CEO of Microsoft AI: AI Will Transform Business and Work Like Never Before! - Mustafa Suleyman

Routine work is going to go away. That's going to free us up to be creative as entrepreneurs. I'm expecting people to declare that it's like them and their AI starting this new company. You don't have to learn the language of a computer to interact with it anymore. We're just about to embark on a completely new phase of the digital journey where computers are now learning to speak our language. It will increasingly be able to see what you see. It now lowers the barrier to entry to get access to support and information, and that is what is going to be the level up.

It is great to be here. Thanks so much for having me.

Yeah, it's a great question. I mean, I think that to be rational, you have to be both. And if you just sit on one side or the other, you're probably missing an important part of the truth. Because I think wisdom in the 21st century is being able to hold multiple confusing or contradictory ideas in working memory at the same time and navigate a path through those things, which doesn't leave you sort of characterizing those who disagree with you or somehow see things slightly differently in a negative or unconstructive way. And so it is true that this AI moment is going to deliver the greatest boost to productivity in the history of our species in the next couple of decades. That to me is unquestionable. I don't think that makes me an optimist. I think that makes me a good predictor of the underlying trends. At the same time, it is also going to create the most change in a disruptive, positive and negative version of disruptive way we've ever seen, right? And that is going to be incredibly destabilizing to the way that we currently understand the world to be, the way that we work, the way our politics operates, fundamentally what it even means to be human. So we're about to go through a revolution, and I think that, you know, in that world, I'm both optimistic and there are parts of me that do feel pessimistic.

Yeah, so in the past, the goal has been to invent things and have science make our lives calmer and happier and healthier. And the race has been how do we unlock these capabilities as fast as possible. We want to create and produce and invent and solve challenges, and that's what we did with steam and oil and electricity and wind and food systems. And we've seen this explosion of creativity certainly in the last couple centuries, but that's also been the history of our species, right? So the goal there has been proliferate, spread it as far and wide as possible so that everyone can enjoy the benefits immediately. Now, the thing that I've speculated about, which is the containment hypothesis, is that if we do that this time with AI technologies in a completely unfiltered way, then that has the potential to empower everybody to have a massive impact on everybody else in real time. So we aren't just talking about spreading information and knowledge, that's one part of it, but increasingly your AI agents or your copilots will be able to actually do things in the real world and in the digital world. And the cost of building one of those things is going to be zero marginal cost, right? So everyone's going to have access to them in 20 years, maybe even much earlier. Wow. And that's going to completely change how we get things done, how we interact with one another, and potentially cause a huge amount of chaos. And so a little bit of friction in the system could be our friend here, and that's really all that I'm positing in the containment hypothesis: containing things so that we can collectively as a society think carefully about the potential future consequences, the third-order effects. That seems like a rational thing to do at this point.

Yeah, good question. I mean, there are lots of potential ways that this technology gets misused, and in some respects, this concentrates power. It makes it easier for a small group of people to see what is happening in an entire ecosystem, understand the faces and the images and the patterns of massive crowds of people walking around in our societies. And that certainly can be used for a lot of good, right? I mean, we want to have a police force that is capable of instilling order and structure and so on. On the flip side, that also makes it easier for authoritarians, despots of any kind, to identify minority groups that they want to get rid of, right? And it'll just be much easier to find the needle in the haystack. And again, that has amazing benefits because you can catch the bad guy, but it also is potentially scary. And so each one of these steps forward has that balancing act between the harms and the benefits.

