Microsoft Research Podcast - 042 - Internships Ahoy! with Kirsten Bray, Wei Dai and Sara Beery

Episode Date: September 19, 2018

At the heart of any vibrant research community, you’ll find a diverse range of scientists. You’re also likely to find a robust internship program, like the one at Microsoft Research. This summer, MSR welcomed another stellar group of interns who had the opportunity to learn, collaborate, and network with colleagues and mentors who will impact their lives for years to come. On today’s podcast, you’ll hear the stories of three of these interns, each of whom came to Microsoft Research from a different field, with a different story and a different perspective, but all of whom share MSR’s passion for finding innovative solutions to the world’s toughest challenges.

At the heart of any vibrant research community, you'll find a diverse range of scientists. You're also likely to find a robust internship program, like the one at Microsoft Research. This summer, MSR welcomed another stellar group of interns who had the opportunity to learn, collaborate, and network with colleagues and mentors who will impact their lives for years to come.

On today's podcast, you'll hear the stories of three of these interns, each of whom came to Microsoft Research from a different field with a different story and a different perspective, but all of whom share MSR's passion for finding innovative solutions to the world's toughest challenges. You're listening to the Microsoft Research Podcast, a show that brings you closer to the cutting edge of technology research and the scientists behind it.

I'm your host, Gretchen Huizenga. Kirsten Bray is a grad student who lives at the intersection of psychology and technology. She joined researcher Asta Roseway in the HCI group and spent the summer making temporary tattoos a little smarter. Kirsten Bray, welcome to the podcast. Thank you for having me. Tell us a bit about your personal story. It's an unconventional one for this industry.

What's your background and what got you interested in high-tech research? So it's actually kind of funny. I have a degree in psychology and I got into human-computer interaction by spending a lot of time in my school's makerspace. I attended Spelman College and we had a kind of up-and-coming innovation lab where we were still trying to kind of piece together that space and encourage students to explore the intersection of the arts and technology. You've got these elements of understanding human behavior and human thinking and then they converge where you blend them together to make more usable and more engaging types of technologies. It was really interesting kind of even imagining the possibilities of something like that and the affordances going into a field like that would give me

as far as being able to maximize the existing knowledge that I have and the existing passions that I have. You say you just graduated with a degree in psychology from Spelman. Yeah. So are you in graduate school now or what are you what's your situation? Yeah. So I'm actually going into graduate school in the fall. I'll be attending DePaul University and I'll be starting a human computer interaction program. So I'm hoping I can get the technical skills there that I need to kind of bridge that gap from psychology to actually go deeper into human computer interaction. Wow. So this internship is a like gap summer. Oh, yeah. It is the perfect kickoff point for me right now, because I feel like a lot of the

thoughts and the interests that I had as an undergraduate were kind of just theoretical. I'm just like, these are the kinds of things that I'm interested in. This is the kind of space that I feel like I would work well in based off of the stuff that I know. And now it's actually time to get in and understand what a human computer interaction research environment is actually like. And then understand what kind of skills I need to get to be in a space like this and be productive in a space like this. How did you come to be an intern at MSR then this summer? I had been working on a project as an undergraduate. I was working with a professor who was trying to work on a device that would allow you to take pictures through a microscope using your smartphone.

And so while I was working on that project, I was trying to understand like, okay, what is the user experience process going to be like? What are users going to have difficulties with? How can we improve this device so that it can be the most usable for our students? I didn't really know where to start with some kind of research like this. And so I asked my professor for information about anybody who would be able to help me out with understanding how to gather this kind of data. And so she bounced me over to Melissa Boone, who is a Xbox researcher.

And by the end of our conversation, she told me that MSR had opportunities for internships. And so I just took a leap of faith and decided to apply. And so when I sent him my application, one of the researchers I wanted to work under was Asta. And I'm very interested in art and technology intersections. And so I asked to work with her, and she had seen my portfolio and my resume

and thought that I might be a useful addition to the team. And I was excited to work with her. Tell us about the project then that you're working on for this internship. So I am now actually working on the smart tattoos. They're tattoos that connect to a Bluetooth-enabled microcontroller that takes in capacitive touch input. And so they allow you to interface with different devices like phones or computers. And they can also be used as a power strip for an LED if you just want to have like a kind of decorative, yeah. If you just want to be decorative, you can do that.

