is this idea that it's a general purpose technology. A GPT, which allows you to do in all kinds of economic activities. But the other characteristic, which I find super interesting, is AI is an invention for making inventions. And so when people make the analogy of AI to the industrial revolution, I wish it's correct. People often think of it as a GPT, a general purpose technology. But I think if you're looking to these inventions as a method of invention characteristics, the Enlightenment may be a better analogy. Because what we did at the Enlightenment was discover new tools, instruments, telescopes, microscopes, the scientific method, things that allowed us to do science in a way that we could not have done before. Look at things we could not have looked at before. So I think AI has these two remarkable characteristics. And I think the impact on science is just, I'm so excited about that. But also it raises questions about, so what's the role of scientists? Are they, is there all there? It's a bit like the work question. Is it to formulate hypotheses? Is it to explore them? Is it to validate them? Is it to ask questions that I haven't been asked before? But I think the possibilities are super exciting. James said AI adoption is very wide, but still very shallow. In a typical job, people use AI for only about a fifth of their tasks, and mostly as a helper rather than something they actually work with. That's the gap. And the people who close it are the ones whose work becomes more valuable. That's why we put together a free 30-day plan to help you go from using AI like a search engine to actually working with it with a step-by-step guide on what to set up and in what order. It's available through my newsletter, Future Proof. So subscribe using the link in the description. And you'll get the 30-day plan for free. When do you think we're going to see a scientific breakthrough that's going to affect everybody's lives? Because for now, there are some amazing breakthroughs when we're treating very rare diseases and babies who were just born, right? But there hasn't been a massive effect on the population where we're like, oh my god, this is things to AI. I think we're starting to see that. Now, the question is that people think of it as, oh, thanks to AI. So I'll give an example. In healthcare, for example, in health, we're now able to do so many diagnostics that we couldn't have done before. Either faster, earlier, and see and pick up diseases or evidence of illnesses that we couldn't have done before. Can you give an example? Yeah, so we just published early this year. One, the largest studies ever done on breast cancer detection. We did this with the NHS in the UK. I think it involved over 180,000 women were part of the year-long study. And in that case, we're able to use these tools to do to pick up evidence of breast cancer that I would have been missed before and picked them up much earlier. And we've been. By looking at the mammography. Yeah, and using AI tools to observe those and evaluate those. We've done the same thing in the core rectal cancer, in lung cancer, and also in tuberculosis. I mean, tuberculosis is interesting because something like 30 to 40% of the people in the world, we have TB, go undiagnosed. And that's mostly because they live in communities where the infrastructure is just as in there to diagnose them. And so being able to use these tools in those places, we did a study that involved. We studied in Zambia and India. And we were able to show that, no, you can actually make accurate diagnosis for tuberculosis in those places. Now, in all of these examples, well, those people say, thanks to AI, they'll just think, oh, I got a great diagnosis and they discover that I have this disease. So a lot of those effects are going to be a little indirect. It's the same thing with another era where spending a lot of time on is trying to help communities be resilient. Imagine if people would not get surprised from natural events and natural disasters. So we've been working down a list from floods to monsoons and cyclones and weather and even earthquakes. One of my favorite examples, just because it's at scale now, is in floods. There's two kinds of floods you want to worry about, in most cases, which are riverine floods and flash floods. And in those cases, it's always been known that if you could make a prediction six plus days in advance, the impact on people would dramatically have. Just because people can move out of the way and so forth. And starting three years ago, we did a first experiment using our AI tools in India and it worked. And now we've expanded that to over 150 countries, in places where more than two billion people live. But again, in those cases, I think those people, those people experiences, they'll probably get an alert on their phone to get out of the way. There's a flood coming. - So not necessarily bridging it back to AI. - Yeah, but the connection back to AI may not be as direct. And I've given another example recently in the Venezuela earthquake. We've been experimented with the idea of using our, you know, Android systems with AI sensors, but also AI systems to make predictions. And in that case, I think this is reported in the New York Times as well. We're able to send alerts to about 12 million people ahead of time. In some cases, a few seconds, in some cases, a couple of minutes ahead of time. So we've been working our way down everything from wildfire predictions to monsoons, to cyclones, to hurricanes, where the next one of our models Thank you. in the in Melissa, in Jamaica, and we were able to send it. So, those are examples where the ability to make these predictions, are extraordinary. Yeah. And the impact of people, this is not in the far future, it's right now. What about the medical examples that you talked about? Are they available in the US? Well, like, we use this technology in the US. So, we're starting to show, for example, one of our tools called AMU, which is an assistive tool that we've tested with a variety of hospitals and clinics, Beth Deakness, Israel, at Harvard, and other places. We now involved