Well, one way of thinking about it is that intelligence is the thing that has made us productive and successful as a species. It's not just our muscle or our brawn, it's actually our minds and our ability to make predictions in complex environments, to solve hard problems. This is really the essence of what makes us special and creative. We learn to use tools and we invent things. So that intelligence, that technique of being able to predict what's going to happen next, that is actually something that we're increasingly learning to automate and essentially turn into an intelligence system that can be used by everybody. So what does that mean? That means, okay, well, we now have this very thing that made us smart and productive and created civilization, that very concept, intelligence, is going to be cheap and abundant, just like energy. Frankly, oil has turbocharged the creation of our species, and we now have 7 billion people on the planet largely as a result of the proliferation of oil, right? The next wave is that everybody is going to get access to personalized, real-time knowledge and a companion, an aid, a coach, a guide, a copilot that is going to help you get things done in practice. It will help you create and invent and solve problems and get things done. And so that is a massive force amplifier. So for everybody who wants to solve world hunger or invent new energy systems or solve the battery challenge so that we can really unlock the power of renewables, we now are going to have a super intelligent aid at your fingertips that's going to help you work through those problems.

Yeah, you're totally right. And in many ways, the iPhone's impact was just as hard for people to predict. Like if you said to people in 1985, like, what would the internet enable, very few people would have predicted mobile phones with real-time communication, video streaming, a camera and a microphone in your pocket, right, on the table almost in every room. That would have sounded terribly dystopian and scary. And yet, because of the friction, because it's taken us a couple decades to really figure it out and get it working in practice, we've created boundaries around these things, right? We've got security, we have encryption, there are privacy standards, there are real regulators that make meaningful interventions, and there's public pressure and there's a reaction, and it steadies the kind of arrival of these technologies in a really healthy and productive way. And it's kind of mind-blowing, you're right, like who would have thought what the iPhone would have done.

Thanks for the question. Human rights is core to what I believe to be the solution for a peaceful and stable society. I think like your success and my success is probably largely a function of the privilege that we have to live in a society not at war, have access to a great education system, be able to get access to health care, and really just get the chance to be who we want to be, instead of growing up in societies that are maybe in the middle of a refugee camp, going through war, in the middle of complete chaos, families that get moved on, you know, month after month or year after year and essentially end up being refugees. And I think that more than anything, we have to have so much more empathy and kindness for people who have been through that struggle. Like we tend to, there tends to be a sort of demonizing of refugees or migrants as though they're like coming to steal something from our stable worlds, but in fact they're actually fleeing insane hardship. So we have to be way more compassionate and forgiving and kind to those kinds of people and remember that those people are just like us. Like our societies could fall apart in the US, in the UK, and Europe just as their societies happen to be falling apart at this moment in time. And you know, we have to extend a hand of friendship and love. And the human rights framework taught me that because I grew up as a pretty strict Muslim and I realized, you know, sort of in my late teens and early 20s, that it was too narrow a kind of worldview. You know, it prioritized being Muslim over being human, and it didn't make sense to me that there wasn't gender equality. It didn't make sense to me that, you know, people who chose to get with a member of the same sex suddenly got demonized and they were like, you know, evil somehow. That none of that made sense to me, whereas a human rights framework respects everybody as equals. And I just can't see how that isn't the right way to live. So yeah, I became an atheist and secular and, you know, a big believer in these kind of rights frameworks for everyone.

Well, think about AI as a force amplifier. So the question is, which force is it amplifying, and which frame is it placing on the amplification? So what set of values, what ideas is it putting out into the world? What are its boundaries? What does it not do? And so technology is fundamentally an ethical question. It is clearly about how we reframe our culture and our ideas and our entertainment, our music, our knowledge in the world, right? So it was very obvious to me from the very beginning that we started DeepMind that we were going to have a huge moral responsibility to think about what it's going to be like to bring these agents, these copilots, into the world. What would their values be? What would they not do? Like, what are their limitations? And so that always was a big motivator for me and was a big part of the kind of structure that led to our acquisition by Google and subsequent efforts over my time at Google and since then have always been very focused on that question.

