CEO of Microsoft AI: The Next 10 Years Will Change Humanity Forever | Mustafa Suleyman

0:01I'm not giving up. I am selling the building. The final season of FX is the bear. The restaurant is flooded. Everything's either going to be okay. Or not. We are outgunned and we are outmanned. We have each other. FX is the bear. The final season. All episodes now streaming on Disney+. When you need to build up your team to handle the growing chaos at work, use Indeed Sponsored Jobs. It gives your job post the boost it needs to be seen and helps reach people with the right skills, certifications and more. Spend less time searching and more time actually interviewing candidates who check all your boxes.

0:46Listeners of this show will get a$75 sponsored job credit at Indeed.com slash podcast. That's Indeed.com slash podcast. Terms and conditions apply. Need a hiring hero? This is a job for Indeed Sponsored Jobs. Are we in an AI bubble? This is Mustafa Suleiman, co-founder of DeepMind, one of the first companies to teach machines how to think. Today, he leads Microsoft AI, one of the magnificent seven shaping the future of our world. It's the smartest, most capable technology we've ever invented. It's improving faster than anything we've ever seen. And it's very important that we keep reminding ourselves of that.

1:25Like in 25 years time, a large portion of the population are going to struggle to compete in the workplace with AI. With that power comes a question none of us can ignore. Because the same intelligence that builds our future could also rewrite what it means to be human. These things don't suffer. They don't feel pain. They're just simulating high quality conversation. So we've got to be very careful about that. AI isn't just learning, it's starting to live with us. Your personal AI is just going to have ambient awareness of what you're trying to do and talk to you in real time. It is going to remember everything in the future and be perfect if you choose for it to do that.

2:02So if it remembers everything, what's going to happen to our brains? They don't sleep, they don't forget, they don't feel, but they're learning faster than we ever imagined. So the real question is, what happens to us next? Mustafa, welcome. Thank you so much for being here. Thanks for having me. the first question everyone's talking about it on youtube right now are we in an ai bubble uh yes so no we're not i don't think so i don't think so i think um when you kind of think about it we are creating something that is truly magical um intelligence is the thing that has made us successful as a species.

2:49And we're now distilling that into a smaller and smaller unit that can be spread all over the world and that is going to be cheap and widely abundant. So that's just a remarkable thought. Yeah, but we kind of had the same thought, you know, during dot-com crash, like we're inventing internet, blah, blah, blah. But there are some companies that, you know, their valuations are much higher than their revenues. and like everyone's talking about nvidia investing and then that revenue going back to nvidia what do you think about that no i i think the value that we're going to produce in the next five to ten years is going to be unprecedented so you don't think we're going to experience something like you know 2008 covid i mean who knows but i think if you think just focus on the fundamental value that's being created this is the best prediction engine anyone's ever seen it's the smartest, most capable technology we've ever invented.

3:41It's improving faster than anything we've ever seen. It has got more easy to shape and control, not less. Three years ago, we thought that it was going to get more chaotic and more disorganized, so we weren't going to be able to sculpt it. And now we're producing beautiful, powerful, amazing experiences that are surprising us every month. So yeah, I know I'm genuinely very bullish. Okay. So you're not worried. You're good and we shouldn't be worried. That makes me feel better. Okay. Let's talk about AI becoming more conscious. You talked about it in one of the podcasts that we perceive AI more and more as human.

4:21How fast is it happening? What's going to happen in the next few years? It is getting more human-like. It's getting more accurate. It's getting more fluid, more smooth. And that is unprecedented. At the same time, it's definitely not conscious. It isn't aware of itself. It can't speak about itself. It's very different to what it's like for a human to have a subjective experience and to feel pain and suffering. And it's very important that we keep reminding ourselves of that because consciousness is the basis of our rights-based framework. And it is the thing that gives us responsibilities as citizens in societies that allows us to vote, which allows us to be subject to the law, which creates order.

5:06Like, you know, we can't begin to attribute this quality to a new species. That would be terrible for our species. But we saw stories of Google, the story that you told in your book of Google engineer who was convinced by AI that it was experiencing sadness and it was afraid of being shut down. True, true. Yeah. And that was an anthropomorphic projection. People, you know, just because it has some human qualities doesn't mean it has all of human qualities, including subjective experience. And I think we can prove that it doesn't have subjective experience, that it isn't going to emerge this kind of self-awareness.

