Behind The Tech with Kevin Scott - Mira Murati, Chief Technology Officer, OpenAI

Episode Date: July 11, 2023

Mira Murati is one of the most innovative tech leaders of our time. She has helped to scale OpenAI to the company it is today and lead teams to build innovative technologies and products including ChatGPT, DALL-E, and GPT-4. In this episode, Kevin and Mira discuss how she became interested in the sciences and technology, her innate sense of curiosity which led her to a career in artificial intelligence, and how OpenAI is deploying their technology responsibly for developers and consumers.   Mira Murati | OpenAI   Kevin Scott   Behind the Tech with Kevin Scott   Discover and follow other Microsoft podcasts at aka.ms/microsoft/podcasts.

. We're working on something that will change everything. Will change the way that we work, the way that we interact with each other, and the way that we think, and everything really, all aspects of life. Hi, everyone. Welcome to Behind the Tech. I'm your host, Kevin Scott,

Chief Technology Officer for Microsoft. In this podcast, we're going to to Behind the Tech. I'm your host, Kevin Scott, Chief Technology Officer for Microsoft. In this podcast, we're going to get behind the tech. We'll talk with some of the people who have made our modern tech world possible and understand what motivated them to create what they did. Join me to maybe learn a little bit about the history of computing and get a few behind the scenes insights into what's happening today. Stick around. Today we have a super exciting guest with us, Mira Marotti.

I've had the pleasure of working very close with Mira and her team at OpenAI for the last several years. Even though I've had all of these opportunities to interact with her, it was so interesting to hear more about her story, like how she grew up, how she first became interested in first mathematics and then physics and science and like where, like this intense curiosity that she had from childhood eventually led her. And I think there were just some amazing nuggets in our conversation. So just can't wait to dive right in.

So let's get at it. Mira Marotti is the CTO of OpenAI.. Mira Marotti is the CTO of OpenAI. She worked as an engineer and product manager, most notably helping to develop the Tesla Model X. She joined OpenAI in 2018 as the VP of Applied AI and Partnerships, and has since been promoted to CTO.

During that time, she's helped bring AI products like ChatGPT, DALI, and GPT-4 public, and has partnered closely with our team at Microsoft to integrate their technology into our products. It is so awesome to have you on the show today, Mira. Thank you so much for joining us. Thank you, Kevin. Excited to be here. I'm going to learn a lot about you today that I don't know, which I'm super stoked about. So I would love to understand how you got interested in science and technology in the first place.

You know, it started with math. When I was a kid, I just gravitated towards math and I would do problem sets all the time, and then eventually I did Olympiads. And I loved doing that. It was such a passion. And, you know, I grew up in Albania. It's a small country in Europe. And this was during the transition from totalitarian communism to this liberal capitalism.

And when I was two, the dictatorship regime fell, and it was anarchy overnight. But I think one thing that people misunderstand about these communist regimes is that when everything is equal, there is this really fierce competition for knowledge and education is everything. And so that's kind of the setting that I grew up in. And I was just always very hungry for knowledge and the pursuit of knowledge. But in a place where there is this constant regime change and everything is uncertain, I gravitated more towards the truth and science, something that felt steady

and you could get to the bottom of. And also the sources of history books or other books are sort of questionable. History kept changing. So I think maybe just intuitive and natural gravitation towards sciences and math was amplified by the circumstances in which I grew up in. And so from a very young age, I was super interested in math and physics and continued to pursue them until university. And were your parents mathematicians or scientists? No, not really.

They actually taught literature. And so it was just an organic interest towards math and science. I mean, coming from the West, one of the things that I... I'm a little bit older than or a lot older than you, I think. And one of the things that struck me growing up, where I also was interested in math and science and programming fairly early on, was that there was this competitive nature between the liberal democracies of the West and some of the Russian coalition that knowledge itself, particularly science and mathematics and technical knowledge, were one of these things that were highly valued both here and there at the time because it was a way to just sort of compete in whatever

contest it was that we were playing. I don't know whether it felt like that in Albania or not. Yeah, very much like that. I just love doing all these Olympiads, whether it was chemistry or biology or math. And when you're a kid, you don't really think about that. It was just a passion. But looking back, I can sort of see the circumstances. And also just keep in mind that there wasn't access to a lot of tools or entertainment. And so a lot of it was just out of boredom as well. Boredom actually, I think, is a very powerful motivator to go explore and really pursue frontiers of anything. And, you know, in the first years of my childhood, Albania was incredibly isolated, like North Korea is today.

