Microsoft’s AI MIGHT Be the Most Revolutionary Thing

‍

Soundar Srinivasan is the Director of the Microsoft’s AI Development and Acceleration Program in the Office of Microsoft's CTO.‍

The program’s mission is to increase the AI capabilities of Product Groups and develop the next generation of AI leaders at Microsoft. The team works on a variety of AI topics, with a recent focus on Generative AI (including LLMs), and AIOPs. Soundar drew from his experience as the Director of AI Product Challenges at Robert Bosch while he built out the team, including establishing a strong culture of Learning, Diversity, and Inclusion.

Soundar obtained a PhD in noise robust speech recognition from the Ohio State University in 2006. Before joining Microsoft, he worked for Robert Bosch LLC in various aspects of machine learning and AI, including leading a central team of experts that provided AI as a service to internal business unit partners and incubating new AI products. On the personal front, he is an avid sports fan and hiker and enjoys the offerings in the New England area with his family.

‍

Host:
Today on the RH podcast on AI, we have Sundar Srinivasan. He's the director of Microsoft's AI development and acceleration group. I ask myself, is this what I want to do for the next several years? I explore opportunities. If you like the podcast in general, just subscribe, and I'll see you at the end of the video.

I know that you're a director at Microsoft for AI development and on the acceleration program, and it's quite a unique opportunity. I would just like you to describe what it's like for you to contribute on a daily basis and what it's like in your department. What are you guys trying to achieve at Microsoft?

Sundar Srinivasan:
Yeah, so Microsoft has been investing in AI for a really long time. Our program started roughly about six years ago. The goal was primarily twofold: one is a recognition that Microsoft needs a lot more AI talent, a lot more expertise, but there was also recognition that AI is going to disrupt not just the tech industry but all industries. So, there is a lot of competition, but there’s not a lot of expertise. You’d expect systems to have set up to graduate, but being a relatively new field, there’s not a lot of expertise in the industry either.

In assessing the competitive landscape, we wanted to create something differentiating to attract people to come to Microsoft. That was one of the motivations, and this gives them an advantage point to AI opportunities, but also team culture, location, product interest, tech stack, what you will, at Microsoft. It allows them to go after their passion.

So, a lot of what we do is on the product side, innovating for new capabilities. We don't typically take on day-to-day innovation, and we also don't do long-term innovation; that is the scope of research. So, we're in between, where we take what we think is possible, but it's not clear how or if at all in the near term, meaning like 6 to 24 months, and we find a way to accelerate that innovation.

So, it's a combination of those things. Basically, that's pretty much what my team does. On one hand, we help people through this experiential learning framework, get very deeply acquainted with the practical aspects of developing these world-class AI systems. On the other hand, we help product teams break the latest research into engineering and product definitions for these products.

Host:
It sounds like a mix of applied AI and almost like an education platform for different departments.

Sundar Srinivasan:
Yeah, I mean, I wouldn't call it a lab because sometimes the lab has a notion of being separate and doing its own thing. We don't sit and think of stuff to do on our own. We basically look at the challenges that product teams want to collaborate with us on, and we look at the next generation of challenges that they have. Then we find where we can add value by accelerating our expertise.

We don’t just have scientists on the team; we have engineers in equal proportions, and we have a significant amount of product management. So, basically, what we’re looking at is new opportunities and seeing whether we have expertise to help the product team. It's very applied.

The educational aspect is less about teaching in the classical sense. It's more about accelerating your career. By virtue of working on these hard problems, you have to push the envelope. There’s technical learning that happens, but also these are the most important problems for the company. So, you get exposure to the leadership of the organization, you learn how to communicate to executives, you learn how to work in an interdisciplinary setting, and you learn the practical aspects of launching a product that goes to millions of users.

So, there’s a lot of learning from that perspective.

Host:
So, okay, so let’s if we slice that up a little bit um well you know I basically want to figure out how you make that possible in a way right because there's so many different aspects to it um from it's almost like you know it's a pipeline it sounds like you can tell me if it sounds like a pipeline from research to pract to practicality but not just from a technical point of view but also from this is these are these are all the ways you're going to have to interact with uh other parts of the organization that makes the makes whatever you do make possible or implementable or practical so there's uh there's a you know sort of like a wide breath of skill sets that at the end of this that you're going to have to ingest whatever the the cycle is um so that that's a complex thing to structure um you know and it you know it's almost like a mix of MBA and a technical degree so how you know again it's less pic but uh yeah no fair enough more applied so more you know and and I would a shorter time frame so I'd like to like slice up how do you make that possible so what's what's the first thing you would you would say you think of when you're trying to get a person from uh from zero to to your goal so what's you know what's the underlying uh principle or acction