But if you want to have some functionality equipped with it, like turning on lights or calling people from your phone, then you can totally do that just using a simple tattoo. And from your arm. Yeah. Just like on your arm and like. Or wherever. Yeah. I mean, we've had a lot of back tattoos. We've seen arm tattoos. I've actually had some fun experiences augmenting people's existing tattoos with the smart tattoos and

building off of tattoos that people already have because people already have been putting things on their bodies. And so it's like, why not add functionality to that beauty? It's been a really exciting project to work on. It's definitely right up my alley with this intersection of creativity and understanding user experiences and understanding new technologies. I'm kind of exploring the design aspect of this and like exploring like, okay, what more can you do with the actual visuals of this? How can you make something that's functional, but also has this beautiful form? And so it's a very basic, simple interaction, but it allows people to kind of see what's possible and understand that there aren't really very many limitations with this kind of technology as far as what you want it to look like or

what it can do. So let's drill in a little bit there, because it is interesting to me when you, you know, you see someone with a sleeve tattoo and you can add functionality to that. Where could this go? What might an application beyond just, you know, turning on your lights or whatever? Have you dreamed about that at all? phone numbers, but you can put on like a cute little temporary tattoo sticker and be like, hey, like if you get lost, call mommy, just tap your tattoo. It'll just call somebody. We had a hackathon project that was really cool where somebody was designing the tattoos where if you touch somebody else's tattoo, then you could exchange music files with them. And so just having that kind of, yeah, having that kind of affordance where you've

got like all sorts of different possibilities, whether it be like just fun or a little bit more inclined towards like practical usage, there are so many different possibilities. And so it's like, what would you do with a button or a trackpad if it was on your skin? I know somebody came to our office and was talking about how people's tattoos, like they always have different stories about their tattoos. And wouldn't it be cool if you could just touch somebody's tattoo and maybe it could display the story of what this thing is about. And so it allows people to communicate better. It allows people to be a bit more creative with their interfaces and actually choose what they want to say with their technology.

So this hackathon you talked about, you guys called it the Hack-a-tat. Oh, that was so fun. So we had basically a two-part challenge and we did a workshop first and we did that over in the garage and had a bunch of people attend and just get their feet wet with experiencing what working with smart Tattoos was like and seeing what kind of ideas they could come up with. And so we took the challenge and extended it out from there and created this Hackitat challenge for the hackathon and had people do the same thing with a little bit more time

and asked them to use these materials and come up with new designs and fresh ideas. And so we got two things out of this. One, we got all sorts of cool ideas about just what people were actually interested in. And we also got a good feel for what people are actually experiencing when they work with these things, because we're hoping to release a smart tattoo kit and actually allow people to play with these and really be engaged with the making process. Yeah, yeah. Who could get the kits? Our hackathon winners will be actually getting some kits to play around with

and hopefully we'll be able to release a second generation kit that has a smaller, more easy to use board and more easy to use connectors. And hopefully that kit will be something that can be launched and turned into something bigger that's more available for public use. Still ambitions for the future, but I'm very hopeful. We've gotten a lot of good feedback. And so I'm excited because I really want to see this after I leave.

Like I want to just go to a store and see, hey, there's this kit. Because I would totally use that. Right? And then you could say, I was there at the beginning. I saw this. Talk about the people that you're working with. Who's your advisor and the group you're in?

What have you learned from them in this experience? My mentor is Asta Roseway, and she's been great. She's full of creative energy and just an awesome person to work with. I think she's always been very supportive of the things that I've been doing and encouraging me to kind of look at things in new ways or try out different things. And we've also got Paul. He's our engineer on the team. And so he does the software side of things. And so he's been actually helping me with one of the other tattoos that I'm doing where it senses where you're touching. He's actually doing the programming for that. And so I quite appreciate him. Like, I feel like everyone who I've worked with here has been very positive. It's always very positive energy. Like, everybody's very

receptive to new ideas. Everybody's willing to help. And we just have a lot of fun. So, Kirsten, what's next on the horizon for you? Immediately, you're going back to graduate school at DePaul. And what do you hope to get and do? And where do you want to land? What are your dreams for the future? I've seen so many different possibilities. I think being at MSR, obviously, academia is looking a whole lot more appealing. Research. People are really very intense at MSR. And I love it. I love that intensity, like where people are just so attached to this project and they want to see it through to the end. And so I like that spirit and that vibe of just creating and exploring new things. And so I would definitely like to go into research sometime in the future, but I'd also like to go into industry and kind of experience what it's like to be a designer on the product half of things