Yeah, great question. I mean, this is a term that gets banded around quite a lot in kind of nerd niche circles, and it's sort of broken out a little bit now. But I mean, basically, the idea of a general-purpose intelligence is that it's capable of learning in any environment, and over time it will end up becoming a super intelligence. It will get better by learning from itself and learning from any environment that you put it in, so much so that it kind of exceeds human performance at any knowledge or action task. So it could play any game, control any physical infrastructure, learn about any academic discipline. You know, it's going to be a really, really powerful system in the future. But I do think that's a long, long way away before we get there. There are just going to be regular AI systems or AI companions that can talk to you in the same language that you or I talk to one another. I'm sure lots of people have played with various chatbots like ChatGPT and Copilot, and you can see it's getting pretty good. Like, it's still a little off, but it's pretty knowledgeable. We've been reducing the hallucinations quite a lot lately. We've added the real-time access to information so it can check up on the news and it'll know the weather and it'll kind of have a temporal awareness. And those systems, I think, are going to be around for a pretty long time before we have to worry about a big AI. But you know, that is probably coming.

Good question. So not quite. Like, I think of a hallucination as the model inventing something new, right? And so sometimes you want it to be creative, like you're actually looking for it to find the connection between a zebra and a lemon and a New York taxi, right? And if you ask it that question, it's going to come up with some creative, poetic connection that links those three things. But then times you just want it to be like super to the point and not wander off and talk all kinds of creative, poetic nonsense. You just want it to give you the facts. Yeah. So that's a spectrum, and what we're trying to do is figure out from a query, like given the thing that you're asking the model, should we put it into a more creative mode or a more precise and specific mode that's more likely to be accurate? And that's kind of the hallucination challenge.

Well, narrow AI is the kind of conversational companions that I was describing. Like, it's limited, it doesn't improve on its own. Okay, so you sort of train it within some boundaries and then it gets good at a specific number of tasks. Like, it's good at great conversation, maybe it can do document understanding, maybe it can even generate a little bit of code, but you're kind of, we know what its capabilities are to some extent. Okay, that's a narrow version of AI when we use it for a specific purpose, right? A more general AI is going to be one where it has things like recursive self-improvement. It could edit its own code in order to get better. It could self-improve, or like it would have autonomy. It could act independently of your direct command essentially, or you give it a very general command, you know, and it goes off and does all sorts of sub-actions that are super complicated. Like, you know, maybe even invent a new product and create a website for it and then set up a dropship for it and then, you know, go and market it and take all the income and then do the accounts and so on. I mean, I think that's kind of plausible in say three to five years, before 2030. I think we'll definitely have that, and it might well be much, much sooner. Could well be like a lot sooner.

Yeah, for sure. I mean, we are going to have to contend with a new type of software, right? Historically, software has been input in, input out. You type something on Airbnb and it gives you a search result page, or you know, you play a piece of music on Spotify and that comes out as intended. And software so far has been about utility, it's been functional. The goal is for it to do the same thing over and over again in a very predictable way. And it's been really useful, right? We've created trillions of dollars of business value out of it and unbelievable social human connection and knowledge and all the rest of it. It's been incredible, actually. But we're just about to embark on a completely new phase of the digital journey where computers are now learning to speak our language, right? You don't have to learn the language of a computer to interact with it anymore. I mean, you can, it's still important to be a programmer, but it can actually understand your language. It can understand the audio that you give it when you do a voice input, the intonation, the inflection, the volume, the pace, the pauses, right? It will increasingly be able to see what you see. So not only will you take a picture and it will recognize what's in the picture, it will have complete screen understanding of everything that you're doing in your browser or on your desk or on your phone, frame by frame. You're browsing Instagram, it's seeing everything that you're seeing in real time, talking to you about the content of what you're interacting with on Instagram, on TikTok, you know, when you're reading the news, whatever you're doing. So that's a profound shift in what it... that's not a tool anymore, right? That is really starting to capture something meaningful about what it means to be human because it's using the same language that we use to understand one another, right? All those subtle cues that take place in social bonding of human relationships suddenly is going to be present in that dynamic. And it's a really big deal. And I think even though we've had a couple of years of large language models being out there and people get to play with it in the open source, I still don't think people are quite grasping how big a deal this shift is about to be.

