Subscribe
Share
Share
Embed
Crafting a Future of Synergy in the AI Era
A series that zooms into the intersection of artificial intelligence (AI) with daily human life. It features interviews with leading AI experts, thinkers, and tech innovators, offering viewers a behind-the-scenes look at the advances in AI. With a conversational tone, the film tackles the big moral dilemmas and the future implications of AI, all while keeping things relatable. It's not just tech talk; it's about what it means to be human in an age where computers are getting smarter. This series presents a vision of a future where humans and AI could work side by side, changing the game for everyone. It's an accessible take on the potential teamwork between man and machine, suggesting a world of possibilities when we embrace AI as a part of society.
Welcome at another episode of Human Hearts Silicon Minds. Today we're interviewing Rob to learn more about the intersection of AI and humans in the startup scene. Hi Rob, welcome and thanks for joining us here at the San Francisco Microsoft garage. It's nice to have you here. Can you please introduce yourself? Well, hi I'm Rob Ferguson. I'm the head of AI for Microsoft for startups. Before I joined Microsoft I was a CTO of a few different startups all the way from really small all the way through exit. I've had two exits including one of YC's top 10. And I was also the VP of engineering for Machine Learning Unicorn. I've been doing machine learning for about 20 years and my passion is helping other startups be successful in the cloud because I had that same success in my own startups. I want to help other startups have this same success. So you've been using machine learning tools since it became popular? Back when I had to build them myself. Okay, so I had even built open source libraries to do things like support vector machines and neural nets when you couldn't actually call one and you'd have to write it in C++. So how do you see the changes from then to now? It's immeasurable. When we think about machine learning it's really funny because when I was doing all my graduate work the entire effort was making it as efficient as possible because we thought for sure there's no way that you're going to be able to use these techniques that require a ton of processing, a ton of storage. You're going to have to come up with some other breakthrough to make it more efficient in order to be able to teach a computer to use data to learn new things. But then in practice that's not what happened at all. You know around 2006 the cloud came and we got really good at breaking down really big problems into small composable parts. And so you can think of like instead of if I were to do a jigsaw puzzle by myself versus have my friends do different parts of the jigsaw puzzle and then we all put it together in the end. After 2006 there were a number of innovations but they all involved the opposite of efficiency. They involved using more computers. So you know around 2012 deep learning really came into the forefront and deep learning said we'll just let the computer start guessing at different features rather than us come up with them ourselves. And then along came transformers which was something along the lines of well you know how we can do auto complete with a dictionary? What if instead of using a dictionary we just say just guess the next thing based off of things that you've seen before? And then from there it got really wild because we blew the lid off of scale. When you look at language models like a GPT it would take over 255 years to build that model on a home computer. And so it really absolutely requires this massive scale that comes with the cloud. And with it really strange things start happening. You know by the time you've observed five billion examples these models get good at translating between different languages. When you get to six billion examples, eight billion examples they can start auto completing answers to questions. GPT has at least you know 150 billion examples that it's trained on. And these effects of scale just can't be understated. My partner is a physician and in older fields you never have an example where experts are surprised by something that's coming out. You know you work for many many many years to figure out what the next thing is going to be. But here I don't know an expert that wasn't absolutely floored at the tremendous responsive scale when building these neurological models. And as a result it's just an incredible time to be alive. You just see the most incredible innovations and the pace of change is absolutely fantastic. And so I just can't understate how rapidly that has changed in my career from building tools by hand, training models by hand. You know if you want to have a really expensive bill you have your machine learning engineers do all your data labeling. But that's how everything started. And now all of those things are built at scale using unsupervised techniques. In a way that is only possible with a combination of cloud technologies and really superior tools. With the power of the clouds for everyone can have supercomputers. So that's what's right you actually can build an on-the-fly supercomputer which is just incredible. You know if you in graduate school if I wanted to get time on the school supercomputer there's like a whole process where I could get like a day if I had a grant. And then you would like prepare your experiment well in advance so that you knew that it was going to be