Soundar Srinivasan: 

Yeah so it all starts with identifying where the opportunities in the organization are uh at any given point in time we can tackle roughly about 8 to 10 opportunities um so it starts with what those opportunities look like what type of uh technical skills they map to or what type of innovation needs the map to and a lot of the things that we get uh as a support we unfortunately have to say no to um it has to be a mix of things that are highly important for the organization things that uh we can uh we can contribute but also because we like mentioned we on a outside P sort of like this solution of uh lab right so we don’t have deep domain expertise into the products that's that's why we collaborate with the product games um so we also have to be cognizant of how much onboarding we need to do into the toades how much uh tomade expertise we need so it's sort of finding the sweet spot first and then um um I feel like the uniqueness of the program uh that we have is in exactly the ability to onboard people and get them to deliver and get them to collaborate in an interesting manner we have uh not only Engineers on the team we also have uh AES and barers on the team they um do the um heavy lifting of scoping uh the work um providing ongoing feedback clarifying uh planning the different options so there are uh it it runs like a typical aend development process so from that perspective it's not different except that this um it is innovation so it is not regular product development so there is uh we have C for like failures and H for exploration uh uh overhead um 70% of the work that we do uh tend to have be very successful that's um that's way more than what we understood so when we started we thought okay if we get like a 30% success that would be great um but uh I think it goes to careful uh wetting that we have opportunities and over time we um we have built this muscle out like we know how to honor people we know how to collaborate uh we actually because we uh started out in um New England area in Microsoft so um product cent man is and the West Coast so we actually started uh doing the hybrid uh and jamort collaborations day before um Everybody forced to do in the co um so um lot done muscle we going the last few years all Bones on set

Host: 

So it seems like uh you know you're maximally flexible you've been maximally flexible on almost all um on on all axis for um the way that you build things and and how you integrate it to even the way you work so that's a you're experimenting on on all kind you've been experimenting on all different layers in all different ways

Soundar Srinivasan: Yeah that was actually one of the things that excited me about taking on this role it was an opportunity to build this ground up uh and you're going to get those opportunities in the last couple days so I was super excited uh but exactly if like we had opportunity to experiment on all these dimensions and you out like what works and uh what um you know what are things
We need to continue to get better at what do you consider your vetting process for innovation? This is one question I wanted to ask you because you talked about making sure that you get the right things that make sense for you. So what's, what do you consider that? You know, you said that that accounts for most of your success, making the right choices, and I think that in life, that's generally true. So how do you make that happen? How do you, uh, how do you think about making choices for innovative solutions? Because a lot of people are thinking about how to do that for their lives when they want to make something, you know, how viable is it?

Soundar Srinivasan:
Yeah, so I think most of research or innovation comes down to three exes: like, why this problem, why now, and why us? So if you have good answers to all those three, then I feel like, you know, that's the best we can do. It's still, uh, public risk. If there's no risk, then, you know, I would say it's a regular development. It's not, uh, innovation. It's not research or at least like breakthrough, uh, type of development. So we start with, uh, why now, and, uh, some of it is, uh, there are capabilities, for example, generative AI is now enabling, which didn't exist before. So that could be part of it. But also, maybe the market is a lot more ripe for the opportunities for innovation now than before. Uh, that sort of goes into the market-fit side of things. Um, and, uh, why this particular technical problem? So why can't the problem be solved in another manner? Even dealing with data utilization, like there is now with the generative cycle, um, every problem looks like a generative item. Of course, that, uh, that's not, uh, necessary or useful. Um, so we try to also look at different options and find if this is not only the most cost-effective option but also if this is the most customer-centric option? Is this, uh, user-centered? And the last thing is, uh, why us? So why do we believe that starting with, like, a Microsoft level? Why do we believe Microsoft can succeed with this innovation? More details of, do we have that expertise on the team to deliver, but also, do we feel we have an opportunity to do a good collaboration with the product team? Because sometimes the product team is, um, running so fast, and they have to fight so many fires with their existing product and existing service, um, that maybe now is, um, they don't have the time to do the collaboration with us. So, um, all those things have to align, uh, for us to have a reasonable chance of success, and even with that, like, as I said, like, not everything succeeds. So these are, uh, these are good prerequisites that you have, including addressing security concerns and understanding the challenges of AI that can arise in these situations.

Host:
But you also talked about, uh, you bring on board and offboard talent regularly because, uh, so, um, are there situations where you simply still don't have the talent for one reason or another?