And so there wasn't much inflow of, you know, entertainment or anything really besides books. So books were this entire universe. And back then, I just searched everything in books. Now we've got all these powerful tools in our fingertips and can do anything really. Yeah. I mean, what you just said that boredom is a very useful thing, I could not more strongly agree with. And I think it's really interesting that we seem to, as a society, have decided that boredom is bad and it is a thing to minimize. And it's one of the things that I struggle with my own children. I've got a 12 and a 14-year-old, and they don't have the same capacity to be bored as I did when I was a child.

I didn't grow up in Albania. Like, I'm sure like it's probably unfair to even make this comparison. But, you know, like I grew up in rural central Virginia. We had three television channels and, you know, like I was bored a lot. And like most of my life was, you know, in books. And it was, it was a very useful thing to, you know, I got focused very quickly on things that were substantive. Yeah, I think exactly that. Like exercising that ability to stay focused on something and, you know, reflect on information or distilling this information further.

And a lot of math is like that. You just need to sit with the problem forever and kind of exercises that muscle and faith that if you sit with it, you'll discover something. Yeah, for sure. I mean, I don't know about you, but I've even had hard math problems that I've worked on in the past where I was so obsessed with them that I would dream about them. And I sometimes would even wake up and I'm like, oh, finally, like I've got the proof for that interest that you had, which sounds like it was sort of innate or in air and cultural, just from the circumstances of how you grew up. But how did it get nurtured? Some of this stuff is hard. And so did you have mentors or teachers? Were the schools good?

How did you or maybe I should ask a different way. So at some point, whenever you are trying to do something substantive, like things get hard enough where you get stuck. So how did you get yourself unstuck? Yeah, for me, my teachers, when I was growing up, they were extremely supportive. And it was sort of unusual circumstances because I think today maybe less of that would be available. But back then, I don't know, maybe they saw something in me and they really wanted to help me pursue my interests. And, you know, often in class, I do completely different problem sets because I was bored

with the usual curriculum. And I'd still sit there with everyone, but they were very supportive of me doing something entirely different. And I was also lucky that my sister is a year and a half older than me. And so when I'd get bored with my stuff, I would go and look into her books. And then when I do her books, then I'd find other books. And my teachers were very helpful with that. I think that was probably the most helpful thing. I always knew there was something else.

There was more to pursue. There was more to learn. And then when I was 16, I was fortunate to get a scholarship to study abroad in Vancouver and Canada, where I did my last two years of high school. And there was that was a big opportunity to get outside of Albania and study in an international school with people from many different countries. That was a great opportunity for me. So where did computers enter the picture for you?

It was quite late, I would say. Maybe when I was a teenager in Albania and, you know, internet was slow, but I already thought about intelligence a lot. More through math and solving problems and just like how the theory of how the world works and trying to explain a lot of things through math or physics even. But I was always interested in how the brain works and intelligence more theoretically and at abstract levels. But I would say that the art of what I pursued was more in the theme of trying to apply my knowledge and try to apply technology to really hard problems that in some way makes our lives better. And when I was in college, I was studying engineering because I thought this was the

best way to apply my knowledge to actually solving real problems in the world. And when I was studying engineering, I was very interested in pursuing ways to bring sustainable transport to the world and also just sustainable energy in general. And so my senior project actually was building this hybrid race car. It was fun, but also we wanted to do something that felt really hard. And so instead of batteries, we used super capacitors and really trying to push what was possible. And obviously that was not something that you could build in production, but it was pushing science and seeing what's possible.

And that's why thereafter, I went to work at Tesla and I was really passionate about sustainable energy and sort of doing my part in bringing sustainable transport to the world. And that was a very exciting time about 10 years ago at Tesla. That's awesome. So what type of engineering did you study? Were you an electrical engineer, mechanical engineer? Something different? Yeah, I studied mechanical engineering. So a lot of hands-on stuff, software, but also hands-on. And what was your favorite thing about, I mean, because you're doing something very different now, like mechanical engineering is quite a bit

different than running a software engineering team. I love mechanical engineering. It's funny enough, I built my entire career on software engineering, but most of what I do in my free time is mechanical engineering and mechanical design. So I like what, what, what attracted you to that in the first place, other than, you know, the sort of the sustainable, like that, that it was a lever onto doing something in sustainable energy and like, how, how was that different than what you do now? Yeah, I think back then I probably saw it

more as a more tangible way to change things. And it just felt it didn't feel abstract. It felt very, you know, just very tangible. You make a change and you see it and you see how it affects reality. And I was always sort of a thinker. I would explore different things. It was hard.