able to run in the amount of time that you had. And if it didn't complete you just got turned off. Now you can just rent one by the minute if you want to. It's a completely different world. Yeah it is. Now AI is already impacting quite some people's lives. How did it impact your personal life? You know in a bunch of different ways. You know one really simple example was my partner had a lot of student loans. And we used AI to write grant applications to come up with like you know ways of paying off student loans. Because a lot of people don't apply for grants they're all over the place. And it's really time consuming and tedious to do. And so we used AI to write a grant application. It worked and student loans will be forgiven as a result of it. So I can think of almost no aspect where I'm not using AI in my daily life. I use voice transcription every day. I use generative AI to help build images and diagrams for talks that I give. I use AI when I broke my hand and I needed help completing the different tasks that I was doing as a result. I try to integrate AI as much as possible and stay on top of the state of the art as much as I can. And as a result it's just absolutely fascinating and fantastic. When I first started working on this stuff you would be years before you would see an implementation. Now I see new things coming out on Twitter. It's a different world. Now you try to stay up to date with everything that is there? I try and very unsuccessfully. Even though I would say I read an AI paper most days. Okay, now that's us as tech people. How can we achieve this also for the general public? To be able to make sure that they also get all those things that are available. Yeah, you know it's fun. I think you have to have a learner's mindset and just be really curious. Because as you're curious you don't have to stay perfectly on the state of the art. The truth is the best time to get involved in AI would have been 20 years ago and the next best time is now. Right, like the best time after that is now. Just get started right away and have that learner's mindset. I really enjoy tutorials and things that we have on Microsoft Learn. But also I'm like an avid YouTube addict. I love AI Explained and Two Minute Papers. I spend a lot of time just enjoying where are people having discoveries. And then I spend a little bit of extra time just trying to look at an implementation. See what startups are doing and go and have a lot of conversations with people like yourself. So yeah, can we tell those people to start reading papers then? Just the general public or how should we handle it? I think the YouTube channels and the Twitter read-ups and podcasts will do a better job. You can follow people like Ali Miller. You can follow people who do a really good job of sort of staying up. And then breaking it down into a way that's a little bit easier to understand. Okay, now the opposite of it is that also many people believe or understand a little bit of what Hollywood is telling us about AI, singularity, super intelligence. What's your opinion about that? You know, I understand where a lot of the concerns come from. And I also understand that people may not fully understand where the state of the art is. A lot of the state of the art, if we're honest, is more like a really fancy auto-complete than it is something that reasons or plans. So for example, if I were to take a glass of water and I had ice cubes on the bottom of the glass of water. And AI might be able to say, hey, those ice cubes are in the wrong spot, but doesn't reason, oh, they need to float. And the biggest concerns that people have when they're talking about AI in public who are experts is that AI kind of does what you tell it to. And that can have some really weird results. Now there are catastrophic examples that are very unlikely to come true. And then there are more reasonable examples of what you expect. And the result that sometimes AI practitioners are not necessarily themselves ethical. But I know at the same time that there's these entire new fields that come from this. My job never existed before now. In fact, I know many other people whose jobs haven't existed before now. My goddaughter, she makes like assets on Minecraft as like her summer job. You know, she sells other people designs online. I had a paper route. She's selling like Minecraft t-shirts. It's a different world. The reality is that AI ethicists, AI safetyists, these are actual jobs that we do. And we start building other regulations and things to make these systems safe at scale. Microsoft has a tremendous responsible AI team. But where the concern comes from is sometimes misplaced in thinking, oh, super intelligence is going to create, I don't know, the terminator or something. But the other things that can go unnoticed can be even more pernicious. And that's stuff like the Rubicon has been crossed. We had a period of time where we could have potentially labeled all information that is AI generated and had some regulation around it. But now we're already having AI models being trained by AI. We're already having AI data out in public. We don't necessarily always have the right protections to help people understand those things. Like what happens with a deep fake? How do people create misinformation? These are really people problems, more so than super intelligent problems. And regulations, safeties and protections that make sense for us as a society have to come as part of it. And that creates an entire new class of