Blending Research and Innovation in AI

Sheikh Shuvo: Thanks so much for joining us this morning, Egor. We're really excited to dive in and learn about you and your background.

Egor Pushkin: Absolutely. Sheikh, it's a pleasure being here.

Sheikh Shuvo: So the very first question I have for you, Egor, is one I like to kick off with all of my guests. How would you describe your work to a five-year-old?

Egor Pushkin: Yeah, I have a seven-year-old and a 10-year-old, and every once in a while, they explain new concepts that I never knew existed. So, let me give it a shot. My work is around making it so that we can interact with computer systems and talk to them just like we are talking right now. So, instead of figuring out what button to press and how to navigate through an application, like when you're playing on an iPad or doing your homework, you could just tell the computer what you want, and it would respond and do what you're aiming to accomplish. For example, you can just say, "Hey, computer, buy me a board," and then your dad will spend hours explaining to technical support that it was a mistake.

Sheikh Shuvo: It sounds like that might be from past experience.

Egor Pushkin: Not far from that.

Sheikh Shuvo: I still remember when I first introduced Alexa to my house. I have a six-year-old and an eight-year-old. They found some audiobooks they loved and figured out how to buy them. That just changed my total experience of it. Tell us about your career story. What were some of the inflection points along the way that led to where you are now?

Egor Pushkin: I started in academia a long time ago. I was working on processing multispectral satellite images with early neural networks, pre-convolutional neural networks. Back in the day, one thing that we realized was the open source community, virtually speaking, was nonexistent. Libraries were in the dark age of neural networks. You couldn't just pick up a TensorFlow library off the shelf. So one thing that I started working on is the project that you noticed, the Neural Laboratory, which was essentially a visual environment for building and constructing neural networks. With the SDK, you could use it to export the final product. While doing that, I realized that I had a very deep passion for structural software engineering, which was quite different from the research work I was doing back then as my primary occupation. So I made the decision to pivot to the industry. I found myself joining a local real-time location sharing startup here in Seattle, Vlims. I joined shortly after the founders, Microsoft veterans Brian Chasso and Steve Miller. We sort of started this journey from scratch. We built a company from the ground up. That was an interesting and powerful experience that I went through. Picture our technology, real-time location sharing, being integrated across literally all pieces of modern consumer tech: smartphones, wearables, modern automotive dashboards, pretty much anything that had GPS and networking ability. That gave me a very broad spectrum of technologies to work with and exciting companies to interact with. We visited research and development centers of Mercedes, BMW, went to Samsung. So from that perspective, the breadth of applying this technology was vast.

Egor Pushkin: I could not possibly imagine a better place to be during that consumer revolution. At that point, a lot of pieces of consumer tech were revolutionized. We are not dealing with legacy automotive dashboards anymore; it's a giant screen in almost every vehicle right now. And, in that gig, there was one turning point worth mentioning: transitioning from being a purely consumer company to an enterprise play. That was very interesting from the perspective of learning as well. It's a significant shift in the paradigm and the mentality of how you build software, how you make it available to customers.

So, at that point, Williamson was focusing on this last-mile experience in delivery and any technician showing up at your door, giving you exactly the time when they would arrive. That was a phenomenal learning exercise. After spending about 10 years with that place, I realized it was time for me to review my career, to review what I was doing.

I looked at the overall journey, where I started, and what was happening in the industry and academia around me, and noticed there was something quite significant happening in the world of NLP. It was the end of 2017, early 2018, shortly after the fundamental papers on transformers were published, and I realized there was something in there. So, I had an interaction with one of the leaders at AWS, Vladimir Zhukov, who was heading a group of services dealing with all things language. They invited me to be their technical counterpart. So, I joined AWS in 2018 and spent a couple of years diving very deep into all things MLP and all things human language. Shortly after that, I got reached out to by Oracle. One of the leaders of Oracle Digital Assistant invited me to join the team, with the perspective of having a bigger scope, being able to apply myself to all things machine learning, the whole data platform, and the data science platform at Oracle Cloud. So here I am, in this group we have a portfolio of big data, managed open source data services, data science platform, and a group of vertical services covering all modalities of data: language, vision, speech, document, predictive services, and recently announced, generative AI service.

Sheikh Shuvo: Well, that's so much to dig into there. That's a really fun evolution. Going back to your earliest experiences that you mentioned when you were in Belarus, working on neural nets almost 20 years ago, way before it was cool. Looking back at those experiences as you were designing and training neural nets then, what's something that you didn't think would be possible that now, with the field having changed so much, is a reality?