Mechanical engineering is hard, but it's also very, very fulfilling. And there was always a software component. So like in a hybrid car, you've got the entire system and And it's not just the mechanical engineering part, there's always a software component, the electrical engineering component. So it's kind of a little bit of everything. And I always was attracted to sort of like complex systems. And, you know, when I was at Tesla, I got more and more interested in autopilot, in the promise of it, and also what we could do with AI and computer vision to completely change the way that we travel. And so that got me more and more interested in AI and what it could do in the

world, what sort of changes it could bring. I didn't necessarily want to become a car person. I always had this curiosity for different things. And I was very curious about how AI would affect the way that we interact with machines and how we interact with information in general. So at the time, I got really interested in spatial computing and just interacting with information and complex concepts in a completely different way that we interact even today, really, with a keyboard and the mouse, which is just so limited. And so I thought that AI and computer vision would help us really change this interface of interacting with information. And I imagined virtual reality or augmented reality where you can almost touch molecules or you can get a sense for chaos theory or gravitational waves.

And that is such intuitive understanding of concepts, complex concepts versus when you read it on a page. It's almost like as intuitive as grabbing a ball and, you know, getting a sense of projectile motion, even if you don't know the laws of physics. And so I thought, wow, this can really change the way that we learn and the way we absorb the world. That feels so true to me. I mean, I think one of the things that I really appreciate

about the modern world that we live in right now is you have things like YouTube, where if you are trying to understand a thing, there are so many people trying to explain that thing in so many different ways that if you are determined enough, you can find someone explaining the thing in exactly the right way for your particular brain to understand it quickly. And that was always my struggle. I could learn very quickly, but I, I don't think I learn exactly

the same way that other people learn. And if like, I can get the right conceptual hook on something that I've got it. And like, I can even understand like the things that like before I got the hook were too complicated. It's one of the things actually that it really excites me about what it is that you all are doing at OpenAI with these agents because the agent, if you are trying to get it to explain something to you, is infinitely patient and it's sort of adaptable. It will explain things to you in the way that you need it to explain things to if you're willing to have a conversation and tell it what it is that you need it to explain things to if you're willing to like have a conversation and tell it like what it is that you need. That feels very powerful to me. I completely agree. Yeah, it's one of the things that I'm most excited about

with this large language models and just generally deploying the AI systems that we're building in the real world. Yeah. So let's go back for a minute before we get on to all of the exciting AI stuff, which I'm sure is what everyone wants to hear us talk about. I want to hear a little bit about Tesla. So what was it like working there? And you had a pretty big responsibility there, like at the end where you were the head product manager for the Model X, which is like one of the most amazing innovative vehicles that anyone's ever created.

So for you not thinking of yourself as a car person, like you helped make one of the most disruptive cars that the world has seen, like maybe in the past, you know, 40, 50 years. So like, tell us a little bit about that. You know, Tesla was an incredible place. And in some ways, actually, I find it quite similar to OpenAI now where you have, obviously, it was much bigger and working on something very different but this high density of very talented smart people that are just

so passionate about what they're doing it's almost you know it's almost like a spiritual pursuit everyone believes so hard in what they were doing and that being the most important thing. And that is just so powerful when you're working on really hard problems. And in the case of Tesla, it's transforming an entire industry versus creating many new ones as well as transforming them. And, you know, it was incredibly hard, but also just invigorating and so fun. And I learned so much in a short amount of time. I don't think it's normal, you know, to build a car from zero to one you know in just three four years it's a very short time uh these things usually have this like very long

uh life cycle or timelines in in terms of uh design and prototyping and then production and so on. And, you know, one of the things that I learned at Tesla was that there is always some different way. Even if it seems impossible, there is always a way. There is always a different way. And, you know, in product in general general there's this kind of two ways of building product where you have the really really polished stuff and then this way of kind of hacking and iterating um and you know getting a lot of feedback from your user base and customers and just iterating quickly on that um and tesla i would say was in between kind of doing both. And that was

incredible. I mean, just the first time of operating like that in an industry that is so established. And so I learned a lot, especially from, you know, just the power of being creative and thinking originally, um, and just really changing everything, um, uh, and, and questioning what you know and questioning why, why things are done a certain way. Um, and I also just, that was the place where I started getting really interested into the power of AI and how we could change everything that we do. So in a sense, it was, you know, in my career, it was the place that really catalyzed my interest in working in AI. And then, of course, after working in VR and AR, I just thought, okay, intelligence is really the fundamental property of how the world is going to change.