in a nationwide trial. It's an assistive tool for doctors. Another AI-based tool developed called for detecting what's called diabetic retinopathy, where you can look at the eye as well to predict all kinds of diseases. We've actually applied that in more than 600,000 cases in India, and now we're working at the partner, the goal to scale that up to be able to do six million assessments. The Tobaculos' example that I mentioned, now that we've demonstrated that it works with a year-long study, we're now working the various providers in Africa and other places to actually start to deploy that and use that. So, a lot of these things are here and scaling. That's why these are such important examples, but often the attribution of those things to AI is often indirect or at least a step removed from the people who are experiencing those things. But there's still so much more to do, by the way. It don't get me wrong, but I think I worry that when we think about the AI as contributions to science and health and all these areas, it's always positioned as in the far future. Someday, we will do X, and I say, "No, no, no, no, no, no, we're starting to do that today." Of course, there's a lot more we will be able to do in the future, but don't, this isn't some future promise. I was trying to see those effects today. What is the wildest scientific moonshot you think is available now because of AI? Oh, what happened to the protein folding? I mean, that was a 70-year grant challenge that ideas first posed by, you know, Anfinson in his global lecture, like in the '70s, so in the '60s. So that's a moonshot that came to be real. What about future moonshots? Something you're excited about? One thing I'm excited about is the day we are able to build a prediction model that helps understand the dynamics in cells. Cells are very interesting. We don't quite yet understand. Nobody's ever has yet built a fully complete virtual model of the cell of how cells works, the perturbations that have been in the cell level. So that's still, and there's a lot of great work going on, by the way, so folks at the Broad Institute and a few other places are trying to build the cell atlas, and we're contributing to that. But once we're able to build models that allow us to make predictions about the dynamics in cells, that's going to be pretty exciting. So I'm excited about that. What's your timeline for that? Oh, I mean, that's the problem with breakthroughs. Scientific breakthroughs is hard to make predictions about time, because it's easy to make engineering predictions, but when it involves scientific breakthroughs, that's hard. I'm also pretty excited about, by the way, another area in science we haven't talked about, which I'm excited about, is the intersection of quantum physics and AI. So we have a team, the quantum AI team. The reason I mentioned that, so that's a moonshot, obviously we're trying to build the quantum computer, and I think we're making progress, and you'll be seeing, and you'll see in probably some big milestones we've had from the building of quantum computer point of view. But one of the things that's happening alongside our working quantum computing is we're now able to do extraordinary science alongside trying to build the quantum computer. So we've had some breakthrough results in physics in understanding the structure of black holes, in understanding electromagnetism. In fact, we've even demonstrated some new results in traversable wormhole possibilities. And also in our last year, we had a couple breakthroughs, we were able to do some work that relates to medical science, where we were able to understand the structure of some very hard to understand molecules and their dynamics, which is so important for some of the work we need to do in science. So I think you're going to see, even from that arena, breakthroughs that are relevant, not just for biomedical research, but also for other parts of science. So I'm pretty excited about the moon shots on the quantum AI side too. Yeah, it is really fascinating. I feel like it's happening gradually now, but then we're going to probably have a moment in history where it just speeds up immensely. I'm convinced of that. If you asked me, you were talking about bats earlier, I'm willing to bet that when we, you know, ten years from now, when we look back and say, what are some of the biggest things that came out of AI? I think this intersection of AI advancing science in ways that are hugely important for humanity and helps us to tackle questions in health, in well-being, in crises, and resilience. I think we'll probably be some of the biggest things that will have happened. Now, we also have to be super responsible because just as much as we're making progress on the science of it and the advances, it does raise some complexities we should pay attention to. So there'll be some risks related to bio and a lot of this work in biotech and in chemistry poses some risks. So the responsibility there will be super important as well so that we're doing this in a safe and responsible manner with regards to the science we're doing with AI. Yeah, which is another huge topic. Do you have a particular scientific discovery that if it's made in your lifetime, you'll see? Oh, wow. I've got quite a few. I mean, there's a whole bunch of mathematics. I mean, we've seen a lot of progress in mathematics these days. I mean, I think we forget to solve in the human hypothesis that'll be great. I'd be, you know, I'd love to see that. It's been one of the longest-tinding problems and challenges in mathematics. We've seen some of the things happening in mathematics. So that's exciting. But I actually think for me, the ones that I'm, I'd love to see in my lifetime have to do with the structure of biology and of life and our brains and our bodies. So these questions around understanding cells, understanding the dynamics of cells, really understand what's happening at the genomic level. Those, because I think if we make progress on those, it's going to lead to so many more discoveries. And one of the things I'm excited about even with the work on an awful fold is