Soundar Srinivasan:
Yeah, um, there is, um, because of the nature of this process, there is a gap between, um, the identifying opportunities for innovation to bringing on the talent, uh, and sometimes that, um, that entry doesn't work out, um, well. Um, there are, um, there are cadences in the recruiting cycle, so if you miss that particular window of opportunity, that can affect things. Not to say that it can't be overcome. This is just, uh, um, practical aspects of, um, people being onboarded. Um, then the other aspect, one of the reasons why I think the program has been very successful is, um, the innovation, uh, is continued and put in a good place to succeed after it leaves our, uh, program. Uh, we are the people who go with the innovation to the team, uh, and continue the innovation, not the product teams. So it's, uh, I've seen it not just here at Microsoft, but my previous work, uh, but also speaking to my colleagues in other industries, other companies. One of the big challenges of AI for tech transfer or innovation is there's a mismatch in expertise between the folks, uh, creating the innovation and the folks consuming the innovation. This is where AI in healthcare can play a crucial role, especially in addressing such mismatches with the right expertise. And one of the models that we apply through this, uh, talent transfer, so not just tech transfer but the talent that goes with the tech, um, that's passionate about the innovation. So they have a lot of emotional stake in making sure that the innovation feeds, um, and that, uh, allows the ability for them to follow up in the product game. Has been a, a significant contributor to success. It's almost like you have your own startup and even if you get bought, you still want to remain on board at, uh, see it succeed, so there is, um, there's that, um, personal connection that we tap into. Data utilization is also a key factor in ensuring this success. Generative AI plays an essential role in transforming ideas into scalable products, enabling new opportunities for progress. Security concerns are also integral, ensuring that all data and innovations are protected through robust systems. So, yeah, I know you also had some experience that I think it's Robert Bros and, uh, how did that this opportunity emerge from that, uh, from that experience that, you know, being able to get into Microsoft and lead these teams?

Host: 

How did Robert Bosch sort of, uh, prepare you and, and then sort of make it so that you were the right person to, to be on the, on this, on this project?

Soundar Srinivasan: 

Yeah, so, um, and BOS, we were, um, one of some of the earliest folks doing AI and when we started, uh, working on, uh, these types of technologies, there were a lot of questions regarding, uh, the fact that, uh, these systems can fail and these are not, uh, you know, failing at the rate of, uh, one in a million, right?

That's the type of areas that the company had previously produced, so they're the world's biggest automotive supplier. So, you don't want your car to fail at the same rate and your, uh... So, I got a lot of, uh, insights into, as I say, like what, uh, what works in research and what the, what the gaps are and bringing those innovations to the product team. So, in fact, one of the things I did there, uh, was to more out of research with, uh, based on recognition of the need that the innovation needs comprehensive and interdisciplinary, um, support which, um, you know, research and own is not able to provide.

So, that was one of the key self-discoveries that, uh, I had, uh, that was helpful in my previous role, which I was able to bring to my current job. And the other one, uh, most recently, what I did before joining Microsoft, uh, was, uh, what we call key product challenges, sort of like a product mood chart. So, for example, in the B2B that me, uh, in the autonomous driving space, how can we bring autonomous vehicles to urban environments? This is what, like, 2016, where we were seeing a lot of successes with, um, highway-to-highway driving, but it wasn't clear, like how do you I, uh, with these vehicles in, um, crowded open departments like in New York or LA or Chicago or San Francisco?

Soundar Srinivasan: 

Uh, and of course, not to mention much more congested cities in, uh, Hero.

Host: 

I can't imagine, uh, I can't imagine it, uh, you know, AI on was driving in, in France or something.

Soundar Srinivasan: 

It's a, it's a bit of, it's a different sort of driving, yeah. And definitely, the technology has, uh, come a long way now, but at the time that was, uh, exactly one of the goals that we pursued, like how we act a way that by bringing in all these, uh, disciplines together in a focused manner of course, a lot of, uh, a lot of success in these type of wood shocking effort comes down to prioritization first.

 So, we were, uh, I got in that initiative. I got, um, exposure to that as well. So, all of, uh, those learnings I brought in and, of course, in Microsoft scale, the question was not with the B2B. The question is like, what are the right top personalities and what is the right way to, um, care and develop and retain our talent? So, there were slightly different set of questions, but, um, I think that experience, uh, played well in, uh, creating this pretty much a caned up, um, and from what I know...

Host: 

Your PhD was actually in speech recognition, so how, how, you know, you seem to be, you've done quite a lot of innovative things over in general and you don't seem afraid to sort of jump into a different, uh, cycle. So, I guess that theme is consistent, but, um, what made you go from speech recognition to, uh, sort of almost a different kind of modality now, which is a visual representation? So, what made you go in that from, you know, from sound to, to light, basically?

