 From theCUBE Studios in Palo Alto in Boston, connecting with thought leaders all around the world, this is a CUBE Conversation. Hey, welcome back, you're right here, Jeff Frick here with theCUBE. We are still getting through COVID. It's a hot August day here at San Francisco Bay Area. I think it's 99, somebody said in the city, that's hot. But we're still getting through it. We're still reaching out to the community. We're still talking to leaders in all the areas that we cover. And one of the really interesting areas is natural language processing. And it's a small kind of subset where we'll get into a little bit more detail with a very specific place within the applied AI world. And one of my very good friends and CUBE alumni who's really an expert in this space, he's coming back for his second startup in this space. And we're joined by Ben. He's been chung, the co-founder of Ogmagod. Did I get that right, Ben? Ogmagod. That's correct. That's right. Great to see you again. Thank you for inviting me to the show. Well, I love it. You know, one of the topics that we've been covering a lot, Ben, is applied AI. Cause, you know, there's just so much kind of conversation about artificial intelligence and machine learning. It's kind of this global big thing. And it kind of reminds me of kind of big data or cloud. In the generic, it's interesting, but it's really not that interesting because that's really not where it gets applied. Where I think what's much more interesting and why I wanted to have you back on is, no, where is it actually being applied in applications? Or where are we seeing it in solutions? And where is it actually changing people's lives, changing people's days, changing people's behavior? And you seem to have a propensity for this stuff. It was five years ago. I looked, July, five years ago we had you on and you had founded Genie, which was a natural processing company focused on scheduling. Successful exit sold that to Microsoft. I think they baked it into, who knows? They probably baked it all over the place. Left there, now you've done it again. So, before we get into it, what's so intriguing to you about natural language processing for all the different kind of opportunities that you might go after from an AI perspective? What is, what's special about this realm that keeps drawing you back? Yeah, sure. Yeah, I mean, to be honest, it was not anything pre-meditated. I kind of stumbled on it. Before this, I was more like an infrastructure guy, spent a number of years at VMware and had a blast there and learned a lot. Then I kind of just stumbled on it because when we started doing the startup, we didn't intend it to be a AI startup or anything like that. We just had a problem that my co-founder, Charles Lee, and I really wanted to solve, which is to help people solve people's scheduling problem. But very shortly after getting into and start looking at some of the use cases, we thought that the easiest way is to communicate with people like humans due to help them do the scheduling. And that's kind of how I stumbled on it. And it wasn't until that I stumbled on it that I realized that it has a lot of traction to me because throughout my whole life, I'm always very interested in the human emotions of it, how humans relate to each other. And that's always been a hidden side project thing. I do traveling to figure out those stuff and get a little bit of that. But once I started getting into this field, I realized that there's a lot about it, about humanity and how humans communicate, that it was kind of like a hidden interest for me that now suddenly coming out and it kind of just got me hooked. Right, that's awesome. So one of the things, and we'll just get into it, is people are a little bit familiar with natural language processing probably from Siri and from Google and from Alexa and increasingly some of these tools. But I think you kind of rapidly find out beyond what's the weather and play a song and tell me a joke that the functionality is relatively limited. So when people think about natural language and they have that as a reference point, how do you help them see that it's a lot more than asking Siri for the weather? Yeah, there are a lot of capability but also hopefully not offensive to some of the tech visionaries just as a guy who's dealing with it every day. There are also lots of limitation. It's not nearly to the degree of refinement like what might be in preach out there saying that the machines are gonna take over everything in one day. We have a lot of struggles, there's very basic stuff with machines. However, there has been definitely a lot of breakthrough in the last few years and that's why I'm dedicating my life and my time into this area because I think that it just, there's gonna be huge amount of innovation continuously going in this area. So that's at the high level but if you talk about in terms of artificial intelligence and in general, I think I have my own understanding. I'm more like an applied guy rather than academic. So what I'm gonna say might make some academic screen because I'm more like an everyday practical guy and try to conciliate these concepts myself. The way that I view is that artificial intelligence is really try to help mimic some human capabilities that originally thought that it's the domain of human. Only humans are able to do it but machines now try to demonstrate that machine can do it as though the humans could. So, and then usually people get that mixed up with machine learning. To me, it's actually