And so then I got more and more interested on less the application side of it, but really understanding what general intelligence meant and how we could build it and how we make things go well for the world if we do build it. Yeah. So before we move on to AI, what's a, if you can share, an interesting technical problem or technical thing that you learned on the Model X, something that was tricky or interesting or different? So many things. So many things I could talk about the Falcon doors, but I feel like this could be, they could be problematic.

Maybe at a high level, we can sort of talk about that. So like that is, that is an interesting design choice to make. So obviously a brand new thing. And like, you know, as an engineer, like, I don't know the details of the implementation, but like, I can imagine where some designer has this idea that I want to do this thing. And then some engineer has to go decide or figure out how to make the thing work. Just in general, how do you balance those two things. Yeah, there are a lot of things about the Model X that felt just really pushing the envelope and just never been done before, especially in that kind of car. And so, you know, like the doors were a feature like that, or the HVAC system, the HEPA filter. And it always required kind of bringing together the whole team,

all the parts that would be working together. So design, engineering, manufacturing, the software side of the team, or maybe if it was relevant, the electrical engineers, and really bringing together all the pieces so you could design it together versus hand it off and then go back and forth or design something that could not be manufactured. So that was very powerful in working with teams that have different backgrounds, domain expertise, figuring out how to design something that has never been done before, adopting new ideas, but also very quickly kind of killing old ideas and moving on to the next one and just figuring out the right problem to work on at the right time. Yeah. I mean, I think that is an incredibly important thing. So this idea of like, you do your work and then throw it over the wall to

like the next person or team and the change that has to go do the next thing is, you know, there's a certain sort of efficiency that you can get from doing things that way. But if you're trying to make something brand new, it's very difficult to like have these sort of waterfall processes like that. I mean, there's so many jokes about, you know, like one of the things I was going to ask you about as a mechanical engineer is

like, Hey, did you spend any time in the machine shop? Because like there's this tension between mechanical engineers and machinists. It's like, Oh, you like gave me this print and like, there's no way to make it. Or, you know, there's the tension in software engineering between the product managers and the engineers. Like, you know, the product manager says we're going to go do this thing and the engineers are like, oh, are you

crazy? And so like it usually works better when everybody is like in the conversation. So it's super interesting to hear you say that's how you all did your work. Yeah, totally. And yeah, it's funny that you mentioned it because as a mechanical engineer, I was often machining my own parts just to understand sort of the constraint limitations

and also just the challenges of doing it. And it was very similar at Tesla where the design engineers were often on the floor fitting, testing the parts and just working very closely with manufacturing engineers. And I think that, like you said, it's key to innovating at scale past a certain size of the company. It's difficult to innovate if you're just throwing things over the wall. And, you know, like bureaucracy can kick in or processes. And, you know, as they grow, companies can lose their vision and sort of stop pursuing new ideas.

But, you know, if you kind of cut through that and minimize sort of the layers of processes and things or hoops that you have to jump through to get something done or bring some new idea, then I think it's much easier. So that was something actually quite critical looking back that I learned working at Tesla. Yeah, I was listening a long while ago of an interview that like Elon was doing where he was describing this thing

that was happening, not with the Model X, but like another one of the automobiles where they were having a really challenging time getting something manufactured. And as soon as he started like asking the right questions, it turned out that like the problem wasn't solving the problem of how to make this particular thing actually manufacturable. It was like, why did this thing exist at all? It was just completely unnecessary that it got designed the way that it got designed. And like the real fix wasn't to like go solve the nasty hard problem because the thing itself was a little bit arbitrary

and it's like changed the initial conditions and then the problem gets easier to solve. And so I think that's one of the things I admire a lot about Elon is like this first principles thinking of always like being able to sort of step back and ask the right questions about why are we doing a thing the way that we're doing it and what is necessary and what is not. Yeah, I think that's incredibly important, stepping back and not,

I mean, having the ability to be immersed in details and dig deep when you need to, but also stepping back and asking the right questions and having sort of this high degree of adaptability in the team for an intolerance for ambiguity. Because especially when people are extremely experienced, they have a certain way of doing things. And so you kind of need to be adaptable and also, you know, kind of believe and disbelieve things at the same time. And, and that's, those are hard, hard qualities

and traits to sort of sit together. Yeah. And then there's just something about big organizations. Like, you know, organizations should only be big if the nature of the problem that they're solving for their stakeholders requires you to be big. Because bigness is almost like a flavor of entropy that forces some of this stuff to happen where, just because of the complexity of the whole, no one has all of the details in their head. And so you can find yourself trapped in just feverishly working as hard as you can on the details of something. And if you could pull all the way back, you would just find that the thing that you're working so hard on is completely

unnecessary. And so it's one of the great things about the size that OpenAI is at right now is you sort of still institutionally, and the complexity of things, you have less of that weird entropy that happens to big organizations. And so the thing that I found is you just have to fight against it super hard because if you're not pushing back against this thing, you're just letting people entirely optimize for the narrow thing. Like, it just metastasizes into confusion, basically, and people optimizing for the wrong thing. Yeah, and certain momentum just carries on. Yeah.