Soundar Srinivasan:

Yeah, uh, I was thinking less in, uh, terms of modality. Uh, um, my recognition was the, uh, machine learning that is behind, uh, it is much more universally applicable because it's light, because it's light or because of the, the, because of the math, I would say the math, uh, right, uh, and also the, uh, some of the, um, compute tech that goes into it, that's somewhat, uh, similar. When I finished, uh, my, uh, PhD, it was still the peak of AI, um, where it was not as widely recognized as a dominant field. One of the opportunities for innovation at that time was to explore the applications of AI beyond its traditional scope. As the field progressed, the challenges of AI became more apparent, particularly in terms of understanding how AI could scale while addressing ethical concerns. So, there's that personal connection that we tap into, US. Yeah, so I know you also had some experience, I think, at Robert Bosch, and how did this opportunity emerge from that, from that experience, that you know, being able to get into Microsoft and lead these teams? How did Robert Bosch sort of prepare you and then sort of make it so that you were the right person to be on this project? At Bosch, we were among the first to delve into data utilization for advanced AI systems. When we started working on these types of technologies, there were a lot of questions regarding the fact that these systems can fail, and these are not failing at the rate of one in a million, right? We also began to address security concerns early on, particularly with the increased adoption of AI in critical infrastructure. These issues were a key consideration in the generative AI projects that followed, where the need to create models that could generate meaningful data was balanced by the need to ensure secure, ethical outcomes.

Host: 

Right, right.

Soundar Srinivasan: 

That's the type of areas that the company had previously produced. So they are the world’s biggest automotive supplier, so you don’t want your car to fail at the same rate. I got a lot of insights into what works in research, and what the gaps are, and bringing those innovations to the product team. So, in fact, one of the things I did there was to move out of research with the recognition of the need that innovation needs comprehensive and interdisciplinary support, which research alone is not able to provide. So, that was one of the key self-discoveries that I had that was helpful in my previous role, which I was able to bring to my current job. And the other one most recently, what I did before joining Microsoft, was what we call key product challenges, sort of like product mood charts. So for example, in the B2B space, how can we bring autonomous driving to urban environments? This is around 2016, where we were seeing a lot of successes with highway driving, but it wasn’t clear how do you handle these vehicles in crowded open departments like New York, LA, Chicago, or San Francisco? And of course, not to mention much more congested cities in Europe.

Host: 

I can't imagine it. I can’t imagine driving in France or something.

Soundar Srinivasan: 

Yeah, it’s a bit of a different sort of driving. And definitely, the technology has come a long way now, but at the time, that was exactly one of the goals that we pursued: How do we act on that by bringing in all these disciplines together in a focused manner? Of course, a lot of success in these types of shocking efforts comes down to prioritization first. So we were—I got hands-on with that initiative. I got exposure to that as well. So all of those learnings I brought in. And of course, in Microsoft scale, the question was not with the B2B side; the question is like what are the right top personalities and what is the right way to care and develop and retain our talent? One of the key opportunities for innovation was the ability to harness cross-disciplinary approaches that hadn't been fully explored yet. As we moved forward, we realized that understanding the challenges of AI and how to adapt and apply it in a practical, scalable way was crucial. Additionally, data utilization became a significant factor in refining and improving our processes, especially in terms of personalizing the way we approached talent development. There were also security concerns to address as we integrated new technologies, ensuring that the data we were using was protected from any potential threats. Furthermore, leveraging generative AI opened up new avenues for creating more adaptive systems that could assist in these efforts, providing us with a powerful tool to drive forward.

Host: 

So, there were slightly different sets of questions, but I think that experience served well in creating this pretty much a roadmap. And from what I know, your PhD was actually in speech recognition. So, how—you know, you’ve done quite a lot of innovative things over in general, and you don’t seem afraid to sort of jump into a different cycle. So I guess that theme is consistent. But, what made you go from speech recognition to almost a different kind of modality now, which is a visual representation? So, what made you go in that direction, from sound to light, basically?

Soundar Srinivasan: 

Yeah, I was thinking less in terms of modality. My recognition was the machine learning behind it is much more universally applicable, because it’s about the math, I would say, and also the compute tech that goes into it. That’s somewhat similar. But when I finished my PhD, it was still the peak of AI, where it was not particularly fashionable to be in AI, but definitely also the peak of speech recognition, a lot of progress there. I asked myself, is this what I want to do for the next several years, or do I take the technical ability that my education has afforded me and explore other opportunities? In fact, even before autonomous driving, one of the first areas that I started researching at Bosch was on small-scale sensor network applications. This was when small sensor networks were very new, but practically, an example of this is—before you had smartphones, we didn’t realize that WhatsApp and these sensors would be a thing. Now, you see there are all these applications where your phone can detect your performance, but back then, many physical parameters weren’t being collected. So, what can you do with the data that previously wasn’t available? Similarly, there are all these sensors embedded in your environment: motion sensors, video cameras, humidity sensors, or other things. It’s a different source of data, but the core question was how do you bring in knowledge from these data sources? From an AI in healthcare perspective, these are quite similar types of problems, where data from various sensors could be used to improve health monitoring. So from a technology perspective, these are quite similar types of problems. Obviously, location domains were different. Like when I was doing my PhD, it was really about making hearing aids smarter. There are so many possibilities for AI in healthcare now that the data from wearable devices or environmental sensors could be used to better understand and treat health issues.