quite different thing. Artificial intelligence is just like what I mentioned. Machine learning is just a technique or a science or way of applying like to leverage this capability, machine learning capability in solving these artificial intelligent problems to make it more achievable, to raise the bar on it. So I don't think we should use them into changeably artificial intelligence and machine learning because today machine learning is the big deal that are making the progress first. Tomorrow it might be something else to help improve the artificial intelligence. And in the past it was something else before machine learning. So it's a progression and machine learning is the very powerful and popular technique right now to being used. Now within artificial intelligence, I think you mentioned that there are various different domains and topics. There is like object recognition, use of image processing, there's speech detection, there's video and video, what I would call action or situation detection. And then there's natural language processing which is the domain that I mean that is really in that stage of where we seeing quite a bit of breakthrough but it's not quite there yet. Whereas versus speech detection and image processing actually has done a tremendous progress in the past. So and you could say that like the innovation there is not as obvious or as leapfrogging as the natural language processing. Right. So some of the other examples that we know about that are shared often from machine learning are say the visual thing. Can you identify a Chihuahua from the blueberry muffin which sounds kind of funny until you see the pictures they actually look very, very similar. And they always stated that Google in their Google photos has so many pictures, such a huge and diverse data set in which to train the machines to identify a Chihuahua versus a blueberry muffin or you take the case in Tesla if you've watched any of their autonomous vehicle stuff and their computer vision process and they have the fleet, you know, hundreds of thousands of cars that are recording across many, many cameras are reporting back every night with natural language processing you don't have that kind of a data set. So when you think about training the machine to the way that I speak which is different than the way you speak and the little nuances even if we're trying to say the same thing I would imagine that the variety in the data set is so much higher and the quantity of the data set is so much lower that that's got to be a kind of special machine learning challenge. Yes, it is. I think if people say that there is we are the cusp of being able to understand language in general, I don't believe that we are very far away from that. And even if when you narrow scope to say like focus on one single language like English even within that, we're still very far from it. So I think the reality at least for me speaking from ground level kind of person try to make use of these capabilities is that you really have to narrow it to a very narrow domain to focus on and bound it. And my previous startup is really that, you know that our assistant to help you schedule meetings that assistant doesn't understand anything else other than scheduling. We only able to train it to really focus on doing scheduling. If you try to ask it about a joke or ask anything else, it wouldn't be able to understand that. So I think the reality on the ground at least from what I see of a practical application and being successful at it you really need to have a very narrow domain in which you apply these capabilities. And then in terms of technology being used broadly in natural language processing, in my view, there are two parts of it. One is the input which people sometimes call natural language understanding. And then that part has actually very good progress. And then the other part is the natural language generation meaning that the machine knows how to compose sentences and generate it back to you. That is still very, very early days. So there is that break up. And then if you go further, I don't want to bore Jeff here with all these different nuances but when you look at natural language understanding there are a lot of areas like what we call topic extraction or entity extraction, event extraction. So to extract the right things and understand those things from the sentences there is sentimental analysis knowing that whether a sentence expressed is somebody's angry or some different kinds of emotions. There is summarization meaning that I can take sets of texts or paragraphs of texts and summarize is with fewer words for you. So, and then there is like dialogue management which manages the dialogue for the person. So there are like these various different fields within it. So the deeper you look, there is like more stuff within it and there's more challenges. So it's not like a blanket statement say like, hey, we can conquer on this. And if you're digging deep, there's some good progress in certain this area but some areas like, you know, it's really just getting started. Right. Well, we talked about in getting ready for this call and kind of reviewing some of the high level concepts of and you brought up, you know, what is the vocab? So first you have to just learn what is the vocabulary which a lot of people probably think it stops there but really then what is the meaning of the vocabulary but even more important is the intent, right? Which is all driven by the context. And so, you know, the complexity beyond