the fact that what that work has done is opened up the canvas to many more scientists. So today, awful fold, I think we reached a milestone a couple weeks ago. It's now being used by four million researchers around the world, because you made those predictions open and available. And these are researchers in 190 countries. And what's exciting about that is many of them are working on previous diniclector diseases that would have been too costly to even tackle. So the fact that we've expanded the canvas for people can our work on problems, I think that's exciting. So I think for me, part of it is giving tools to scientists to expand the wide range of possibilities to solve humanity's problems. Because I think so much of many of them have just been ignored. Absolutely. Because we have no time. We have no time. There are two costly people in various places around the world don't have the infrastructure to be able to do this work. But if we can make these predictions openly available so that more scientists get to work on these things, I think that's going to be exciting. Yeah, it's a community thing, right? Oh, I absolutely. I mean, I think it's also the reason we know where I think, you know, we have to do everything we can to make all of this, you know, available, accessible to everybody on earth. It's actually one of the reasons I like out. You know, Google's founding mission right organized was information making university accessible and useful. So the more people can use this and access all these benefits, we talked about all these things in health and, you know, we didn't even get to education in all these areas. One of the moonshots that we have, you're asking about moonshots, is that around languages and accessibility because it's one thing to do all these things in health and education and work and learning. But often you want to be able to do that in as many languages as possible. So we have a moonshot, our current moonshot on languages are allowed to get to 1000 languages. Last week we actually hit a milestone, which I'm very proud of, where we now are at 300 languages that we can handle, starting to make sure that's available in all our products. But 1000 would be great. In fact, if you wanted to get all the languages spoken on earth by everybody, you'd have to do something like 7,000 languages. Yeah, I was reading a research with some languages. If you ask something in that language and AI does a translation, sometimes it's up to 60% difference in meaning, which can be critical when it comes to a health question, for example. Oh, absolutely. And in fact, it's even more challenging than that. In many languages, we often think of singular languages. Often there's a lot of dialects in single languages, where there's showcases over 100 dialects of a single language. And so we're doing all this work, for example, in India, with the Indic languages to understand the various dialects and all the languages we're doing some work in Latin America and in Africa and many so even understanding that the dialects underneath the singular language is what meaning, how things actually mean. I think there's still a lot of work to do. It is. But that's a very moonshot. Yeah, absolutely. Can you talk to me about AGI? We're hearing so many, I think, Demis, Hassabis at Google IOS said we're at the foothills of AGI. What does it mean for a regular person? One of the things that's always challenged about AGI is this. People have so many definitions. So many definitions. In fact, you know, some level, if you'd gone back to, you know, the now famous dot with conference in 1956 where the term AGI. I first came into being, if you actually read the definite what the scientists would go together, the Dartmouth and 1950s were trying to do, it was, if you read that with a modern eye, so that sounds like a definition of AGI, even though they didn't name it that. And the term AGI itself, I think one of people was a big popularizer of that term is actually my colleague Shane Legg back in 2001 or two, where he first tried to talk about that and that was initially as a way to distinguish that between a more general kind of intelligence and set of capabilities compared to more narrow systems that we were building at the time where that's for natural language processing or there's that vision, machine vision. So those are an attempt to describe a general set of capabilities and at the time I think, so I think where he comes to AGI, I think of it as a spectrum. So there's a minimal definition you can think of where you say no, no, you're talking mostly about the kinds of things that Shane was talking about, or even the Dartmouth conference where you're describing the general cognitive capabilities that a typical human has. With all the peaks and valleys, I mean, humans have all kinds of peaks and valleys and stuff that I don't know how to do, but some humans maybe you can't, cannot be able to do. So there's a jaggedness to it. And I think if that's your definition, you could argue we're close to it if you look at it in hindsight, but there's some other higher definitions, the demonstrates, for example, likes a much more higher definition, which is a more scientific one, which is more complete in the way. I think on that basis we're much further away. So I think the AGI question usually when people ask me, I'll just start with what's your definition that I have to give you a prediction if I know what definition you're working with. But I think the reason it's important to think about it is, you know, regardless of how close we are to it, as we approach it, the effects of it are going to be quite profound, I think. I think we're going to have to think about, you know, what does that mean for, you know, everything from the risks and complexities associated with it to, boy, all the amazing things we can use it to do, science, discovery, all these extraordinary things we can use the technology to do, but also even deeper questions about what does it mean to be human in an age of AGI? What is it? I can do everything. I can do most things in no way, in everything. What do we, how do we think