‍

Host: 

What do you mean by smarter?

Soundar Srinivasan: 

Well, when you're at a party, and let’s say you have a hearing aid, typically, a lot of what these hearing aids do is amplify the volume. But amplifying the volume can be quite annoying if you're at a party, right? Like, not only do you have someone else that you’re trying to speak to, but there’s a lot of music, cutlery noise, and maybe someone yelling. You don’t want everything to be amplified. That becomes more of a problem. So, one of the things we observed was people would turn up their hearing aids because the value they were getting wasn’t optimized. It was just amplifying everything, which wasn’t ideal.

Host: 

Right, yeah.

Soundar Srinivasan: 

So, how do we—but as humans, we don’t have the technology in our hearing aids to filter out background noise when we go to a party. We can still have a conversation even in the loudest environments. Of course, it’s not perfect. Maybe you have to ask the person to repeat a few times, but it doesn’t get to the point where you can’t have a conversation at all.

Host: 

Yeah.

Soundar Srinivasan: 

That was one of the applications, and the other was, as we experienced just starting this podcast, how do you create noise-robust speech recognition systems? How do you, with a field microphone, get rid of background noise and echo?

Host: 

Yeah, exactly.

Soundar Srinivasan: 

And sorry about that—I was going to say, like, the way I was approaching this was by trying to understand how our auditory systems work. What makes human hearing so robust, and why are these artificial systems so fragile? What can we learn in terms of cross-architecture that we can bring into AI systems? Turns out, this was in the early 2000s, long before deep learning and all that stuff, but we had many similar ideas. We did realize, though, that for the attention addition to speech recognition, we used networks, but a lot of it was also based on exact algorithms.

Host: 

Say the last part again? What kind of models?

Soundar Srinivasan: 

Hidden Markov models, or HMMs, are a good way to think about these as extensions of finite state machines. Except, you know, being in a state isn’t deterministic. It depends on probabilistic transitions. It’s kind of like neural networks in a way.

Host: 

Yeah, exactly.

Soundar Srinivasan: 

Yeah, there’s a lot of good connection to deep learning theory and the mathematics of training these STMs—they all carry over. Fairly B. Okay, so I mean, you've done all this and sort of been able to transition at least quite, you know, from the foundational idea is sort of the same but, you know, from sound to light. How are you, when you think about onboarding people into this space and getting them to understand it quickly, like, how do you teach quickly, even if the person is quite technical, to the point where it has utility, and even the utility to the market? I mean, that's, you know, that's a very... There's a lot of people trying to do that right now to be honest, that the wave is here promising, you know, AI expertise in a short amount of time for a nominal fee. But, you know, how are you actually able to do this consistently over time, and how have you been able to do it?

Yeah, we don’t teach AI. I think what people get out of working with us is how do you make these innovations happen. Right? So there is a real aspect—people bring in people who have that theoretical expertise and we don’t place a premium on two different things. One is good theoretical knowledge, because one of the other things that we are seeing right now in the space, for example, and that was always true in the case of AI, is tools and technologies change very rapidly. So we put little to no premium on doing a particular framework, doing a particular programming language and things like that, but understanding the path behind it. But also, you know, not everything that we do is simple. In other disciplines, do we have a good sense of how do you create these distributed systems, how do you create these memory or computer-adaptive systems? On the product side, do you have a feel for innovation? Do you understand how to take the right risks? Do you understand how to go from zero to one, that type of customer journey, and identify the right opportunities? But also, how do you go through the journey to make it so? I think those are a lot of things that we... I think there’s still an education in one way or another.

Host: 

Yeah, so no, that's why I think, like, that’s sort of like our ping process and then when they come into the program, they get to practice this, right? And of those, they succeed, some of those they... we have learned from our...

Soundar Srinivasan: 

Um, so, there are many things that actually work well for us. One of them is the diversity of Microsoft products, to go from the entire stack of the cloud to things like the productivity suite of products and gaming, not to mention hundreds of other products. So that's a lot of diversity, and we can use that diversity to not only explore but also learn quickly. So we don’t have to necessarily be focused on one product and learn only from what works in that product and with that particular customer base. So we see a lot of patterns, a lot of things that we can do and patterns that we can extract from different domains, different groups of technologies, and learn from our PE as well. So, I think that is one helpful thing.