and like work in that kind of field. So I feel like it's an open box for me right now, but there are certain things, like I definitely do want to get my PhD. I would like to, right now, I'm looking at human computer interaction PhD programs and also perhaps digital media programs. But there are so many things, like I feel like there are so many more possibilities

right now. And so I am open to it all and hoping for just lots of new experiences. Kirsten Bray, we are excited to see where you're going to end up. We'll keep an eye on your career and your work, and hopefully we'll see you on the cover of Wired Magazine one day. Thank you. Wei Dai is a math whiz with a need for speed. He spent his summer working with researcher Kim Lane in the cryptography group to accelerate computation on homomorphically encrypted data using GPUs. Wei Dai, welcome to the podcast today. It's great to have you here.

Yeah, thank you. It's nice to be here. Tell us a bit about your personal story. So originally I'm from China. I did my high school and university in China. And then I come here to Worcester Polytechnic Institute at Massachusetts for a master's degree. And during my first semester there, I performed well in the class, and they offered me to do a PhD in the lab, and I accepted. And ever since, I've been working in this field of cryptography. Did you know from the get-go that you wanted to do cryptography?

Let's just say you're good at math. Is that the natural place that you go? Or, I mean, what got you interested in that particular field? Actually, before taking that course, I don't know what's the meaning of cryptography, the English word. So after that, I figured out, okay, cryptography is not about solving some puzzles. It's actually about using mathematics to build a magical scheme that protects the data. So I was very interested and amazed. Where are you in the process of your degree?

Yeah, I'm still in Worcester right now. I'm about to finish my PhD. So after this intern this summer, I'll go back to the school, start planning for my dissertation and schedule for defense. And in fact, Christine Lauter, my manager here, she is going to be on the committee board. What is your dissertation going to be on? So previously I've been doing research on this homomorphic encryption. And also I have a specialization in doing it really fast on GPUs. And this is going to be the theme of my dissertation. How did you come to be an intern at MSR?

It was a really good chance that last year in summer, Microsoft Research, the crypto group, actually hosted one workshop for the standardization of homomorphic encryption. And they invited a lot of very famous scholars to Microsoft Research at the campus. And I actually, I came with my advisor, and that's how I met Christine, Kim, and everyone in the group. And we introduced our work. Everyone was thinking, okay, we never know on GPU, it can be so fast.

And Christine got interested. She asked me, so would you like to do an internship next year, summer? So we scheduled it out. And that's why I came here. Well, give us a little overview of homomorphic encryption and what its purpose is. Basically, it's not just any cipher or

encryption scheme. It does not only do the encryption protector data, but also allow computation on the encrypted data, which wasn't possible before. So just to clarify, if I encrypt data and then I want to do something with it, as you say, operate on it, traditionally I've had to unencrypt it to be able to do something with it, as you say, operate on it. Traditionally, I've had to unencrypt it to be able to do something with it. Right, yeah. With homomorphic encryption, you don't have to decrypt the data, but you can perform computation, and whoever did the computation returns you the result in an encrypted form,

where you are the one who owns the data. You are the only one who can decrypt the result. So you've got the key. Nobody else gets the key to decrypt it, to be able to operate on it. And then you can do this homomorphic encryption, operate on the data, get it back, and then you can see what happened. Yeah, correct. Tell us about the project that you're working on, maybe starting with the problem of the speed of homomorphic encryption to begin with? Ten years ago, people started building homomorphic encryption schemes mathematically and also doing software or hardware implementations. At the beginning, the encrypted data are so large

compared to clear data, and also the computation on encrypted data was so slow. Microsoft Research has started this library called the SEAL library, which is a simple encrypted arithmetic library. So the goal of building this library is to simplify how to use homomorphic encryption and push for more efficient computation. And for the past three years, the performance, the security, the user experience, everything has been improved by a lot. And they would like to see how can we accelerate the library on the GPU, a different hardware. And we also have another intern who's doing the implementation on FPGA, which is we are really trying all the hardware devices to see what we can get.