about meaning, I mean, this is these questions that we've been thinking a lot about. You may have seen recently that Demis Shane and I announced this setting up of the deep mind institute, but that's work we've been doing for the last few years to start to think about these deep questions about, so, and we'd love to just engage others and think together on how do we think about these questions? I think these are going to be so profound. And also the speed of it, right? If it happens in the next hundred years, then we have plenty of time to think. If it's the next five years where AI models advancing so fast, like how do we adapt? Yeah, I think the speed question is so critical, because I think right now we're starting to see in a whole range of techniques, kind of speed up this process of, of advancing these capable everything from recursive self-improvement and all these different things that, you know, we're now starting to be able to do that will probably speed things up. But I think there's still many breakthroughs that are going to be needed, many more advances in my view. That's still going to be important to get there. But I think we have to think about how do we do this responsibly? But safely, because, you know, we talked earlier about the risks to, you know, work and jobs and so forth. But AGI would probably pose other kinds of risks to the questions around alignment, questions around control. Those are all things we're going to need to think through and make sure we're being thoughtful about how we build this technology to make sure we're still, you know, we solve alignment, we address alignment. I think what I worry the most about Marina is the idea that we shouldn't let the advances in AGI capability and performance outrun our ability to control, build this technology safely. So I think that's the, we have to have those two things happen in tandem. But if the capability is outrun, our ability to build this technology and deploy it safely, I worry about that. Do you think we're at a stage where capabilities are outrunning? We should make sure that doesn't happen. I mean, we're trying very hard, at least on our teams, to think about, you know, how do we keep building our ability to build the technology safely, making sure that, for example, there's now a whole range of questions that are emerging around, you know, if you think about, you know, a lot of the reasoning capabilities of these systems, we do that through these channel thought and reasoning and traces. How do you make sure we don't, as a term now, people are using a lot of this idea of serial depth opacity? How do you make sure we can actually still see what the model is thinking all the way through? So we don't, there's some, there isn't some other part of it going on where the thing is going on. If we don't have a way to understand that or see that, that's what we all have to do. And I think that becomes super important when we start to have these agentic systems where we may set a benign goal. But then the sub strategies or sub goals that the agent then takes on to achieve that goal may be things that we may or may not want. And so unless we understand it all the way down as much as we can. So I think that's what I mean by we have to keep these two things in tandem. Yes, advanced the capabilities, they're useful, the science and everything else. But also make sure that doesn't outrun our ability to do the safe, responsible safety work and security work we need to do. Do you agree with some statements that people say we need to start slowing down now, or do you think we're not at the, well, you know, I always worry when people say that, just what do they mean? I think if people mean, let's not let these two things out, pace each other, then I agree with that. Slow down the development, the model development because it's getting so complex and we can't trace what the agent is thinking. Yeah, either slow down or speed up the safety and responsible work we're doing. I think to me, the key thing is to keep the two things should keep, should not outrun each other. That's the pacing, I think, that I think we need. And so that's the critical thing in my mind. I think we need to make sure that we are doing all the work we should be doing to make sure we build in these systems, responsible and safely. Is there an AI risk we're underestimating now? We're talking a lot about 10% chance of destroying humanity. That I've heard so many times in the past few weeks. We're talking about systems escaping. Is there anything else we're not talking about enough? We're trying to spend a lot of time thinking about this. I think one of the things we don't spend enough time on is just at least in the public dialogue is in how these systems are now interconnected to so many other systems. I think we're often we think about AI models in isolation. But keep in mind that we now connect in these systems, I mean, for useful good reasons, to grids and other things and other systems, I think, you know, this so much benefit from that, for example, we were able to manage grids more efficiently, manage electricity, manage more efficiently. But at the same time, that also raises some new questions about security and safety of infrastructure. So I think thinking about these systems, these AI capabilities as systems, I think is an important lens we should bring to this as opposed to thinking about, oh, if I evaluated my singular model, how's my model doing? I think we do a lot of that. But we don't often don't do the system wide evaluations and understanding of the effects. Also keep in mind that some of the intelligent capabilities, maybe the results, not of a single model, but a collection of models working together. So I think these more systemic kind of questions, I think are important both on the positive side. It doesn't therefore what that could mean. And think about us as humanity, I think we're an extraordinary species, mostly because of all of us, right, and all of complementary intelligences. I think it may for or be when we get to age, I won't be the result of somebody's singular model achieving that. It may be