Host: 

Yeah.

Soundar Srinivasan: 

The second thing is that people trust Microsoft. You know, we place a premium on not exploiting customers' data. I mean, now we are very transparent about what we’ll use the data for, so there's a lot more trust in some of the other corporations. And we can tap into trust and understanding better. We have access to real-time interactions with our users, which allows us to learn their trade paths, and sometimes we find that the core idea that we had doesn’t really meet what the customers need or want, so we can quickly pivot. So, I think I would say, from a challenge perspective, it's less of, although right now in generative AI, everybody’s also figuring out what the right set of orchestration tools are and how to do this efficiently and cost-effectively. But more generally, AI has always been a very expensive powerhouse that...

Yeah, I think the biggest thing that we tackle is the use of scenarios, and how this will be set up for the POS. That's the biggest thing. Over the years, we've done so many new innovations, and we've had the opportunity to go back and look at the past. What can be improved, or what has worked in the past? The program has a vibrant group of people who are now on the other end of the innovation. They continue to support the program through feedback, which is valuable.

Host: 

I'd like to dig deeper into how you pick innovations. Do you have a framework for deciding what is worth pursuing? Obviously, you have skill sets to consider, and that’s the easier layer, but how do you assess whether something can enter the market and truly be useful? How do you make that decision, and how do you determine if it will provide 70% utility for the things you’ve picked so far?

Soundar Srinivasan: 

I’d say a lot of our process isn’t necessarily unique. We start with primary research, looking at market research reports and insights that we can rely on. We look for trends, stability, and interest. We try to understand what products people are asking for. It's not just about generating an end product, but about breaking things down. For example, people are looking for productivity assistants. That’s an insight we work with. So, from there, we go a bit further. We gather feedback from our collaboration partners, those who have more direct access to end customers on a regular basis. We also look at gaps in our product portfolio, and analyze what our competition is doing. In the context of AI in healthcare, we also consider how these insights could impact the sector and what specific needs healthcare providers may have. As we dig deeper into these trends, we align our innovations with AI in healthcare solutions to ensure we're addressing the most pressing demands and staying ahead of the curve.

Host: 

So you’re looking at what other players in the space are going after. What insights do you gain from startups targeting these needs? How does that impact your product strategy?

Soundar Srinivasan: 

Exactly, what startups are targeting can indicate gaps in our own product offerings. Of course, at Microsoft’s scale, we can't pursue every opportunity, so we focus on those with significant market potential. We also assess our go-to-market strategy. There are a lot of successful applications in the consumer space, but Microsoft’s big play is in the commercial or enterprise space. So, we have to consider if the way we succeed in those domains aligns well with our commercialization strategy.

Host: 

That makes sense. So, another question I have related to innovation: How do you ensure that you’re able to get something done in a short amount of time? Innovation is difficult because it’s not always easy to say, “Okay, I’ll leverage this discovery within two months.” That’s a tough thing to do consistently. So, how do you approach executing innovation, especially when you have to do it quickly?

Soundar Srinivasan: 

Well, it starts with understanding what is the role that we play in this whole product development process, right? That is primarily in addressing the technology gap. So, what that means is we become very focused on what exactly is the thing that we have proof that can work or prove that it doesn’t work. We don’t take on a very broad scope. By restricting the scope, I wouldn’t even call it like an MVP in a classic sense because we don’t go fully to the market either. We get to the point where we can do some initial feedback, but we are primarily just testing that whole idea: Is this feasible? Is it feasible with the current state of the art or with maybe an obvious extension? Once we demonstrate that, then we’re done. So, I think that focus is what allows us to be very fast, and we approach it like a product development process. We have a lot of ideas backlog that we look at: What is the next thing we could do in the next two weeks that allows us to either validate or learn more about what we’re tackling in this product space? So, I think it’s staying disciplined, and we also typically set ourselves a hard timeline of six months. Sometimes we extend it to as much as a couple of years, but I would say most of the innovation that we do, we can show visibly in about six months. But we can show that this doesn’t work, and of course, there’s a lot of this, fair enough.

Host: 

So, I want to move the segment into a little bit about how to get started in this because I think a lot of people are interested, at least now, in AI. You sort of got into the market after your PhD in a bit of a dry zone, and you made it out of that zone. But I think now, people are super interested in AI and how to get into it. Young people are thinking about how to get into it, and it’s sort of interesting because you got in when it wasn’t popular. Now, it is. So, what’s the fundamental thing that you should have when you want to get into this field, regardless of what experience you have in the market after you’ve learned what you need to learn?