5:43That's just a sci-fi fantasy. And I understand why people sometimes go there because we've been schooled on sci-fi for as long as forever. And it's a natural assumption to make, but these things don't suffer. They don't feel pain. They're just simulating high quality conversation. And that is magical and it feels amazing. But that doesn't mean that it is something, it feels something internally to be one of these models. So we've got to be very careful about that. Some people talk about our kids having real relationships with AIs in 10 years, like getting married to AIs. Do you think it's something that could potentially happen?

6:21And if yes, how can we prevent that? I think the history of the human condition proves to us that if it's possible, someone will probably do it. You know, people do crazy things, but the majority of people don't. The majority of people are pretty sound of mind and pretty sensible and want to live healthy, happy lives. And so if some people choose to marry their AI in 20 years time, I'm not going to kind of sit here and necessarily judge that or say don't do it or do do it. If you ask me my opinion, I don't think that these should be treated as entities that are of equal moral significance to humans.

6:57I think that they should be in service of humans. They should work for us. That's why we invent technology. You know, I love the technology behind this microphone because it amplifies our voices and distributes them to millions of people on the Internet. That is a technology that is in service of humanity that's doing good. That's what we want. And I don't necessarily think that an idea of superintelligence that is autonomous, that can self-improve, that can set its own goals, that can act independently of humans, that to me doesn't feel like a positive vision of the future. It would be very hard to contain something like that or align it to our values.

7:39And so that should be the anti-goal. That's not what we're trying to build. In my opinion, we're trying to build a humanist superintelligence, one that is aligned to our interests on our team in our corner, backing us up. That's the great project that we're engaged in at Microsoft with Copilot. But then again, a lot of people are talking about AGI. Is that similar to superintelligence? Yeah, you can think of AGI as a step before maybe superintelligence, but roughly speaking, they're used fairly interchangeably. Because Demis, your ex-co-founder mentioned we could achieve it by 2030. Can you talk about what it means for society?

8:17Yeah, I mean, I think achieving human level performance at most tasks, I wouldn't say all knowledge tasks, that feels quite likely in the next five years. I think it's a fairly grounded prediction. You know, if you think about it, these models already do many tasks better than a human. I mean, you know, summarization, translation, transcription, research, document writing, maybe even arguably poetry or some parts of literature or whatever. So you can see that they're taking steps towards being as good as a human at being a project manager or being a marketing person or an HR person or having a tough conversation with somebody about a diagnosis for a medical condition they have.

9:00So it is going to fundamentally change work in the most profound way. it's going to change the type of work that we do. You know, 10 years ago, it wasn't possible for you to do the job that you're now doing, but with streaming content, you don't need a big team around you. You don't need to be part of a big institution. So we've democratized access to the power of broadcast with the arrival of the internet and streaming and podcasting. That's amazing. Now with these models, we're democratizing access to intelligence. So you don't have to be a really privileged, wealthy, educated person that has access to capital, to be able to found a company, to hire a great team.

9:43You're going to have a team of intelligences around you that are the best lawyer, the best doctor, the best teacher, the best project manager. And that is going to create unbelievable amounts of competition because the distance between an idea and the realization of that idea is going to collapse. That's an amazing thought. People are just going to be thinking new companies into existence, new products, new pieces of poetry. That's amazing. You said people. People as in you and me. You're going to have an agent around you that is going to be able to execute on that idea. You know what my concern is with super intelligence?

10:23That AI is going to come up with ideas. Because why do we need to come up with ideas if AI has all the information. If we're talking about a marketing business, for example, it can identify gaps in the market, come up with a website, launch the ads, do the business. Do you think that might happen? It'll definitely happen. It'll definitely happen. You're going to have a team of agents around you that can do more and more of the tasks and arguably do them autonomously. But that comes back to what we were saying before about containment. Containment is the project of making sure that these AIs have limited scope and check in with you and are accountable to you and are working on your behalf.

11:03And so the creative challenge for technologists is to say, what are the limitations? What are the guardrails? What can it not do? When does it have to seek your approval for an autonomous action? And we've done this many times in past moments of technology, right? We invented the combustion engine almost a century ago, And that over time became cars and trucks and so on and so forth. But the amount of regulation that we put around those vehicles to make them work for us in service of humanity, it's unbelievable, right? Seatbelts and vehicle emissions and streetlights and freeways and speed limits and driver education.

11:44There's so many different components that go into that. And I think this is going to be no different. We're going to have to figure out what are all the regulations and guardrails that go into managing autonomous agents so that they always work for us. Let's imagine superintelligence. What does my day look like then? If AI is capable of so many things, what is this thing that is very specific to humans and is going to stay with humans? Yeah, I think that great technology helps you get bad technology out of the way. And that's the test. like it is kind of annoying to use a graphical user interface today on your desktop if you think about it if you just take a step back you go to your computer and you've got all these windows with different colors and different brands and different menus and information hierarchies and different settings and you have to learn each app for its own sake and we don't think of it as advertising, but when you open your computer, you kind of have a billboard of all these different competing apps that are like, use me, try me.