Egor Pushkin: I'll briefly start with the opposite and then I'll get back to your question. What was certainly at the top of my mind, and something that I was able to predict, is neural nets growing in size. We already saw that, even during the dark ages of neural networks, ever since their inception. Those were structured as layered formations, so people kept on adding layers, and as technology progressed, as computational power grew, it was pretty easy to picture that they would be growing bigger and bigger in size. Back to your question, what I certainly did not predict is generativity being achieved so quickly. And that's something that happened over the course of the last couple of years, really. Up until that point, we were working on purely task-specific machine learning, and this paradigm shift to zero-shot learning and few-shot learning, currently powered by large language models, is quite significant. And that's effectively our journey to which, at this point, is not too far ahead, I would say.

Sheikh Shuvo: During your time at Glimpse, you mentioned working on so many different things, and really growing the company there with your work on mobile, automotive, and wearables. What do you think the influence of that has been as you develop systems now?

Egor Pushkin: Actually, there was one learning that helped me throughout my career while building Glimpse, and especially when focusing on our consumer play. There was a quite significant realization: when you put up a product, customers interact with it in a quite different manner than how you picture it. Effectively, when building it, you're an expert in the UX, in the flows, in how you expect people to get what they want out of it. And most end users, our customers, walk into this just as users, without intrinsic understanding of how the product works. So this obsession of being able to drive the customer through the journey that your product offers effectively drives the way I think about design, not just on the technical side of things, but also how that is expressed in the form of UX and in the form of some product concepts. These days, the tie to machine learning is that conversational AI and the newly introduced paradigms allow us to lower this barrier of entry. And that gets back to something that I mentioned in the very beginning, almost literally allowing customers to just walk in and say, "Hey, I want to do that." And the product would adjust, and the product would navigate the customers through, at times, very complicated and convoluted UX, which is usually the case for enterprise software.

Sheikh Shuvo: So, it's almost about giving up control and empowering the customer then. Looking at it from that perspective, and seeing how the technology is being used in the wild, across your time at many different companies, it seems like you're developing new things and helping customers figure out how to use them, either in your time at Amazon or now at Oracle. What have been some of the barriers when companies decide to go from a proof of concept to actually launching it in production? What are some of the technical and maybe cultural changes that need to occur when you're working with AI systems at scale?

Egor Pushkin: All the time. At a very high level, the story sort of diverges depending on the type of customer you're dealing with. In a lot of cases, when enterprises get into the topic of building machine learning systems from the ground up, they set up teams, they use data science platforms and hardware offered by the cloud, and they inevitably become experts in the process. They set up the life cycles and data acquisition processes. Over time, they build this expertise. So, as it's perceived externally, their challenges are mostly infrastructural and technical in nature.

The second category, which I would say is the most important and the one I've dealt with most in my past experience, is enterprises adopting managed services and vertically integrated services. In that case, they come in as product experts, experts in technology, but not necessarily in machine learning. So for them, the challenge of setting up machine learning operations becomes the first one. They interact with a service that changes under the covers so that they make a call, and in a couple of weeks, a new model is released, they make a call and get a slightly different result, which is the nature of probabilistic software. They naturally go through this paradigm shift in understanding how to build an application against a service like that, how to evaluate it, how to test it, how to make sure that their tests pass reliably. So it's all things machine learning, I would say. And the second one is evaluation and data. Generally, there is a big challenge, even for experts, for companies that are building machine learning software. Setting up the data pipeline and the data acquisition process is a very significant undertaking for a company that does not have it in its DNA and just getting into machine learning. It's a pretty big step.

Sheikh Shuvo: Is that something that's solved by adding more types of talent to the customer team? Or is it more about just retraining the existing team on how to work with new things? Are there any best practices that you've seen in the deployments you've made?

Egor Pushkin: Companies approach it differently, but in terms of best practices, they will almost certainly and unavoidably introduce new disciplines and employees working with data. One thing that is working really well is having those groups driven by domain experts. So, when you're preparing the data, irrespective of what measure of quality you actually choose for your particular task, just in general, the quality of your approach to training, evaluation, and testing data will determine the success of your initiative. So having domain experts and practitioners being deeply involved in setting up those processes makes a very significant impact.

Sheikh Shuvo: That makes sense. Shifting to what you're up to now, could you share some of the projects you're working on at Oracle? Maybe a particular project you're excited about?

Egor Pushkin: One of the biggest areas of focus for me right now is AI technologies. We recently launched our first AI service at Oracle Cloud. Now the work goes in a number of directions. We are growing the functionality of the service, with a number of launches already lined up. In parallel, we're working on adopting it across internal customers within Oracle, and external enterprise customers working with us. So that's generally a very exciting space to be in, with LLMs and the power of generative AI tech. What makes Oracle particularly interesting is, picture a company with data and enterprise software in its DNA. Now, this place that is already highly innovative and moves hand in hand with technology advancements, picture that place adopting generative AI. Making generative AI a part of its DNA, alongside data and enterprise expertise. I see a lot of potential in that across all domains, across our Fusion applications, NetSuite, and our new healthcare play that we joined just a couple of years ago.