MELANIE WARRICK- So the speed up of the runtime is what you're looking for. YUFENG GUOIERI- Yeah, exactly. The faster, the better. And according to our experience on a GPU, usually you can achieve at least 20 times speed up. And if the scheme is really well designed, you can reach two orders of magnitude, which is Henry times.

And on FPGA, you can achieve something similar. So tell us about your library. This is a fairly unique, innovative thing that you've been working on. Yeah, so suppose I'm a data company. On the client side, I can provide them software to encrypt their data or doing decryption later. And the library itself also provides service for the data company who can design an application of machine learning or genomic data analysis. They provide the data analysis service to the clients in an encrypted form.

Sure. With the help of the SEO Library, everything can be simplified. They don't have to understand too much about the home of encryption. They just know this is realizable. And using our library will make your life much easier. MELANIE WARRICK- So how has that played out?

I mean, is this available yet? Or is it still in the research phase? Or where is it? JIEBO LUOFENBERGER- So the research never stopped Or where is it? So the research, it never stopped. So what we do is that we keep releasing new versions of the library. And the library is currently published under Microsoft license. And we do have a lot of clients and other groups who are using this library internally at Microsoft Research.

What do you hope to see in this area in the future? We do see a promising future in this area because it's still fairly young, over nine or 10 years so far. And now because there are an urgent need for privacy protection and there's regulations and laws making towards this direction, I think

home office encryption has a much brighter future. Tell us about your advisor and the group you're working in. What have you learned from the people you're working with this summer? In our group, we have four, five full-time employees, and also we have five new interns this summer. So mainly, I'm the one who's working on GPU. We do have a lot of meetings together because different people have different opinions

and they might be able to see, for example, my colleagues will see what did I do wrong in these procedures or somewhere. And they can correct me and give me another idea. That's very helpful. So it's group constructive criticism. Right, right. Who's your advisor?

So Kim Lane, he's the one who makes this SEAL library really useful right now and ends up being used by many other groups. Like we have people working on machine learning, people working on compilers, and everyone is curious about how can we use this, and they are trying to use it right now. Hopefully we can come up with some joint work and end up in a product. Yeah. So when you came here, did you bring this project with you, the GPU speed-up scenario for homomorphic encryption?

So, accelerating seal library on GPU has been planned by the Microsoft Research and also in our group. However, in this field, it requires some special skills. Not many people in this area have actually done GPU implementations. I'm one of them, and I was lucky to have the best performance out of it. When I came here, they just told me, okay, you are specialized in this, please do this. Just- There you go.

Yeah. Tell us a bit about the performance, like that you've alluded to, that you had the best performance. What are we comparing that to? YIUAN LIU- Traditionally, we were just comparing, suppose we have an algorithm. I implement it on a GPU and achieve probably more than 100 times speed up compared to running it on a CPU. Even compared to the best CPU implementation, it could be 100 times more.

And now this is not just a simple algorithm. It's a whole library. And even the algorithm itself is hard to implement. It involves a lot of research work because talking about how to do this computation is one thing. Actually, adding two numbers on a computer is different. And different hardware is a new area. This summer, I've been working on accelerating seal.

We are expecting more than 20 times at least. Hopefully, again, faster the better. What's next for you, Wei? This fall, I'll be focusing on finishing my PhD and doing the dissertation and the defense. After that, I would like to go in industry. And Microsoft Research actually is one of the best destinations for me, if I can come back. My bet is yes. Thanks. I hope. So, but a career in research continues. Yeah, I feel this area is too promising to leave right now. I would like to stay in this area and continue my research.

I think right now internally we have been doing something very meaningful and we do have some prototypes that is driven by home-ramp encryption and it showed very good result. Not ideally practical, but good enough for you to consider using it. Wei Dai, this podcast has been a very good result. Thanks for coming in and sharing this with us. It's so interesting. Thank you. Sarah Beery may be the only MSR intern who was also a professional ballerina, but her

passion now is using technology to help protect the environment. She joined researcher Dan Morris and spent her summer building AI models for motion-triggered cameras used in wildlife conservation. Sarah Beery, welcome to the podcast. Thank you. You have a really, I say this to a lot of people, but you really do have an interesting, unique, unconventional path to this place. Tell us a bit about your personal story, and then we'll get to the details of what you're doing later. But I want to hear about you. Like you said, I definitely had an unconventional path to tech. So I grew up here in Seattle, and I moved alone at 16 to Atlanta to dance at the Atlanta Ballet. Ballet was my entire life.