responsibilities and things that folks like myself working in AI need to be thinking about all the time and how to use things responsibly. So I think you always want to think of AI as silicon plus carbon. It's not one without the other because it's not as though we have a super intelligence where it's really cheap to run. It costs millions and millions of dollars and requires hundreds of individuals, a massive coordination and work with the top companies in the world to build one of these models. It's not like a genetic algorithm that you can just let go and reinforce itself and doesn't cost any money. It's estimated that GPT might have cost $20 million by some estimates. That doesn't happen on accident. AI doesn't run away, but people can use it irresponsibly. And in the worst case, it is a source of disinformation. And that is something that we simply have to make a choice to say, we don't tolerate that behavior and put in meaningful rules around the responsible use of AI. You work a lot, well, mostly with AI startups. Have you seen AI startups like, who, what are you guys doing? You should not be doing this. I've seen some funny behavior from very early stage startups. One of the things that I encounter all the time is person that says, you know, I was really into crypto and now I'm really into AI. That is a very, very important thing. That is a very enthusiastic space to be in. And honestly, there are even very good startups who that is their story, but have maybe not considered some of the bigger picture and of building AI systems and like how to build compliance systems and ethical systems and things like this. And so it's our responsibility to make sure that we're continuing to create those teachings and help people understand the difference. With knowledge, I very, very rarely encounter a company who I think has bad intentions. Most of these folks actually want nothing more than to build things that work really great for the people who are there. At times, we find people who have unintended consequences of their behavior, such as ingesting copyrighted data sources and not making sure that the people who are using that are aware of what's happening. That happens more frequently by accident than on purpose. And so we have to learn how to build these systems in a responsible manner. Which is not always easy because as a startup, you have always a particular goal and then not always thinking about the things that are around it or the consequences are linked to it, of course. You know, one of the things that I learned through my own journey was that there are like great organizations to help you out, do their job. And they help you out, do this. In my own startup, a lot of times I thought I had to do everything myself. And, you know, I kind of knew a few things that clouds would do for me. I knew I could get startup credits. I knew I could get like a roadmap review. But I never thought about just engaging in other forms of partnership. And frankly, the clouds really know where are the places where people go awry, when thinking about building something that's like an enterprise class application. And so they're really effective ways to partner, even if you're a startup. And there are so many programs to help them. I just think that a lot of times they just don't even know. So, you know, that's what we do today. We tell people, hey, we can be a great thought partner as you're going through and thinking about your system. We have an expert network that helps people connect with other people to really understand where are the potential pitfalls? What are those trades that I might not be aware of? Are there any startups you've worked with that really, well, not shocked you, but like how AI really has helped them grow exponential? So many. You know, I think of companies that truly achieved virality, like CopyAI, you know, Gen Zs of the world, whatever is after Gen Z, I'm not sure. I don't actually use TikTok, but sometimes I'll get TikToks from like my nephew and things like this. They really have to think about packaging out these materials and, you know, CopyAI said, hey, we're just going to give you like different titles for your Instagram post or your TikTok video for you to be able to go through and try. And so it had this immediate viral success of people who could just like, you know, test at some immediate scale, different copy that they might use in a different place. And as a result, it kind of creates a community of people who are really interested in, you know, that potentially going viral, hitting the audience that they want to hit. And so we've discovered inadvertently that community ends up being very important. When you think of a company like MidJourney, they exist entirely as a community in this weird world that people hardly even think of as a startup because they're off on a Discord server where you can like talk to other people about building the image that you want to have. And it's an incredibly engaged community and they create incredibly beautiful aesthetic pictures as a result of it. I see community being an element in generative AI all around the world in a lot of different places. The biggest factor is that when you have a tool that will kind of do whatever you say, sometimes you have to have a little coaching on what to say. And there's a little bit of writer's block and the community aspect is sort of like, oh, this person said that, that's kind of similar to what I want to say. I'll say this and then you might make a