vocabulary is super high and extremely nuanced. So how do you start to approach algorithmic, algorithmically, excuse me, to start to call out these things like intent or I mean people talk about sentiment all the time that's kind of an old marketing thing but when you're talking about specific detail to drive a conversation and you're also by the way converting it back and forth between voice and text to run the algorithms in a text-based system I assume inside the computer not a voice system. How do you start to identify and programmatically define intent and context? Yeah, just to share a little anecdote, like one of the most interesting part of since I started this journey six years ago and also interesting also very frustrating part is that especially when I was doing the scheduling system is that how sloppy people are with their communication and how little that they say they communicate to you and expect you to understand. And when we were doing the scheduling assistant we constantly challenged by somebody telling us certain things and we look at it as like well, what do they mean exactly? For example, like one of the simple thing that we use to talk a lot with new people coming on the team about is that when people say they want to schedule it next week they don't necessarily mean next week. What they mean is not this week. So it doesn't, if you like take it literally and you say, oh, sorry, Jeff, there is no time available next week and actually Jeff probably not even remember that he told you to schedule it next week. What he remembered was he told you not to schedule it this week. So when you come back to them and say, Jeff, you have nothing available this week next week and Jeff might say like, well, your assistant is kind of dumb. I like, why are you asking me this question? If there's nothing available next week just schedule it the week after next week. But the problem is that you literally said next week. So if we took you literally, you know, we would cause unhappiness for you but we kind of have to guess like what exactly you mean. So they're like, this is a good example where they're like a lot of sloppiness and a lot of contextual things that we have to take into account when we communicate with humans or when we try to understand what they say. So yeah, it's exactly your point. It's not like mathematics, it's not simple logic. There are a lot of things to it. So the way that I look at it, there are really two parts of it. There's the science part and then there's the art part to it. The science part is like what people normally talk about and I mentioned earlier, you have to narrow your domain to a very narrow domain because you don't have the luxury of collecting infinite data set like Google does. You as a startup or any team within a corporation you cannot expect to have that kind of data set that Google or Microsoft or Facebook has. So without that huge data set, so you want to deliver something with a smaller data set so you have to narrow your domain. So that's one of the science part. The other part is I think people talk about all the time to be very disciplined about data collection and creating training data sets so that you have a very clean and good training data set. So these two are very important on the science part and that's expected. But I think a lot of people don't realize is what I would call the art part of it. It's really, there are two parts to that. One is exactly like what you said, Jeff, is to narrow your domain or make some assumption within the domain so that you can make some guesses about the context because the user is not giving you to you verbally or giving you to you in the text. A lot of us, we find out visually by looking at the person as we communicate with them or even harder, we have some kind of empathetic understanding or situation understanding, meaning that there is some knowledge that we know that Jeff is in this situation. Therefore I understand what he's saying right now means this or that Jeff is a tech guy like me. Therefore he's saying certain things. I have the empathetic understanding that he meant this as a tech guy. So that's a really art kind of part of it to capture or make some good guesses about the context. So that's one part. The other part is that you can only guess so much. So you have to really design the user experience. You have to be very careful how you design the user experience to hide what you don't know so that it's not frustrating to the user or to put car rails in place such that the user doesn't go out of bound and start going to the place where you are not trained for that you don't have to understand it. Right, right. Because it's so interesting because we talked about that before that so much of communication, it's not hard to know that communication is really hard, email's horrible. We have a hard time as humans unless we're looking at each other and pick up all these non-verbal cues that add additional context. And am I being heard? Am I being understood? Does this person seem to understand what I'm trying to say? Is it not getting in? I mean, there's so many of these kind of non-verbal cues as you've expressed that really support the communication of ideas beyond simply the words in which we speak. So, and then the other thing you got to worry about too is you said ultimately it's user experience. If the user experience sucks, for instance, if you're just super slow because you're not ready to make some guesses on context and it just takes for a long time. People are not