Sheikh Shuvo: You mentioned a lot of work with internal customers there. For many internal teams, has there been resistance against using generative AI in their own workflows?

Egor Pushkin: It's actually happening the other way around. There is an incentive to adopt the technologies. For the most part, teams approach it from the perspective of them seeing the opportunity and knowing where the technology can be integrated into their service. We're getting reach-outs from internal teams, helping them to build the features they need, and helping them to go through the initial hurdles of evaluating and adopting, and getting that functionality to enterprise-grade accuracy level. There's a lot of incoming requests that we're seeing right now from across the company, across a wide range of applications. There's a deep desire to actually make it a part of the experience. So, there is a lot of consensus and a lot of potential in the application of this technology.

Sheikh Shuvo: Got it. So it sounds like you have the keys to the toy store and there's a line out the door.

Egor Pushkin: That's a perfect analogy.

Sheikh Shuvo: With generative AI technology evolving so quickly right now, and you having launched Oracle's generative AI services, what was your process to determine the right features and workflows to actually build?

Egor Pushkin: So, it's very similar to what I was talking about. We are lucky enough at Oracle to have this wide range of internal applications, and they gave a very significant boost in terms of initial definitional, then functional expansion. The second wave comes from our enterprise customers. The way we are looking at it internally, the way I look at it, is generative AI technologies and LLMs, to some extent, I almost look at it as an implementation detail. What's very important is to distill out of it the introduction of new paradigms, of interaction paradigms, that those technologies brought to the table, like new ways of conversational interactions, data retrieval, paradigms like REG. These fundamental shifts are here to stay as technology evolves. LLMs may become something else, but I don’t think those paradigms are going away. So in the end, the functional expansion is being looked at through the lens of the introduction of features that are seeing demand, that feed into those paradigms. And the vision in some cases, we got to believe in something. One of those paradigms that is close to my heart is a concept of agents that has been very actively explored in academia right now. It's based on the REACT model, reasoning, the next-in model that was enabled by LLM. It wasn’t possible before. After working on conversational systems, back at AWS, Lex, and the Oracle Digital Assistant, I believe that this paradigm is going to really revolutionize the way we interact with computers in general. So, it is just a part of the vision that we have.

Sheikh Shuvo: Speaking of agents in particular, I know one of the other major projects you worked on at Oracle was launching the clinical digital assistant. Could you share a bit more about that, and maybe what some of the unique challenges were in developing something with such sensitive information?

Egor Pushkin: The sensitivities are hitting the roof over there. Dealing with customer data in general is a pretty important topic. Dealing with healthcare data is taking that to the extreme. I'll start broad and we'll get back to the clinical digital assistant. A piece of my approach that worked really well for me in the past is, when starting a project, organizing the group, organizing the architectural efforts for it, I focus on some of the key elements that are essential, that are most important in a particular piece of software, particular components that we design. In there, it was very easy to identify that data is going to be the backbone of the system. So making the process driven around all things data allows you to integrate all those compliances, regulations, norms, as first-class citizen principles and not afterthoughts. Even though the space is quite challenging, the topic of building software in healthcare isn't new by any means. We were lucky enough to work as a part of the Cerner organization that is now part of Oracle, where there is an enormous amount of expertise on storing customer data as part of a certain millennium product, which is our HR system. Leveraging that expertise, we were able to build this shim, this digital assistant sitting on top of the data and making the life of healthcare professionals much smoother, much easier, eliminating the unnecessary documentation burden while keeping customer data and following all the practices, all of the GDPR, HIPAA, SOC compliance regulations, in full compliance with those.

Sheikh Shuvo: Throughout that building process, what was the testing and evaluation process like? Was there a certain benchmark you were trying to hit as you developed, or was it more about customer testing continuously throughout the lifecycle?

Egor Pushkin: It's a combination of both, really. Usually, the way it works before we put anything in the hands of customers, we define metrics, and depending on the tasks, the actual metrics would differ. It could be as benign as accuracy for some classification tasks. It's going to be something much more in-depth analyzing the output of large language models that we'll go through, looking at the actual accuracy, fact correctness, and other aspects of evaluation. So before anything goes out, even in a private beta, we already ensure that the software passes our thresholds across all metrics. Then we continuously improve it as it gets into the hands of limited preview adopters, and then a more general audience in a public preview, and finally, in generally available products.

Sheikh Shuvo: Along those lines, recently there's been a lot of attention on developing guidelines and recommendations from regulatory organizations, like the White House, around LLM technology to identify sources of bias or potential misinformation. I'm wondering, as you're designing those evaluation processes and metrics, what's been Oracle's, or at least your team's, approach to red teaming and ensuring responsible development?