leaderboard. I've seen lots of different examples of people doing this to just incredible success. And when I look at, you know, like Prisma Labs that made 31 million dollars of profit off of three million dollars of cloud spend, like they're the same company that made the Facebook profile photos that went viral. The reality, I think it's an app called Lensa. The reality is that those opportunities to just all of a sudden have this massive appeal and success are really coming through the community of people who are using it. So community ends up being just this tremendous factor in this adoption of generative AI. And as we as a society learn how to use this technology, which we only just have learned the very basics of. I think the example of Lensa is quite a good one, how it's not the biggest project I think to build because the idea on itself is quite simple. But they made it so public and so marketing wise very well that everyone wanted to make use of it. I also paid for it to just try it out. And they know like you get kind of your reward out of it, which is probably the one profile photo that you use. Maybe you haven't come and turned that into a profile machine, but it really did help you engage with your community of friends and you were all learning about this technology together. Yeah, it was the next Instagram filter you could say. Now what do you expect in the next months? Let's say the next six or not even go further. How will the AI startup scene change? Yeah, you know, it's wild if I'm just really honest. This is going to be the year of generative video. And as an insider, I get a glimpse into the future where other people might not. And I've seen some of the most incredible multimodal and generative models specifically in media spaces that have been traditionally very difficult. Early examples of generative video were a little bit clumsy because they operated like a flip book. So you can imagine I generated picture, another picture, another picture, another picture, and then I kind of flipped between them. And the result can be a little bit incoherent. Future models do things like use neural radiance fields and be able to project our long light rays or even skeletal models and use key frame animations to go through and to create like a sense of physics attached to the generative model. The end result is that you have a much more coherent image and it can work in a really interesting way. Originally, I think my thesis was going to be like, oh, maybe it's going to be scientific discovery. Like we're going to see the next Nobel laureate come from some generative technology. But in practice, the expensive nature of building some of these models prefers things that have the ability to achieve financial success quickly. We did see over the last three years going from zero drugs and clinical trials that were discovered by AI to 12 the next year. So we are seeing tremendous achievements. And if you know anything about clinical trials, it's just a monstrous amount of work to get a drug through clinical trial. The reality is that we're going to see it coming from many different places, but there are some where there's just the potential to really engage with people and create those big opportunities, both financially, but also from a mind-blowing technology standpoint, get a little bit preferred. And that's where we see generative video in communities just continuing to be an absolutely massive part of the next six months. Okay. I guess with all the video stuff that's also very exciting since you've worked in this space in the past. I did. I did. I started my career working on the editor's side of the world. Working on the editors from motion pictures. Yeah. Okay. Now let's say I'm building a startup. What advice could you give me? There's so many. Just let's say two. Well, my own experience is that there's a lot of different ways to begin your startup journey. And there's not necessarily a right way to do so. You know, some people come as industry experts. They use those industry connections to go and they start their business. But I really liked the advice from Y Combinator and I did do two Y Combinator startups. So I think you want to build something people love. You really want to understand that like your expertise in a domain, like that you want to bring that to the forefront, really bring people's understanding that's there. And don't go alone. And don't underestimate the importance of diversity in that equation. One way that startup founders go a little bit awry is they kind of keep to only their immediate network of people and they wait too long to integrate other ideas. So have an idea you think people will love. Have a diverse team of people who are experts to engage with. Bring in their networks and really pull from those networks as much as you can. And then where possible, use the best AI technologies to save you a ton of effort and time. I would not build software today the same way that I built software five years ago. I would 100% use generative AI technology to help with things like code exploration, simplified customer support cases like categorizing tickets of things that come in. Really invest in those technologies and you could probably move a lot faster than I ever did. Okay. Well, Rob, thank you very much for that advice and for all the other information you shared with us. Looking forward to know more about what is all going to happen in the startup scene. So thank you and hope to speak to you another time. Thanks so much.