12:50It's kind of distracting. It's kind of ugly. I just want to be able to turn to my AI and be like, I need this task done. Solve it for me, buy it for me, book it for me, create it, check in with me at the right time, get my permission, you know, get my feedback, brainstorm with me, whatever it is. But you're going to start to see the operating system, the browser, the search engine, the apps, all of these are going to slip away as your personal humanist super intelligence takes over a lot of the day-to-day work of browsing, searching, planning, booking, thinking, et cetera. So it's no longer going to be a computer, you think?

13:28Yeah, I mean, less and less. I think we'll have different kinds of devices like tablets, wearables. It'll be much more ambient in five to 10 years. imagine it's 2040 and we're in this kitchen what things are absolute they don't exist anymore and what do i have to make my life easier okay so i think there will probably be some kind of household robotic arm probably i don't know if it'll be a humanoid walking around but certainly you would have a a mounted system on your on your uh on your counter here i think you'll yeah it will learn to use all the different appliances so probably you would have many of the similar appliances that you've got today but the ai is just going to learn to use them in the way that you would use them i think some people imagine that the entire thing is going to look totally different to suit the ai but actually it's not like the robot is just going to learn to do the things that you do in the way that you do it so you're still going to want to use a kettle i mean that kettle is you know hundreds of years old we've been you know making cups and handles like that in the same way so i think that stuff will remain the same so the hand or a humanoid what are you leaning towards?

14:36I think the human, I mean, 15 years is a pretty long way out. So it probably will be humanoid. But I think the challenge with humanoid is like the physical, the risk of it being physically in your space. I think it's got to be very, very accurate and very reliable and handling hot things around children and the elderly. And you can imagine all kinds of complications. So I think it will be technically feasible. It will just be a bit of an adjustment to people get comfortable with it what about devices do you think i'm going to have like an ai projector and like tap on the air as it and and get instructions on how to cook things or is it going to be glasses i think that you're going to just have it in your ear so i think that you know your ai is like your personal ai is just going to have ambient awareness of what you're trying to do and talk to you in real time so it will have a camera in my ear yeah and it was oh yeah interesting yeah i don't think anyone's working yeah i mean it's very hard to miniaturize that today so it's still a bit of a way off but by 2040 definitely and you'll have other wearables on you i mean look at how cool this mic is at the moment it's tiny super light yeah um so in time that'll have a camera on it as well and have ambient awareness okay so robotic arm wearables anything else and then i think those that ambient awareness is going to mean that the entire kitchen is just way more proactive proactively ordering the food and just checking in with you like oh i know that you've got a lot of guests coming this weekend should i order all the stuff for the barbecue yeah all of that stuff will just happen seamlessly in the background and the car will go and pick it pick stuff up yeah we'll definitely have self-driving cars by 2040 that basically do most of that yeah or maybe there'll be little bikes or different form factors because you know it won't necessarily need to look like a traditional car are there any markets where you think people should be paying attention to if they want to stay ahead in this ai race what are the most exciting markets i think by far the most exciting new market is medicine um i think at the moment the quality difference between the top 10 and the bottom 10 even in the United States, let alone rest of the world is unbelievable.

16:49The Gulf is probably in order of magnitude. It's huge. And again, that's going to completely collapse because everybody is going to have access to medical superintelligence and it will cost 20 bucks a month. It is going to be remarkably cheap. Talk to me about what you just released, Harvard-backed research, right? Within co-pilot if I ask a medical question, how is it going to work? Exactly. So now, I mean, we find that like about 40 % of our queries each week are health related. 40 %? Yeah, four zeros, pretty remarkable. Like millions of people a day are asking health related queries. So we decided to really focus on improving the quality of the health answers.

17:32And part of doing that is that we ground the answers in citations from Harvard Medical, which is an amazing kind of gift to the world, like the most respected health institution. And we also do the NHS in the UK and we'll be doing other health systems over time. And so people now get really high quality, very reliable, you know, pretty accurate. Like sometimes it makes mistakes. It's not, you can't rely on it yet. You always have to go see a doctor, et cetera. But if you think about it, most of the time in healthcare situation, people are trying to make sense of a complicated technical set of language that they don't really understand.