So let's talk about AI. Let's sort of start with how did you make the transition from Tesla to OpenAI? Because you were in very early. From the beginning, it wasn't obvious at the start that... Not obvious at all that you were going to get to where you're at now. What made you take the leap? Yeah. So after I worked in VR and AR and was really intent on defining the new interface for spatial computing back then, it was a bit early, I think too early for VR and AR. But at that time, I actually got really interested in, you know,

how AI can help us just redefine the way that we interact with the world and we absorb information and the things that we produce and how it affects creativity. And so just, you know, this entire concept of amplifying our intelligence and what that means. And so I was really interested in learning more and seeing where this can go, this idea of pushing intelligence as a fundamental property that can have this very broad universal impact.

And at the time, I wasn't sure what the chances of that are to go all the way to artificial general intelligence, but I was just very interested in figuring out how far we could pursue it. And it really seemed like maybe the last thing that we'd ever work on. And it seemed like the most important thing that I could work on. And it was important to me to work on it at a place that cared about making sure that it goes well for the world. So I joined OpenAI when it was a nonprofit. And the mission of the company was then and still is to make sure that building AGI goes well for everyone in the world

and people can benefit from what it will bring. And obviously, since then, for practical reasons, we've evolved the structure of the company to have it be a limited partnership with a capped profit. So it still maintains the same mission and the nonprofit oversees the mission of the company. But, you know, I just sort of pursued my curiosity and what felt like the most important thing

to me at the time. Yeah. Which I, like, honestly, I think is super good career advice for anyone. Being able to make choices about what you do, where you believe the thing that you're working on is like the most important thing you can make a contribution to, I think is, you know, a thing people don't think deliberately enough about. I think it's so important because when you're working on really hard things,

it's that passion, that innate curiosity is the thing that can pull you through. Yeah, 100%. I mean, like, just really glad you said that, because I say this to people all the time. If you're working on a really hard problem with a bunch of really smart, highly motivated people, it's hard. Like most days, you're failing. So you go in and you're trying something and it doesn't work. And you're frustrated with yourself and you're frustrated with the people around you. And there are only a very small

number of things that you can have that will help you do that day after day after day until you've actually solved the problem and you get something that matters. And if you quit before you solve the problem and you get something that matters. And if you quit before you solve the problem, then you haven't solved the problem. Like, you've got nothing but this accumulated frustration that you've had. And I think one of the very few things that you can have that will get you through is you have to believe that it's the most important thing that you could be doing, you have to believe that it matters. Like money's not enough. Your mom wanting you to do it isn't enough. It looking good on your resume isn't enough. You have to just deeply, deeply believe that it's the most important thing you could be doing.

Yeah, exactly. And it's hard to find that faith and belief. And you almost have to experiment a what your mechanism is for dealing with that frustration of friction and failure. Because it's tough. I mean, I'm sure this is for everything that you've done, because you seem to have repeatedly chosen to do very hard things. Like I know for me, like I repeatedly choose to like do. I mean, it's almost like, you know, the most important thing is almost always like the hardest thing you could choose to do. And so, you know, just being able to like sustain that over time is. Yeah. And because at some point, too, you also like you probably had enough

success from your career at Tesla where you could have chosen just from a success perspective to not do the hardest thing. Well, I could go do something slightly easier than try to make an AGI like in a nonprofit. Right. That's possibly when you put it like that. Yes, yes. In fact, maybe another thing to talk about is like when you are one of the things that I think

has helped open A.I. be very successful is you have really excellent people, folks who are in their particular domain, whether it's figuring out how to ring numeric performance out of a GPU or if it's someone who understands how to do safety and alignment work or whether it's someone who understands how to architect a deep their game in each one of those areas. You also have this mission, like how do you go solve this incredibly complicated problem that not just open AI, but humanity has been thinking about for thousands of years, and how do you make that a reality, and how do you do it in a way where, you know, it, it creates massive benefits for humanity. Um, so, but you've got this third thing that's interesting, which is a way to keep

people focused on moving forward and progress. So like you could have the mission and you could have all of these smart people, but like they could be running in a thousand different directions and like their work could not be accruing to like a thing that's making progress. And I think that's sort of the extraordinary third thing that you all have been able to do. And I don't know, don't know whether you share that same perspective or, you know, like, you know, I'm just sort of curious on your take of what that missing element is. Because lots of labs out there with really smart people spending a lot of money and they've got an interesting intellectual mission and they still haven't been able to make the sort of progress that you all have made.