going to use the thing. So I'm curious on the presentation of the results, right? Lots of different ways that that can happen. Lots of different ways to screw it up. But how do you do it in such a way that it's actually adding value to some specific task or job? And maybe this is a good segue to talk about what you're doing now at Ugman God. I'm sorry, I have to look again. I haven't memorized that one yet. So how do you, I think you, because what you're also doing if I recall is you're taking now an additional group of data and additional data sets in beyond simply this conversational flow. But ultimately you've got to suck it in. As you said, you've got to do the analysis on it. But at the end of the day, it's really about effective presentation of that data in a way that people can do something with it. So tell us a little bit about what you're doing now beyond scheduling in the old days. Sure, yeah, I left Microsoft late last year and started a new startup. It's called Ugman God. And what we do is to help salespeople to be more effective, understand the customer better so that they have higher probability of winning the deal or to be able to shorten the sales cycle. And oftentimes a lot of the sales cycle got lengthened is because of the lack of understanding. And there's also as a, we focus on B2B sales. So for B2B salespeople, the world has really changed around a lot since internet came about. In the old days, it's really about tell it to explain what your product is and so that your customer understand your product. But the new days is really about not explaining your product because the customer can find out everything about your product by looking at your website or maybe your marketing people did such a good job that already communicated to the customer exactly what your product does. But you really to win out against other people, you really like almost like a consultant to go to your customer and say like, I have done your job almost like I've done your job before I know about your company and let me try to help you to fix this problem and our product fit in as part of that but our focus is let's fix this problem. So how would you be able to talk like that? Like you've done this job before like you worked at this company before how do you get at the level of information that you can present yourself that way to the customer and differentiate yourself against all the other people who are trying to get their attention all the other people sending them email every day automatically. How do you differentiate that? So we felt that the way that you do it is really have the depth of understanding with your customer that is unrivaled by anybody else. Now, sure you can do that you can Google your customer or they read all the news report know all the leadership follow them on social media. Right, they're supposed to be doing all the stuff right then they're supposed to be doing all this stuff. And with Google and the internet there's no excuse anymore. It's like how did you not do your homework? You used to have to get the yellow things. Yeah, why did you do your homework? Yeah, some people get beat up by their management saying like, oh, how come you missed this? It's right there, go on Google. But the truth is that you have to be empathetic to a salesperson. A lot of people don't realize that for a salesperson every salesperson you might own 300 accounts in your territory. And a lot of times in terms of companies there might be thousands of companies in your territory. Are you going to spend seven hours follow all these 300 companies and read all the tweets check out the thousands of employees in each of these companies that are linking profiles look at their job listings look at all the news articles. It's impossible to do as a human, as a person. If you do that, you'll be sitting in your computer all day and you will get in the door to have a conversation with the customer. So that is the challenge. I felt like salespeople really put up the impossible task. Because all this information out there you expected to know. And if you screwed up because you didn't check then it's your fault. But then on the same time how can they check all 300 accounts and be on top of everything? So what we thought is that like, hey, we made a lot of progress on natural language processing and natural language understanding and salespeople what they look for it's a quite narrow domain. They're looking for some very specific things related to what they're selling and very specific projects, pain points, budgets related to what they're selling. So it's a very narrow domain. We felt like it's not super narrow. It's a little bit broader than I would say scheduling but it's still very narrow the kind of things that they're looking for. They're looking for those buying triggers. So they're looking for problem statements within the customers that relate to what they're selling. So we think that we can use, develop a bunch of machine learning models and use what's available in terms of the web, what's out there on the web the type of information out there. And to be able to say like salesperson you don't need to go and keep up and scan all the tweets and all the news and everything else for these 300 companies that you cover we'll scan all of them. We will put them into our machine learning pipeline and filter out all the junk because there are lots of junk out there. Like Nike does like, I don't know hundreds of news release probably per week and most of them are not relevant to you. It