actually the collection of agents and models that achieve this. I think these more collective social components of intelligence, I think, is an important area to think of both on the opportunity side, but also on the challenges and complexity side. I remember a few years ago, a couple of years ago, I was co-chairing this body, the UN High-Level Body looking at AI and all the possibilities of AI. And I remember when we were having discussions about the risks of AI. So we went through the typical list that we all worry about, so misuse and misapplication of this technology. And I remember being struck by many of the members of the body from most different global south, say we should add another risk and we said what's that? I said misuse. I thought that was interesting to think about misuse as a risk, but I quickly got it, I think, because I think what they're referring to is there are many communities not just in the global south, but even in the US in developed economies where there's so many challenges already, access to healthcare. I mean, if you go on a county by a county assessment of the 3,000 whatever counties you have in the United States, there isn't equivalent levels of access to healthcare, to education, to all these things. And so if you're in some of these poor communities, and how do you make sure we also don't avoid the misuse possibilities where this could actually make a dramatic difference? Absolutely. When it comes to healthcare, so we have to discuss today. Yeah, to healthcare to schools, I mean, the kids who had promised somebody was going to come teach them STEM skills and all the rest of it had never happened. But this technology can actually be useful in some of those communities to address those. So I think that's another important kind of potential risk that's worth thinking about as well. And because we're here in New York during the UNGA, I was talking to some representatives from other countries and there was someone who posed this question. It made me think, I don't know if you have any ideas on that. But if we're talking about, okay, AI is amplifying our productivity and we talk tax large corporations so that people can participate and employees can gain benefit. We're only doing this in the US because the models have been developed here. So we're taxing the corporations in the US and then we're distributing wealth in inside the US. What happens to someone in Brazil, right? Who's using the same model? Have you ever thought about that? The distribution of the benefits that we're getting with AI across the globe? I think that's such an important question because in my mind, I think the possibilities of this technology benefit everybody, everywhere. So the question is, how do we increase the possibilities of somebody in Brazil being able to participate in this extraordinary moment? Is it through getting more educated, building the wrong capabilities and entrepreneurial capacity to build start companies that that's one way of participating? The other is also to think about, do they get better jobs to become more productive? So the question of, every should have access to participate? Often when I speak to people in various parts of the world, it's actually quite interesting that the attitudinally, there's a more positive view about AI around the world. But often the gaps and issues that most of them will cite are these questions of access or the infrastructure. How do we get to participate? How do we get the infrastructure? How do we get access to the tools to the models? So their concerns often have to do with just access and participation in this. Whereas attitudinally, they tend to be more positive because they see the possibilities in health, in education and entrepreneurship and all these things. I think expanding access is one of the clearest ways in my mind. And that's open models? Open models, languages. In fact, I'm always interested in, in fact, just the other day since you and I are both here. I was in a conversation where we talked about the infrastructure gaps in other parts of the world. People will point to, yes, access to models and compute, but they'll also point to electricity to connectivity. They said, we can have all the access to the models we want, but if you don't have access to broadband connectivity and electricity, you don't even get to start. And I was also struck by the fact that many of them also pointed to language as infrastructure. So many things we need to talk about, wow. So I think the more we can collectively invest, but the team up with others to address those things, I think that expands the possibilities for everyone everywhere. 100%. Okay, let's wrap up with you finishing your phrase. As intelligence gets cheaper, and if you agree with that or not, something gets more valuable. What gets more valuable? Human creativity, ingenuity, entrepreneurship, exploring possibilities, all those things become more valuable. And do you think intelligence gets cheaper or as humans, we're gonna just raise the bar for intelligence? I think both will be true. I think intelligence in as far as we've defined it so far. Remember those the time we used to define do mental arithmetic as intelligence? We got over that with calculators and computers. So I think our definition of intelligence is going to keep evolving, but I think if we're defining intelligence as we've historically defined it, I think that will get cheaper, but I think other forms of intelligence, other ways of knowing, I think are going to become more valuable and we're going to discover those things. We're going to discover those things. Absolutely. Thank you so much James. I could be chatting to you for another hour because it's so fascinating like the scope of things that you cover. Thank you so much for your work. Thank you for having me. It's such an exciting time, but I think we need to hold both things in mind that the possibilities are extraordinary, but the challenges and the risks and complexities are also as important to make sure we work on both things. Fascinating, turbulent time. What a time to be alive, right? Thank you. Thank you so much.