18:08They're probably anxious or worried. So they want to repeat things lots of times. They want to get it explained in a very simple way. And sometimes they want help actually making a decision about who to go see. So we also now have recommendations for physicians. You could go see a doctor, you could see a dermatologist, you could see a physio, and we will find someone that fits your preferences. Like, you know, maybe, you know, you want to see a woman physio who specializes in sports massage because you're a cyclist and you want them to be local to your area. And so we can make that recommendation.

18:43Copilot does that directly inside of the app in chat. That is awesome. Yeah. It's so cool to see that taking off. I love that. I need to ask about my cholesterol levels. We're going to do this today. Like you mentioned that access to knowledge is now democratized. What's going to happen to traditional education? Will we see the standard like bachelor's, masters where we basically spend five or six years just acquiring knowledge? Yeah, I don't think so. I think that knowledge acquisition is going to be a conversation between you and, you know, co-pilot. So for example, another feature that we just launched today, in fact, is this learn live feature.

19:20So you actually have a tutor shows up on the screen and it will lay out a quiz for you on any topic that you like. So it doesn't have to be a school curriculum topic, you might be learning about cacti or Persian rugs. It will give you an education on it, lay out the kind of curriculum piece by piece, give you a nice quiz, present it in a nice graphical user interface. And so knowledge acquisition is about to get completely decentralized and available to everybody. And you just have an expert teacher in your pocket on any subject. Does this mean all education businesses are dead? I'm thinking about my apps.

19:58I think it's the opposite. I think education businesses that adopt these kinds of AI tools and integrate them as quickly as possible, that they're really going to flourish. It's happening like that for everything. I mean, now in Copilot, you can also generate a podcast literally in a minute. It's a great introductory thing. Like the other day I generate, I don't really follow football or like soccer, but I do support Arsenal because I'm from London. And when I go back to London to see my friends, I want to know what I'm talking about a little bit. So I asked Copilot to generate me a podcast on how Arsenal is doing in the season and what the latest is with the transfers and who are the best players and stuff like that.

20:36So just fun, simple bits of information are perfect. And now I'm in the gym listening to a little five, seven minute podcast about that. Going back to knowledge and education, should I be saving from my kids' college or not? They're five and four um so so 15 years time yeah um probably not there you go i don't know it's really hard to judge i'm not sure that we will value a stamford education um that costs i don't know what does it cost 50 grand a year or something Is it more? Something crazy. 50 is a private school here. Right. I mean, I just, I don't, I think that we're going to have world-class expertise on tap that costs 20 bucks a month.

21:28But what if like, for example, right now I'm educated and like I studied maths, economics, so I can have these conversations. If I completely abandon acquisition of knowledge, how can I have an argument? It's not like I'm always talking to my phone, right? Right. Well, you could, you could still argue about stuff. If anything, it'll probably make us all more argumentative because we'll be even more knowledgeable. We wouldn't be exchanging ideas with people, right? Yeah. I mean, but since you base your arguments on things that you've acquired when learning something. Yeah. But you, you, that's kind of my point is that the knowledge acquisition in the traditional form in the classroom, you know, with the textbook, that's going to shift entirely.

22:05So the classroom is going to look more like practicing using your knowledge. So I'd acquire more knowledge if I do it the other way around with AI. Exactly. Okay. Because then as a kid, you're going to be able to talk to your AI tutor at any time during the day. You're going to learn from the AI podcast. You're going to learn from the AI video that is going to be generated. So all the kind of straightforward teaching is going to take place, you know, with the student and the AI. And then when you come into the classroom for your day, the students are going to be talking to each other about the knowledge that they've acquired from their AI and debating and learning to, you know, be more empathetic and be better listeners and adjust their tone and, you know, that kind of thing.

22:49So what do you think parents should be teaching their kids now? I think it's still important to be good at like learning knowledge from first principles yourself and not depending on the sort of AI tutor leading you through. I think one of the most important things from school is the discipline of being able to teach yourself. That's a meta skill. And that comes with friction. So I think as a parent, you still have to introduce discipline and friction into the process. Because if it comes all too slowly and it's always on tap, then there's a risk that the child could just get used to having everything instantly available and doesn't learn from the hard work, you know, the benefits of hard work, which I think are important.

23:36So that's like something to think through now that everything is going to be so seamless. Another question I wanted to ask you, for everyone who thinks that they need to be technical in the age of AI, you have philosophical background and you're head of AI at Microsoft. Yeah, I mean, I think that it's easier than ever to understand what's being built. Like everything is available on YouTube. You know, there are so many courses, the curriculums for all of the major CS classes, machine learning classes are available for undergraduate, master's level even. Is that how you learn? Yeah. Like I've watched, I watched Andrew Ng's course back in the day, which I encourage everyone to look at.