I absolutely loved it. I continued on from there to dance in San Francisco and New York City, actually all over the world. But actually, while I was in Atlanta, I was living really close to Georgia Tech. And I was really broke because I was a ballerina. And Georgia Tech would have these open talks that would have free food. And so I started going to these talks and hearing about all of this really amazing research basically just as a ticket for free food. And I actually wasn't even sure I was supposed to be there. You're probably supposed to be a student, but I lived close enough. It was fine. And I more and more as I was hearing what people are doing, really started to think like, okay, well, I can't do ballet forever. And I always really loved school. Like, maybe I'll go back after ballet and become an engineer so I can help solve these problems for people in the world. And I've always been really passionate specifically about environmental sustainability and conservation. Probably some of that comes from just growing up in the beautiful Pacific Northwest.

I got this idea in my head that I, you know, after I retired from ballet, I was going to go study to be a green energy engineer. I retired actually kind of early for a ballerina, but a big part of that was because I had this dream to do something else as well. And so I went back to school. I went to Seattle U, studied electrical engineering, and along the way found out that I actually loved math. And this was astonishing for me because I absolutely hated math growing up. It was something I never felt good at. And I was in really high-level math classes in school. I took calculus in high school even before I graduated early. But I thought I was really bad at it.

And I wasn't. I'm actually quite good at math. But it took until college and until I had teachers who really encouraged me for me to even think that that was an opportunity for me. I thought I was really going to struggle with the math classes. And instead, with the right teachers and with the right encouragement, I totally thrived. So I added a math degree. Along the way, I found out that actually I was really passionate about research. And so I ended up, now I'm in a PhD program at Caltech doing computer vision research and specifically focused on environmental sustainability and conservation. So where are you in that process right now? So I just finished my second year. I'll be starting my third year in the fall. All right. So talk about your path here

as an intern at Microsoft Research this summer. So as I said, I've been really passionate about environmental sustainability from the time I was a child. And when I actually decided to go to Caltech, I worked with Pietro Perona, who is an absolutely amazing advisor. One of the reasons I decided to work for his lab is that he has shown this real commitment

to projects that are helping the world or helping people. And that was something that was really important to me. After I finished my first year, I was looking for projects that would really further this goal. And I had previously done research with camera traps, which are this really interesting way that technology has been assisting people who study wildlife conservation. So instead of having to do field work or a catch and release type study, which is quite invasive to monitor animal populations, behavior, the density, like what the effects of climate change and urbanization are, they're able to just put cameras out in the field, their motion or heat cameras. Really the bottleneck there is just sorting the images.

So these people go through, they put out hundreds of cameras, they do a large-scale study, and then they spend hundreds and hundreds of hours just sorting the photos by what they see. I was at Caltech trying to figure out what project I wanted to start with, and this group from the National Park Service and the U.S. Geological Survey reached out to Pietro and they said, you know, we're really interested in trying to find a way to use computer vision to automatically go through this process of detecting animals, classifying their species and long term, maybe even doing individual identification. And I was like, actually, I have a lot of experience with camera trap photos already. And that sounds great. So I started building models for them and curating their data set. And then I went to Grace Hopper last year.

And at Grace Hopper, I just so happened to stop by the Microsoft booth, and I was talking to Ossie Roycroft. And, you know, I was like, I don't really need an internship. I'm not really sure that it works with my PhD. But I was curious if you know what's going on at Microsoft Research at the intersection of AI and environmental sustainability. And she was like, oh, you should reach out to Lucas Joppa, who at the time was the head of this nature plus computing group. That was exactly what I was looking for, like somewhere where you could do AI for environmental sustainability. I reached out to him and he was like, actually, we're, you know, it's not been announced yet, but we're going to start this huge new program at Microsoft called AI for Earth. And it was really serendipitous for me, but it's also so inspiring to see that Microsoft

has made this commitment. So this is a company-wide initiative to basically sit at the intersection of sustainability and AI. Hearing about that really encouraged me to say, OK, yeah, no, I'm happy to come work for you. I'll be an intern. I'll help build your tools. And actually, it's kind of perfect because in academia, I can it's widespread and easily accessible for the people the tool was meant for. And AI for Earth is this awesome bridge between that. It gives people like me the opportunity to build models, build tools that can use cutting edge computer vision algorithms to work with this complicated data. And then you actually are able to use this team