You know, I think there are many amazing things about religion. Religion was a way that people made up stories, this is in my opinion, in order to make sense of a complex world that was confusing. And then science came along and showed that actually we really can understand the world through empirical observation and a falsifiable process of coming up with a hypothesis, running some experiments, observing those results, subjecting them to peer review, and then iterating on the corpus of human knowledge. And that's how we've produced known facts that are very reliable because they keep getting tested over and over again. And when something fundamental happens, we change our entire paradigm. So I think that, you know, it's unclear to me what role religion plays anymore in understanding the physical world or even increasingly understanding the kind of social world. I think that, you know, if you look at the contributions in the last, say, century or even couple centuries from great poets and philosophers and musicians, storytellers, creators, inventors, most of them have nothing to do with religion, and yet they've taught us most about ourselves and our societies, right? Yeah. If you really want to understand who we are as a people, as humans, how our societies function, you don't turn to religion anymore, unfortunately. Yeah. So I mean, some people will obviously disagree with that, and full respect to them for their opinion, but that's kind of where I land on it these days.

Well, we were at a time when I had sort of just left Google. I'd been working on LaMDA, which was Google's conversational search AI that we ended up never launching. It then got launched as Bard, and then it became Gemini. And I was sort of frustrated because I was like, these technologies are ready for prime time. They're ready for feedback from the real world. They're not perfect, they make a lot of mistakes, but we're technology creators, we've got to put things out there and see how it lands, iterate quickly, listen to what people have to say. And it was just so frustrating that we just couldn't get things done at Google at that time. And so I realized it has to be done, it's time to start a company. Me and Reid Hoffman and my friend Karén Simonyan started Inflection, and we created Pi, the personal intelligence. It was a super friendly, warm, high emotional intelligence conversational companion. At the time, it was, I think, the most fluent and the most kind of high EQ conversational companion. It had a bit of personality, it had a bit of humor. And you know, I think that it was an interesting time because we ended up getting a decent number of users. We had like a million downloads, about 6 million monthly active users. But the main thing I realized is that some people really love these experiences. The average session length for Pi was 33 minutes, 4.5 times a week. So that ranks it right up there, I was under TikTok, but you know, right up there. And there's not many experiences like that. Yeah. And so I think it gave me a real glimmer into like what's coming and how different the quality of the interaction is going to be if you really get the aesthetic and the UI and the tone of the personality right.

Yeah, I mean, so far you're right, ChatGPT is really a kind of work and productivity AI, and it's great, gives you access to knowledge and helps you rewrite things and so on. But in a way that sort of addresses a small part of our, or one important part of our human needs, right? I mean, the other part of our needs are to make sense of the world, to receive emotional support, to kind of have understanding of our social complexity, right? I mean, we want to come at the end of the day and vent. A big part of what I think people are doing on social media is posting stuff so that they can be heard. People want to feel like someone else is paying attention to them, that they want to feel like they're understood and that they're saying something that's meaningful, or maybe they just want to work through something. So in the new copilot that we've launched just a few weeks ago at Microsoft AI, we focused both on the IQ, so it's exceptionally accurate, minimizes those hallucinations we were talking about, has access to real-time information, really fast and fluent, but we've also focused on the EQ because we want it to be a kind companion. Yeah. We want it to be your hype man. We want it to back you up, right? We want it to be in your corner, on your team, looking out for you, there when you need it. I think people underrate how important that kind of social veges... yeah, you know, that is one of the things that gives middle-class kids a huge leg up, to always have a parent there, to always have a stable family with a sibling or, you know, even a best friend available to you whenever you need it. And I think that we're going to just touch on a little bit of those experiences now and make that available via copilot.