24:15And there's just so many classes. I watched maybe 10 years ago, I think when he was at Coursera and it first came out. Wow. And I've just watched so many of those. So it's very accessible. And I think if you just have a logical mind, you're patient, you're disciplined with your own learning, it is actually very very easy that's great so anyone well basically anyone who learns can become head of ai build an ai company without learning how to code and yeah because the skills now i mean it's still good to be able to learn how to code but the primary skill is in synthesis like the goal is to bring together all the different disciplines of ux and research and product and the aesthetic and also large-scale distributed systems for training big models doing the hardware like you know so i think that people who are multi-disciplinary and can straddle loads of different technical disciplines are going to get rewarded most yeah well this is so inspiring you've released the memory feature right so uh ai is going to remember everything about us i don't know how it's gonna i haven't tried the copilot yet but with other ais i see how it forgets like i just tell it and then i ask the question and for some reason forgets and i have to remind it Is it going to be a problem?

25:22Have you tried it? It doesn't have infinite perfect memory, but it does remember specific durable facts. So it will remember that you have a preference for, I don't know, pineapple on pizza and that you like going running on a Sunday morning or whatever. Or like my measurements. That's the thing that my eyes keep forgetting. Oh, really? Can you choose the size from this website? Oh, I don't have your measurements. I'm like, no, yes, you do. Oh, interesting. And we'll be like, oh, exactly. I do have them. I experienced this forgetting problem almost everywhere. Yeah. I mean, we're working on a good knowledge representation as well.

25:54So you can go and edit the AI's memory of what it knows about you. And if you ask Copilot today, what do you know about me? It will produce a pretty long list. It's not entirely definitive. So it won't be absolutely everything that you've ever told it. It is selective and it chooses what to remember. And you can tell it to remember certain things like please remember my sizing and so on. Yeah. But it is going to remember everything in the future and be perfect if you choose for it to do that. And of course you can delete it and control it and edit it as you wish. So if it remembers everything, what's going to happen to our brains?

26:24I think people said that about calculators, right? Like, you know, people aren't going to have to memorize their times tables anymore. They're just going to use the calculator. But you know what? When I was preparing for GMAT, the test that you have to take for an MBA, they don't allow you to use calculators. And there are a lot of numbers. I remember that two weeks after the exam, my brain was so fast because I trained it to calculate. And then it gradually slowed down and I lost this again. Do you think that that transferred into other areas of your knowledge because you're good at memorizing the numbers?

26:55No, I think it was just related to numbers. It's sort of the risk, isn't it? It's like, you know, I used to be able to memorize a lot of telephone numbers as a kid because I... Because you had to. Yeah, I didn't get a cell phone until I was like 15. I think it was one of the first people to get cell phones in my year at school. And before that, everyone just remembered telephone numbers. And now I can barely remember my own phone number from, you know, my own house. Yeah. Yeah. Do you think it affects our brains or they just develop in a different way? I don't think so because my brain is now stimulated in a lot of other directions.

27:26It's a muscle that you don't exercise, but in turn, you focus on other aspects of your brain. And so I think that we're way better at synthesizing vast amounts of novel information compared to people in the 50s or our ancestors 200 years ago. We're completely bombarded with stimulus. Right. And so I think that's interesting. It's had an interesting effect in some respects has made us a bit more polarized in other respects, it's made us much more empathetic. Like we don't fear, you know, people of color or people of different sexualities or women. And a lot of that has got to do with just being aware of others and spending time with them and watching them on social media and behave in different ways.

28:12And that has driven a lot more empathy. actually. People often focus on how it's driven a lot more disagreement, and it has, but it's also driven so much more understanding and respect and forgiveness and those kinds of things. In general, do you think we're getting dumber with AI or not? No, I don't think we're getting dumber. I think that it reduces the barrier to accessing information and actually makes us a lot smarter. You can ask any question now, and many people are. So questions that were previously going unasked because the cost of going to a library and pulling a book off the shelf and talking to the librarian was too hard.

28:49It took like four hours to get an answer to some question. Like now you can go to the web and then after the web, you can ask a perfect question in a conversational way to a chatbot. So I think that's driving more collective intelligence and more awareness and understanding of our world. Can you talk to me about some killer features that you think people are underutilizing right now with AI? Like you implement something and then you save like an hour a day. So for me, it was asking AI to add things to my calendar, provide my emails, search through my emails. That has saved me a lot of time. Yeah, I mean, I think one of the super exciting features is this thing we're calling connectors.