Suman Kanuganti is the Co-Founder and CEO at Personal.ai. Previous to Personal.ai, Suman also founded Aira. Suman holds his BE degree in Engineering, MS in Robotics, MBA in Entrepreneurship, and ten patents in emerging technologies as well. Suman also founded Aira. Personal AI is a GPT implementation designed to mimic an individual's behavior and to speak like them. In today's episode, Suman shares his goal is to create AI systems that understand and replicate users' communication patterns and cognitive abilities. Suman emphasizes that their language model has time awareness, allowing it to adapt its knowledge base depending on the user's age or point in their timeline. Suman also shares his some learning lessons for AI product creators. Find the full transcript at: https://www.aiproductcreators.com/ Where to find Suman Kanuganti: • LinkedIn: https://www.linkedin.com/in/kanugantisuman/ Where to find Dhaval: • LinkedIn: https://www.linkedin.com/in/dhavalbhatt Transcript:-

Dhaval: So today we have Suman Kanuganti on the show. Sunman is the co-founder of personal AI, a product that he will talk about and, we'll learn a lot more about where his learning lessons bear, what his learning lessons bear and all that other stuff. So yeah, let's get started. Suman, would you mind introducing yourself and tell us a little bit about your product, AI product?

Suman Kanuganti: Sure Dhaval. Thanks for having me. I'm Suman Kanuganti. My background is in engineering. Over 10 years ago I started creating companies. This is my second company. Previously I built a company called Aira, A-I-R-A. My philosophy always has been how do use technology kind, solve, hard human problems. Aira was about using technology to fill the gap of missing visual information for people who are visually challenged, such as blind and low vision. And personal AI is about augmenting people's mind where memory, cognition, and our time is limited. And we would want to augment that using technology by creating a personal language model of every individual that essentially learns to behave and act and learn to be you. So that's a little bit of my background and who I am.

Dhaval: Wonderful. Thank you, Suman. Thank you for that introduction. Tell us a little bit about what personal AI's narrative is like. What is the, who is your target customer? What kind of a problem, specific problem you are solving with that product? And what is the overall narrative? How is it different from your competitors?

Suman Kanuganti: Yeah, totally. As individual people, like on a day-to-day basis, we create and consume a lot of information. Like, you know, a lot of experiences have a lot of conversations. But obviously, 80% of that is lost or forgotten. Our goal with personal AI is to be able to create a model. That actually learns your knowledge, your style, and your voice, more or less like to be a digital version of you. Imagine being able to surface relevant pieces of your own knowledge, on demand whenever you need it. Or imagine, you having conversations in a, chat or text message. Are people talking to you? Where relevant pieces of information when we start facing as you speak. So our intention is to be able to augment, humans with an extension of their own mind. Because one cognition is limited and then two time is also limited. You mentioned about the target market. Our goal is to actually go after everyday consumers. Our intention is to have everybody have their own personal AI. That is trusted by them. The data belongs to them. The model gets trained over a period of time and innovate kind of grows alongside with you. Unlike, public or general models that exists such as open AI, such as, Google or Alexa, which is mostly like trained on public internet of data. Personal AI is a unique model that also uses similar architecture such as GPT, but actually trains on individual person's data. And it does so stylistically, relevantly authentically to replicate as you would. So we are trying to essentially like replicate, your thought process and your mind and give you an extension of your.

Dhaval: Wow. So it is adding the stylistic and tonality and the personal attributes to your, to the replica that you are building for someone which is not there in current GPT or any of those products, right? So there's a bias that. GPT has very confident answers, but doesn't necessarily align with your style or may not align with the way you communicate. And what you are saying is that not only does personal AI helps you do that, but also creates the model in the first place using your personal attributes , did I get that right Simon?

Suman Kanuganti: Yeah, totally. So you will create what we refer to as a memory stack, which is essentially taking all your unstructured data that you ever have in your digital world. Let's say you're having conversations online, you are texting with people. You probably have written a bunch of different knowledge pieces out there. And then we create this memory stack, which is essentially like a digital representation of your memory vault. In other words, we basically break down this idea of structuring your data into these blocks that is associated with time. And imagine over a period of time, as you create an as you learn. Your AI technically would also be training alongside with you. So it's kind of how. Conceptually we have architecture system.

Dhaval: Got it. Now, there are, if you were to dive a little bit into your products architecture or the core engine, the AI, since the audience of this show are the people who aspire to either create an AI Product. Or wanna add AI to their existing product, for their knowledge. If you were to share a little bit without getting into the confidential details about what does a product stack, what does it take to build memory stack? What does it take to build a knowledge replica, a knowledge brain summons, human brain, and human personality into something that you were doing, like what do you call that thing? The entity that.

Suman Kanuganti: Yeah, I'll try to provide answers and then I'll try to provide some contrast technologies that exist out there so that we can wrap our heads around. The first thing is at the core, we are essentially an AI first company. We built an algorithm called. Personal language model. So we call it personal language model. This is in contrast or like kind of opposite in concept to a large language model. If you think of a large language model such as GPT 3 or any other open Language models that exist out there close to around like one 70 billion parameters in our case. Our language models around one 40 million parameters. It revolves around at the core individual's data and not public's data. And you can keep on adding the data to your model so that way it gets more sophisticated in regards to the purposes of your mind. You can go abroad and you can also like, go deep into specific topics itself. So yeah, at the core we build this personal language model for every individual to essentially like mimic the behavior, knowledge, and style of an individual person. And the transformer that we have developed, we call it generative Grounded Transformer. And if you think about like GPT as a generative pre-train transformer, the subtle difference of our transformer is that it is grounded in the personal data of you. And whenever I refer to personal data, is nothing bad. The memory is tag that I was explaining earlier, So every AI response that your personal language model actually generates, it has an attribution, and the attribution is nothing but attribution Back to the data or what"s? Data elements of what memories in within your stack is responsible for creating a particular response? One of the challenges for large language models is that there is no attribution, primarily because it is driven by aggregation of the data. And there is quite extensive anonymization that is involved. So technically, you cannot create that attribution and it's extremely hard to create that attribution. And our goal is exactly the opposite. We would want to have. That attribution, we want to have the ownership and we want to create that value to every individual consumer by creating their own individual model, at the foundation, it's personal language model.

Dhaval: So this generated you, grounded trained model that you are referring to is that, Adaptive with human changes, human behavior changes over the lifespan of the user.

Suman Kanuganti: Exactly. The transformer also has a sense of time. For example, let's say, if you're talking about AI maybe three years ago, How you refer to your transformer, how you refer to your technology, maybe different from your latest and greatest creations or thought process around your ai. So it has a time to decay component. So when you are indeed chatting with your own AI, it normally anchors around the latest and the greatest thought process of. How you would respond. However, let's say if you indeed are contextually trying to fit something from the past that happened like 10 years ago Then, we are probably talking about the first autonomous car, right? And my experience with the first autonomous car, then it will. It'll go back in time and be able to fetch that response for you. So kind of like designed to work very similar or akin to how a human mind would function. You can think about potentially being able to drop your AI at a certain period in time. Given these are small models and given as the data is going in on a day-to-day basis, there is a new version of the model. We are technically able to time travel. Yo...