It's incredibly hard. Like you say, you can have this incredibly talented people and high density of them and they're innately curious and they're, you know, forever in pursuit of discovery and something new, but you need to, there needs to compound, you need to have all the smart people working together on kind of similar or same bets. And, you know, you want to motivate people. You don't hire smart people to tell them what to do, but you want them to be motivated and aligned enough to kind of work on similar or the same things.

And at Opening Eye, I think one of the most important things that we managed to do well was take a bet or take a couple of bets on the things that we believe the most and get alignment on those very early on. And even at the stage of recruiting people, actually, and bringing them in, that's most important and making sure they're really aligned on those things. And it's hard to say no, especially when there is so much opportunity. You could be working on all these different ideas. It's incredibly hard to say no. And so, and you doubt yourself and, you know, it might take a while for these bets to pan out, like the scaling laws and focusing on one large model, a ton of data, a ton of compute, which now it's obvious, but back then, not so much. And getting alignment on that is incredibly hard.

But I think it goes back to this idea of like figuring out how you work on the right problem at the right time and having faith on that. Yeah. I want to double click on this notion of it's hard to say no. It's incredibly hard to say no, because the thing that you're faced with is CTO of OpenAI, and I've had a lot of this over the past two decades, is you will have the smartest people in the world coming to you with very good ideas, ideas that you think are interesting and that you're a curious person and you're like,

oh, that's amazing. I love this. And then you know that that idea is not on the path that you're pursuing. And it might not be the next most important thing to go work on if you're choosing the next most important thing and just saying no. You're also a good person and the people that you work with are good people and you don't want to disappoint them and you don't want them to be sad. And so it's a real art form i think of like in in you know it's two parts it's like having the confidence and the courage to say no yourself uh when you also have your own uncertainties uh like am i wrong like am i making the right call uh and then being able to to deliver the no where it's not a no, it's sort

of a no, but it's no, but here's this other thing that I think if you do that, it will be even more interesting and create more impact. It's hard. Exactly. It is extremely hard. And together with that goes this building the muscle as an organization to sort of learn new things quickly or learn what's not going to work very quickly and adopt what's going to work very quickly and kill the old ideas quickly. It is hard to kill things that are already, you know, maybe working, but they're not working as well as something new that you could be doing. Yeah. Well, look, I think that's another thing that you all do really well. It's very, very, very important is choosing when to stop doing things. For instance, you all had an incredibly great demo a handful of years ago of a robotic hand that could single hand solve a Rubik's Cube.

And like it was, you know, a demo that was trying to get a reinforcement learning system to learn a robotic kinematic model. And so like it's technically interesting work. It's a super cool demo. reinforcement learning system to learn a robotic kinematic model. And so like, it's technically interesting work. It's a super cool demo, but like you all decided like, this isn't on the path. So like, we're going to stop working on this. And like, that's a hard decision because that was a lot of work for someone. It was like their favorite thing in the world. You know, it's like, yeah, people may quit like, because, you know, you stop doing this thing and that's the thing they wanted to work on.

So they're going to go find some other place to go work on it. But it's important, like really important. Yeah, exactly. It was, you know, at the time it was a very big bet that the company was making. And, you know, we had that and D DORA, and we had this inflection point that, okay, what are we trying to learn? How does this fit in on our path to AGI? And is there a better way? And so choosing to stop working on it, there's a better way. Yeah. You said a lot of like very important, profound things. Like you just said something like that, that I think is also very important.

It's like, what are we trying to learn? Like if more people ask that question deliberately, like we would, we would have a much better world and people would have more success. But that is, in essence, I think, one of the things that you all have always had pretty good focus on. It's like you're not doing activity for the sake of activity or doing activity for the sake of proving that you're smart. It's, we have a specific thing we're trying to learn through these things that we're doing. And, and it doesn't have to be AI. It could be product design or it could be, you know, like parenting or whatever. Like just trying, you know, like, what are you trying to learn, you know, through this thing that you're doing?