Egor Pushkin: That's probably the most important one. Let me step back before we get to red teaming. In general, if you look at the process of what it takes to bring a large language model technology to production, it starts much earlier. It starts with the development of the training process, which is a multi-step process by nature. The importance of building a responsible AI system, the entire pipeline that all of those steps lead to, is very critical when you're dealing with enterprise software, especially healthcare software. At Oracle, we pay attention to every single step, every single phase in the development of LLMs. Hence the importance of our strategic partnership with Cohere, which is our primary provider of large language model technology at this point. We work very closely with a group of researchers and data scientists over there to ensure that the training data that goes into pre-training is clean, that we handcraft and build fine-tuning datasets for the alignment phases, for preference tuning, to eliminate as much undesirable behavior as we can. And, of course, it can’t be done just at the data level; the final product must be evaluated. At that point, we get to red teaming that you mentioned. On that end, we use a combination of manual effort, which is quite hard to set up and takes a lot of time, but the importance of that is critical. As the authors of the LLAMA 2 paper state, safety is a long-tail issue. It doesn’t matter if you pass your threshold when you achieve a certain number. The number's not going to be 100 at any point. So it's very important to perform that evaluation, even though it takes an enormous amount of resources in the end. It goes through round and round, producing checkpoints of the model that are better and better in terms of safety. And we've got the company that was automated for routine work, where one of the paradigms that LLMs brought to the table is interleaving and combining LLMs at development time.

Sheikh Shuvo: So there is one LLM that is generating the input for testing another one. And there are those techniques applied to evaluation that are applied to safety checking, which are quite computationally efficient and efficient in terms of human resources and the time it takes for you to go through those iterations.

Egor Pushkin: In practice, the answer is a combination of both. It gives us a lot of data to first analyze how the LLM is performing at a particular stage and then, post-analysis, this data is being fed into the next round of fine-tuning, making it safer as we go through those iterations.

Sheikh Shuvo: Oh, that's fascinating. Especially hearing how LLMs are integrated into the development and testing phases too. Taking a step more broadly, you started your career as a researcher. Looking at the research field now, are there any particular areas that you follow closely?

Egor Pushkin: Obviously, some of those that I already mentioned, like the agents and the chain of thought reasoning, are very close to my heart. I do believe in this paradigm, and I think there is a lot of potential in the topics of evaluating large language models. It's a very significant one. As I mentioned earlier, the success of virtually any machine learning product heavily depends on your ability to properly evaluate the system. The other area I would say is data generically. It's the backbone of the system. And, to some extent, I consider machine learning and the specific techniques used in machine learning a thin wrapper on top of data. As those techniques evolve, and I saw them evolve immensely over the last 20 years, the quality of data plays a bigger and bigger role in the development of those systems and the aspects of data lineage. How the concept of dealing with underlying data plays a significant role in the development of responsible AI systems. Hence, both my interest and my work on the papers that I've published on data lineage specifically.

Sheikh Shuvo: Is that the next paper that your team is publishing then, on data lineage?

Egor Pushkin: Well, there is always something in the pipeline. I prefer to mostly speak about the things that were already published.

Sheikh Shuvo: Cool. Okay. Awesome. The very last question I have for you then is, say I'm a recent college graduate and I'm looking to get involved in the AI space, but I don't know where to start. What general advice would you give for someone like me looking to figure out what type of company to work at?

Egor Pushkin: First of all, I'd start at a high level. In the machine learning space, even if you narrow down the technology to machine learning, there are a lot of disciplines involved. There is a huge technology play; the stack is very deep for the stack of technologies that are used in machine learning and machine learning applications. There is obviously science and applied science, and various levels of engagement and involvement in that space and discipline. And then there is a data play, with a lot of specialties and jobs organized around working with data, again in various capacities.

Egor Pushkin: So something that worked really well for me, and was effectively a highlight of my career so far, throughout the phases, during every single phase that I mentioned in the beginning, now looking back at it, I realize that I was working on something that I was very deeply passionate about. And at first, it was research, and then it was a very deep dive into technology, back in my startup days. And now, all things human language and machine learning in general, that I see an enormous amount of potential in. That gave a very significant boost to my ability to learn those things as I was transitioning across phases. Like, it was incredibly easy for me to dive into NLP. I saw the potential in NLP to get us to AGI, and it just doesn't get more exciting than that. And that drove me throughout these transitions, transformations, and eventually led me to be, I would say, quite successful in all of those journeys.

Sheikh Shuvo: Got it, so follow the excitement and what you get excited about?

Egor Pushkin: Essentially, yeah. Absolutely.

Sheikh Shuvo: Awesome. Well, Egor, thank you so much for sharing about your journey and lifting the veil on some exciting things that you're working on. I'm sure we'll be following you very closely to see what your team is coming up with next. Thank you.