of software engineers who are building APIs and building tools and dealing with the user interface. And, you know, hopefully by the end of the summer, we'll have actual tools that scientists can use to actually start sorting their data. Tell us about the project then specifically that you're working on as an intern here. What I'm specifically focusing on is I'm taking these camera trap actually start sorting their data. Tell us about the project then specifically that you're working on as an intern here. What I'm specifically focusing on is I'm taking these camera trap images and I'm training detectors, which are able to look at an image and not only tell if there's an animal in the image, but actually localize the animal. You are able to train models that work astonishingly well on camera traps, as long as you are using them at exactly the region or even in a region we've dealt with before, but it's a completely new set of cameras, that these models will be able to work for you as quickly as possible.

Yeah. And so what I'm doing is I'm working on trying to train detectors that are able to generalize well to these new areas and these new regions. So I don't know if you've heard of Zooniverse, but Zooniverse is this big citizen science platform for scientists to get crowdsourced data annotations. And 40% of their projects are camera trap projects. And the largest one is this project called Snapshot Serengeti, where these cameras have been out continuously for 10 years in Africa. And I was talking to this woman who has been working with many different camera trap groups all over the world. And I asked her, what's your estimate? Like back of the napkin, let's do a calculation. How many active camera traps do you think there are in the world right now running? And she was like, thought about it

for a second and said at least 500,000, maybe a million. Wow. That's amazing. But if you think about how cheap cameras are, some of these cameras are like a hundred dollars. Yeah. Think about how well you could monitor the effect of a new science policy on your wild animal population. If you had the real-time ability to put out a camera and just track everything you saw and just stream that data, and you could actually build sort of a global scale population density map for all these different animal species. I don't think it's going to happen soon, but I think it's definitely going to happen. And I'm hoping my work can contribute to that. Tell us about your advisor and the group you're in, the people that you're working with. What have you learned? What's been your experience? So my direct boss is Dan Morris.

He is gung-ho, passionate about saving animals. And he has been a really valuable resource for me, just in terms of learning what it takes to build a tool that people can use. He has a lot of intuition about sort of how to navigate the politics of a large company and how to increase visibility and how to basically make sure that people care about what you're doing so that it actually gets done.

And that's something that I hadn't really had to think about as much before. I mean, I have an NSF fellowship, so I basically have the ability to work on what I care about, but I don't have to convince anyone else to care about it yet. But now I'm realizing actually how important that is. I think the ability to share your passions with others and to really make it super clear, like why you're passionate about stuff and get other people excited about it, too. It's probably like the best way to make things happen. Yeah, that's a skill that you will want in the future. What's next for you, Sarah?

Well, I'm going to work my butt off for the rest of this internship. Hopefully take advantage as much as possible of the resources I have here. I'm going back to Caltech right when I finish. Being in a PhD is, it's a long haul, but it's exciting. And it's kind of the first time in my life that I haven't had to be thinking immediately about like, what does the future hold? And I'm pretty sure that if I just keep working in areas I really care about, that I'll be able to find a way to continue to do that after I finish my PhD. And what would you see yourself doing in the future?

I don't know. If Microsoft keeps funding AI for Earth, maybe I'll work here. I would love to continue to find ways to bring AI algorithms and bring the skills that I'm gaining at Caltech to play in areas where you can make a huge impact, but they traditionally don't have access. And hopefully you won't need to look for places that have free food. Yeah. Though I will say, I think being really poor at least once in your life is very good for you because you're not afraid of it anymore. I hear you. What's harder, ballet or high-tech research? Ballet. In some ways, you know, doing research somewhere like Caltech can be insanely difficult, but I don't wake up in pain every day.

And I got to say, it feels pretty good to be able to like get out of bed and not feel like an 80-year-old. Do you still dance? I do. Caltech has a ballet club. A club? Yes, the Caltech has a ballet club. A club? Yes, the Caltech Ballet Club. I take classes on Sundays on a wood floor in the gym. But it's nice to just keep my toes in. Literally.

Literally. Sarah Beery, it's been so delightful talking to you. Thank you for coming in today. Yeah, thank you so much. This has been really fun. If you're interested in applying for an internship at Microsoft Research, visit microsoft.com slash research slash careers.