Let's talk a little bit about, I mean, you all have had a unbelievable, like total run, but like in particular, the past year or even the past six months have been, I think, shocking to a bunch of folks. I've been following what you all have been doing for a while. And so, you know, like, you know, what happened the past six months wasn't so, I mean, it was like surprising to me, but not quite as shocking to folks who, you know, sort of saw nothing, nothing, nothing. And then all of a sudden, like, chat GPT emerges and it like becomes

the most interesting thing in the world. Talk a little bit about that journey because I think Chad GPT is just one point on a long set of efforts that you all have been working on. It's not even the last thing. That's the other thing people probably aren't internalizing, that it is a point on a curve and more things are coming. So how have you all thought about that

in the context of how the public's reacting? The first time that we thought about deploying these models that were just in research territory was kind of this insane idea. It wasn't normal back then to go deploy a large language model in the real world. And what is the business case? What is it actually going to do for people? What problems is it going to solve? We we didn't really have those answers. But we thought, you know, if we make it accessible in such a way that it's easy to use and it is cheap to use, it is highly optimized. You don't need a lot of you don't need to know all the bells

and whistles of machine learning and just accessible, then maybe people's creativity would just bring to life, you know, new products and solutions, and we'd see how this technology could help us in the real world. And, of course, we had a hypothesis, but really it was just putting GPT-3 in the API the first time that we saw people interact with this large language models and the technology that we were building. And that for so many years we had just been building in the lab without, you know, this real world contact and feedback from people out there. So that was the first time. It was sort of this leap of faith that it was going to teach us something.

We were going to learn something from it and hopefully we could feed it back into the technology. We could bring back that knowledge, that feedback and figure out how to use it to make the technology better, more reliable, more aligned, safer, more robust when it eventually gets deployed in the real world. And, you know, I always believed that you can just build this powerful technology in the lab with no contact with reality and hope that somehow it's going to go well and that it's going to be safe and beneficial for all. And somehow you do need to figure out how to

bring society along, both in gathering that feedback and insight, but also in adjusting society to this change. And the best way to do that is for people to actually interact with the technology and see for themselves instead of telling them or just sharing scientific papers. So that was very important. And it took us then a couple of years to get to the point where we were not just releasing improvements to the model through the API, but in fact, the first interface that was more consumer facing that we played around with was DALL-E, DALL-E Labs, where people could just input a prompt in natural language, and then you'd see this beautiful, original, amazing images come up. And then really for research reasons, we were experimenting with this interface of dialogue

where you go back and forth with the model and charge GPT. And dialogue is such a powerful tool. The idea of Socratic dialogue and how people learn. You can sort of correct one another and or ask questions, get really to a deeper truth. And so we thought, if we put this out there, even with the existing models, we will learn a lot, we will get a lot of feedback and we can use this feedback to actually make our upcoming model that at the time was GPT-4 safer

and more aligned. So that was kind of the motivation. And of course, as we saw, you know, in just a few days, it became super popular and people just loved interacting with this AI system. One of the reasons why just me personally, I've been excited about the work that you all are doing is like this notion that you want to really allow a lot of people, a lot of non-expert people to be able to play around with the technology and to imagine how they can use it for things that they think are important is super important to me. And, you know, maybe a little bit of the same is true for you, but I grew up not in one of the coastal innovation centers where things like these AI systems get created. I, like you, did not have computer scientists or engineer parents. And the problems that people have in rural central Virginia, and I'm guessing the problems that people have in Albania, like, you know, some of them are common

across the board, but some of them are like very different. And like some of them, like you can't even imagine, like if your entire worldview is like, you know, I went to Stanford, I got a job at like one of the biggest technology companies in the world. And, you know, like I'm building this technology and like I, you know, I have to imagine all of its possible uses. Like you just can't even imagine what life is like for someone from Albania or rural Virginia. And so I think it's really unbelievably important to like have these things be platforms that aren't, you know, aren't just getting built in a lab where all the consequential decisions get made without any contact with the real world. And that doesn't mean that you have to, I mean, this is the last thing I want to, you know, chat about before we run out of time, but like, you know, it creates this very, very hard problem of like how you do responsible AI, because like you get this big benefit of lots of people participating, but then you get this big bucket of things that you

have to go solve at the same time to make sure that it's not creating a whole bunch of harm. So talk a little bit about how you all think about that. Yeah, that's well put. It's all about these trade-offs and minimizing. You can't have zero risk, but really minimizing those harms and actually really being able to respond quickly and iterate quickly on being able to maybe make changes to the models themselves or introduce tools or policies basically to contain those harms. And that's very difficult because often we're kind of doing all of this in the public eye. We don't have the privilege of doing it behind closed doors.