Yeah, I think you're going to say like, hey, copilot, or whatever you call your personal AI, I'm stuck, what's the answer to this? How should I navigate that? You know, I need to go buy this thing, can you take care of it for me? Will I be available in a week to do this thing? Or like, you know, you're going to basically use it as a way to kind of organize your life and spend less time, you know, sort of being distracted by administration, and more time pursuing your curiosities, right? Especially in voice mode, you know, I think you very quickly just forget that this is actually a piece of technology and it feels like you're just having a great conversation with a teacher that is patient, you know, non-judgmental, doesn't put you down, has infinite time available to you, and will wander off on the path that you choose to take through some complicated question. And it doesn't matter how many times you go back and say, can you explain that again, I didn't quite understand that, what do you mean, no problem, you just get to keep digging in in a completely personalized way. That is going to be the greatest leveling up we have ever seen because it's expensive, you know, socially it costs for me to turn around to one of my friends who knows a lot about something and pick up the phone and say, hey man, can you walk me through this thing? And obviously my friends will do it, but there's a barrier there, right? It's not instant. I'm really asking for something, or it's a cost for me to unload on my friend at the end of the day when I'm frustrated and irritated and I want to show up to my best friend in the best possible way and like have fun and be bright and energetic. And I'm still going to have those emotional moments with them, it's not that they're not going to be there. Yeah. It's just that it now lowers the barrier to entry to get access to support and knowledge and information, and that is what is going to be the level up.

Every new wave of technology leaves us with a cultural shift, right? It's not just static, right? It's going to have some impact. And it may be the case that social media has made us feel more lonely and isolated, and we have to unpack that. Like, why is that? What is that? People report feeling lonely, but another way of thinking about that, I think, is that people feel judged by social media. They feel excluded, they feel not good enough. Why is that? Because I think Instagram really dominated in highlighting a certain visual aesthetic. I need to be big and have muscles, right? I need to get my fashion on point. Oh my God, look at how good a cook she is, she's making incredible food. Look at how she's taking care and he's taking care of their kids, you know? Like, I'm just looking at all these perfect caricatures and it's making me feel insecure, right? So I think that's what's at the heart of it. And what sits beneath that is a UI, a UI that rewards a certain type of attention, right? And I think you can create new UIs, you can create new reward mechanisms, new incentivization to dampen that spirit and create more kind of breadth actually. And actually, I think, you know, the way TikTok is an evolution of that because you don't get as much of that on TikTok. It's kind of, you know, a lot of the comments even are like much more healthy. They're like full of jokes and support and, you know, kind of friendly banter. Whereas YouTube was just spiteful. I remember the YouTube comments back in the day seemed really rough, right? And obviously like X, I mean, I don't know who uses X anymore, but like that's turned into a cesspit. So I just think you have to be conscious and deliberate. I'm sure that when we put out new UI experiences, there are going to be some parts of society that get ruffled by it. And my job and my life's work is to be super attentive to those consequences and respond as fast as possible to trim the edges and reshape it and cast it. It's like a sculpture, and you have to just be paying full attention and taking responsibility for the real-time consequences of it.

So we launched Copilot about a year ago, very much as an experiment to see how people like to interact with conversational LLMs in the work setting. It's pretty incredible to see how Copilot is now embedded in Microsoft 365. So on Windows, on Word, on Excel, there's so many tools and features that enable you to just ask your Copilot whilst you're in the context of your document to summarize something, or create a table, or to create a schedule, or to compare two complex ideas. And so I think it's had a massive impact there, actually, and it sort of doesn't get talked about so much because it's been so kind of embedded and, you know, now it's become like second nature. It's like part of people's everyday workflow.

Well, the next wave for Copilots is these sort of flavor of much more personable, much more fluent, much more natural interactions. So it's fast, it's sleek, it's very elegant in the UI. We've done a lot of work to pare back the complexity. I think that people want calming, you know, cleaner interfaces. I feel like when I look at my computer sometimes, I see like colors of every type, shapes, different kinds of information architectures, and it's just like this blur. And I just need like serenity and simplicity. So we're really designing Copilot to go out and fetch the perfect nuggets of information for you and bring them into your kind of clean feed and really create a UI where you can focus on conversation. So the answers are designed to be pithy, short, you know, humorous, a little bit of like, you know, spice, a bit of energy, and, you know, it's fun to chat to as well as learn from at the same time. And so that kind of interactive back and forth was a big part of the motivation for how we designed it.