Soundar Srinivasan: 

So, I would say the first thing you want to ask yourself is, what is the role that you see yourself playing? Do you see yourself in the role of an AI creator? Meaning, do you see yourself passing the fundamental AI capabilities, or do you see yourself in the role of bringing these AI capabilities to users and improving their experience? Those are very different setups. I think, for the former, you need good foundational knowledge in mathematics, statistics, and engineering. Spending sufficient time in that, whether in a traditional academic setting or an online setting, I think there’s no workaround for that. But if you’re more interested in, like, "Hey, how do I take this innovation and create additional innovation on top, bring new products to the market, new things that users can benefit from or improve the use of productivity?" Then what you want to do is find ways to constantly stay updated, be in the know. In addition to the fundamental education skills you already have, you want to rely on things like podcasts, newsletters, meetups, and conferences. I always find conferences to be a good way to quickly gather vast amounts of knowledge and sift through them to find what I actually care about. So, staying connected to those spaces is important. Generally, right now, for example, I’m finding it very challenging to move through all the innovations in AI in healthcare, so I also have to make time to focus on certain areas. If you’re, for example, in the business of ChatGPT, then you want to focus on AI in healthcare and other innovations. Yeah, those things and there are good options that, um, in the social media now that allows you to at least get a starting point, um, and then you can go ahead and go to the level of detail that you need. Some of the things about AI—sorry, go ahead.

Host: 

So yeah, so you talked about sifting through information. I think in this age, it's absolutely a challenge, it's its own piece of work. So how do you approach that? There's so much—I'm, we're not even talking about necessarily what is and isn't true, that is sort of, is brought out—but there's so much information to ingest for you just to catch up, you know, and people, it's a tough thing. You know, it's—you can't have a hobby, you have to have a full-time commitment in these days, which, and the amount of time in anyone's life has maintained itself to be 24 hours in a day, unfortunately. So how do you, how do you manage that? How do you, you know, with all these newsletters and all that, but these are really just links for you to keep going and learning more, which is fine, but if you just still again don't have that much time, so how do you, how do you—are you able to stay up to date, especially with, you know, the lab or I think you were sort of—you push back on the idea that it is a lab, but I guess a responsibility to manage as well, still keep up to date, right?

Soundar Srinivasan:

I try to be realistic and also focused on the innovation that I really care about. So there are a few DS that I want to follow more closely, and that's around the science of generative AI, that is orchestration of AI application user experiences. Anything else, you know, I have—I don't have the bandwidth to go after. And so I think staying disciplined is basically what helps, and I look for good aggregator sources. Things like surveys or meetups or conferences—I think those tend to be places where you get good summaries brought to you, and that allows you to also cover any black spots that you have in your investigation or in your knowledge base.

Host: 

What wasn't—I’m kind of surprised you talked about conferences as a place where a lot of information is shared for you, and has a lot of value for you. What is it about having those conferences that, uh, you know, is it that they tend to be newer information that comes through, or is it just like—

Soundar Srinivasan:

I think it's a breadth. Like I typically tend to go to, uh—I mean, I would say when I talk about conferences, I’m talking a little bit more fanatical ones, sort of like a, uh, W thing, which can become quite healthy. Yeah, it's a performance a little bit, it’s a show. Exactly, to your point. I think at these other conferences, they bring the latest innovation, but they also bring in diversity. If they’re large enough, there are five or six large teams that each of these events typically have, and each of those type of venues also have slightly different... some are like very theoretical, some are very applied, and some are in between. Some are NLP-focused, some are computer vision-focused, some are more application-agnostic, some are focused on health, or you can mix and match, and you get... I would say the two things: one is the latest, the state of the art, but also diversity that is both important in subject matter.

Host: 

Yeah, right. So if a young person was trying to get into the field, how would you—or, and especially, you know, these things change all the time—but what would you say is the most interesting part of AI you've seen so far? Obviously, generative AI, generally speaking, but what is it that, you know, maybe needs the most work and has the most blue ocean to swim in and find innovation? So what do you think that is?

Soundar Srinivasan: 

I think AI is— I would say right now, the just the cast of AI impacting our entire life. I think what you're seeing is now more like eye-catching type of applications. What I'm excit  has generated in that space. But if you look at, uh, the old innovation from a technology perspective that has happened, that's, uh, a right area. But even, uh, even others, um, other places like, for example, the, the, the, the, we are racing right now, um, that, uh, also feels a little clunky.

Host:

 The current world, like how do you make this more seamless and, uh, how do you, uh, help people, as we also talked about at the beginning, uh, how do you help people see and organize or support people in work? Work and life are also blurring, and people have needs that, uh, for that to be blending fully, like how you support those.