And so obviously with that comes a certain responsibility. But I think actually there is no other way to do it. I think it's the only way to get it right. It does need to be in the public eye and it needs to be in this continuous iterative cycle because the rate of technological advancement right now is insane. And so if you hold these systems back in the lab, the difference between, like, if we had never released GPT-3 or 3.5 and we had just gone out with GPT-4 on child GPT,

that would have shocked the world. It already did. And, you know, we had this continuous development cycle. And so I think that's really important. But one of the things is to really, from each deployment, from every time that we put out a model, we learn something. We learn something about maybe the safety of our systems

in the early development cycle or in post-training or in the product cycle. Safety is really deeply embedded and integrated at each stage of developing and deploying these models. And we're constantly changing what we're doing because we're just constantly learning new things. Like every week, I would say, we're learning something new. And so whether it's, you know, how you think about the data that you're selecting and filtering and analyzing the data early on or about the reinforcement learning with human feedback process that makes these models more aligned or classifiers that we use in production or the tools that we're making available for developers to have control and be able to be in the driver's seat and steer these models.

They're, you know, all these pieces along the life cycle of taking research to production. It's a complicated set of things to manage these trade-offs. But I agree with you. I don't know if there is any other reasonable alternative. And, you know, I think the trick is, like, having lots and lots of inputs that are coming into you, like where you can sort of hear the, what's working, what's not working, what is the scrutiny,

which of the problems that seem substantial or not, and which of the things that people are seeing in some weird permutation of how they're trying to use the product that you never imagined or intended creates. So it is, on the one hand of very exciting, but it's also like a huge responsibility, I think. It is. We're working on something that will change everything, will change the way that we work, the way that we interact with each other, and the way that we think and everything, really all aspects of life. Yeah. I have one last question for you that I ask everybody who's on the podcast.

So I know you probably have no free time given the intensity of the past really year. But I ask everyone what they do outside of work for fun. I love reading and I love going for hikes. Hiking is one of my favorite things to do, being in nature. Yeah. And like we live in a good place for hiking, which is good. Awesome. Well, thank you so much, Mira, for taking time out of an incredibly busy schedule to have this conversation. I've learned a ton and just enjoy this conversation,

enjoy being able to work with you on a regular basis. Awesome. I do too. Thanks so much, Kevin. . Wow. That was a fascinating conversation with Meera Maradi. As close partners, I get to work with Meera and her team all the time, helping to develop some of the big AI systems that they're building and then figuring out how to

safely deploy those unbelievably sophisticated AI systems into the products that we're building. But I learned a ton about Mira today that I didn't know before. I knew she was from Albania, but I had known relatively little about how she first got interested in science and technology in the first place. It was so great to hear about her, her teachers, her always being ahead and having

those teachers who were nurturing the curiosity that she had, her going through her sister's textbooks when she got bored with the stuff that she was working on. Her sister's, I think she said a year and a half older than she was. And then when she got bored with her sister's stuff, figuring out what else there was to learn. And I think you sort of heard at a bunch of places in our conversation like that, what am I going to go learn next? Why is that thing important to learn? And this belief that there is always something more to go learn is like one of the things i think that has driven

mira to such success and that the teams that she's responsible for leading to success uh and i think it's a good piece of career advice for all of us to be just very intentional about how we're thinking about you know the activity that we're doing right now is an opportunity to learn something that will help us get better and better at our jobs and to be more purposeful about how we invest more of our energy in something into the future. It was awesome to hear about her experience at Tesla, which I think has really shaped how she does her job as a leader and how she tackles these sort of complicated things where they are multidisciplinary, intersectional teams where you have to pull a lot of people together with a lot of

different points of view to do some of these super complicated things. And just hearing her talk about her passion for intelligence and what that means for how we're going to interface with complicated bits of technology and how they really have been thinking for a long while about how they take what they do and package it in a way where lots of people can use it. And where you really can unlock the imagination and the curiosity of a lot of other people. So like you're empowering them to use this technology to do the interesting things from their points of view. So anyway, it was just a sort of a fascinating conversation.

There were more tidbits in there. Like I found myself during the conversation remarking on several points where like she said something almost in passing that I thought were like real, super valuable nuggets of wisdom. So I hope everybody gets a chance to reflect on what this conversation really means. And that's all the time we have for today. Big thanks to Mira Mirati for joining us.

If you have anything you'd like to share with us, please email us anytime at behindthetech at microsoft.com. You can follow us on YouTube and on any of the usual places that you go get your podcast goodness. And until then, we'll see you next time.