Soundar Srinivasan: 

So there I think, in the, what you might see the next couple of years are the first type of applications that, uh, happen. And then, um, if you're interested in research in this field, I would say, um, the science of understanding how these work and, uh, why, uh, hallucinations or fabrications happen, I think those are, uh, pretty rich areas to explore. Another one is, how do you mitigate bias, uh, in these systems for you? There's a lot of, uh, talk in the media about responsible AI, but that's really overloaded. Uh, breaking down into things like ability, things like finance, things like de-biasing.

Host: 

Um, and, uh, lastly, you know, a new area that has come up, which did exist before, it's like, how do you make this as secure? We talk about how AI can do this for you, but the onus is also important. Like, um, how do you prevent these systems from being hijacked? How do you prevent prompt injections, attacks, or safeguard against injection attacks? So there's, I think those are just areas we need to explore. So I think there's, uh, the next, uh, say, two to five years, good opportunities to invest yourself.

Soundar Srinivasan: 

Um, the healthcare is very compelling, the healthcare angle is very compelling. Thirty percent of data is in healthcare, and it's, it's not being crushed. Um, the percentage of GDP that, uh, the Western world, um, is devoted to healthcare is just upside in the Western world, because, uh, the issue that the global South is just catching up to reaching on, on the same level. So ultimately, the whole world will need this type of investment.

Host: 

And yet, the general perception is, AI is, uh, start with what seems, uh, needed or desirable. What's the role that AI can play again? Um, AI is often thought of as a new tool, um, as opposed to what I would advocate, is, uh, use AI to take a fresh look at the problem—not as, okay, there are like seven steps to the problem, I can use AI to solve, like step six or step seven.

Soundar Srinivasan: 

I think that's, and, uh, currently, you find a lot of frustration with AI innovation. People try this out, they find, uh, for the cost of investment, it's not as, uh, valuable. So there's a lot of excitement, but there's also some building and pull back, right? So yeah, the AI security problem too, uh, is attached to that. You know, because it is, you know, there's something, you know, there's personal data that, uh, in a way when people use AI, they're going to use it very personally, because the interface is very personal, and so the security problem becomes a little bit higher. Uh, and the accuracy problem still needs to be solved as well. So there are some, they are actually quite some fundamental issues that, uh, a person can come in and really, uh, contribute to, but it is very exciting.

Host: 

So, so like, I, I do, you know, want to keep this, uh, conversation, uh, to a space that where, uh, you can still go and do what you have to do for the day, because it is, it is a Sunday for you. I'm sure you want to do other things. So thank you so much for joining up the, uh, conversation, it was very insightful.

Um, I think one of the, I always want to make sure that people can, uh, be able to reach out to our guest, uh, in a way that suits them, so if there's any place that you would, uh, want people to reach out to you, or that be?

Soundar Srinivasan:

Yeah, I think, uh, personally to reach out to me would be via my, uh, LinkedIn profile. Um, and then, uh, I'll, I'll also drop, uh, some links to where they can.

See work of uh work of my team work uh right of some of the products that we have uh impact um right and you know there's a lot of uh places at Microsoft where uh people can plug into and uh take connected and learn and uh contribute to so and love a few of those um please do yeah the chat Tes yeah if if you actually it's fun if you are there there some you would like to talk about uh like some of the things that people are using right now that you've been able to attach yeah I can uh give couple of examples uh start with one that I like to quote of because actually it was a very first project that we took on and I had we had no idea of this uh suing but if you use uh any of our latest uh office site ofct products whether it's like word PowerPoint Excel you would uh you would find what is called like a context web flow it's like if you work in a particular section of the document uh the menu it customized to what you uh are trying to do at that particular space right and pointed time so there is an understanding of what you're trying to do there the not uh serve you the entire set of uh V options that the uh that was one of the first solutions that R te have paid I'm pretty proud of that because that's actually used by the comp of mil of users already friend congratulations thank you and then that anyway I friend um we have a ma a exciting new product called Loop so if you haven't used loop I would encourage you to go check it out um in Loop has uh thing called co-pilot uh like many other of the Microsoft products are beginning to have now um but that was actually the first one of the very first AI products that we uh you know way Tau right so go check it out and go check out the theil in Loop and uh some of the contextual uh floating uh menus on on the uh Microsoft Suite of products.

Host:
Well thank you so much for joining and um hopefully you continue to innovate and I'd love to see you back on the program and with some new cool things and some of these problems solved.

Soundar Srinivasan:
Yeah absolutely Chris, thanks, it was great speaking with you and yeah I'll talk the information with you and uh encourage your listeners to reach out.

Host:Fantastic, thank you so much. Thanks, take care.