doesn't make sense for you to read all of those. So, but how about we read all of them? And we extract out, you know we, this is called topic extraction we extract out the topic that you're looking for and then we organize it and present to you. Not just we extracting out the topic once we get the topic how about we look up all the people that are related to the topic in the company for you so that you can call on them. So, you know what you want to talk to them about which is this topic or this big point and you know who to talk to these other people. So, that's net net what we do. That's, you know, that's really interesting. It's been a tagline around here for a long time, right? Separating the signal from the noise. And I think what you have identified, right is as you said, now we live in the age where all the information is out there. In fact, there's too much information. So, you know, you should be able to find what you're looking for to your point there's too much. So, how do you find the filter? How do you find the trusted kind of conduit for information so that you're not just simply overwhelmed that what you're talking about if I hear you right is you're actually querying publicly available data for particular types of, I imagine phrases, keywords, sentences, digital transformation initiative, blah, blah, blah. And then basically then coalescing the ecosystem around that particular data point. And then how do you then present that back to the salesperson who's trying to figure out what he's going to work on today? B2B salespeople, they start with an opportunity. So, the opportunity is actually a very concrete word at least in the tech, in the tech B2B sales. We know, we see the 60 stories in downtown San Francisco will validate that state. Yes, so it starts with the word opportunity. So, the output is a set of potential opportunities. So, it speaks to the salesperson's language and say, when you use us, we don't just say, hey, Jeff, there's this news article about trurios, that's interesting to you. Oh, there's a guy that, you know, a trurio that matches what you, the big kind of persona that you sell it to, we don't start with that. We start with Jeff, there are six opportunities for you. Opportunities for you at trurio, let's explain what those things are. And then let's explain to people behind these opportunities so that you can start qualifying them. So, get you started right away in your vocabulary in a package that you understand. So, I think that's what differentiates us. Right, right. And at some point in time, would you potentially just think in logically down the road, you know, have some type of Salesforce API, so it just pumps into whatever their existing system is, that they're working every day. And then it describes, based on the algorithm, why the system identified this opportunity, what the attributes are that flagged it, who are the right people, et cetera. Awesome, so what kind of data sources are you querying? You're designing all product wise. Yeah, as soon as Dave and John watch this, they're going to want to talk to you, I'm sure. But what type of data sets are you querying? There are lots of them. We learned most of it by through the process of working for salespeople, meaning that we work for salespeople, we maybe, you know, quote unquote, stand behind their back and see what they're searching. They're searching LinkedIn, they're searching jobs, they're searching analysts, you know, call transcripts. They're looking at 10K, 10Qs. They dig up various, some people are very, very creative, digging up various parts of the web and find really good information. The challenge is that they can't do this to scale. They can't do it for 300 accounts, because for doing it for one account is very laborious. So there are various different places that we can find information. And it's not, in terms of the pattern that we're looking for, it's not just keywords, it's really concepts, we call it a topic. We're really looking for very specific topics that the salesperson looks for. And that's not just a word, because sometimes words is very misleading. For example, I tell you, one of the common word in tech is called Jenkins. Jenkins is a very popular technology, continuous delivery technology step. But Jenkins also happens to be a very common last name for people. Well, I'm always reminded of our Intel days with all the acronyms, but my favorite is ASP, because you could use ASP twice in the same sentence and mean two different things, right? Average selling price or application service provider back in the days before, we call them clouds. But the nuance is so tricky. So within kind of what you're doing then, and as you described, working within defined data sets and keeping the UX and user experience pretty dialed in and within the Rails, are there particular types of opportunities in terms of B2B types of opportunities that fit better, that have kind of a richer data set, a higher efficacy in the returns? What are you kind of seeing in terms of, great opportunities for you guys? We're still early, so I can't tell you that from a global view, because we are like less than one year old experience quite honestly. But so far, we are being led by the customer. So meaning that there is an interesting customer, they ask us to look for certain topics or certain things. And we always find it to my surprise, because, and that really is a, like I'm constantly surprised by how much is there out in the web, like what you were saying. Customer ask us to look for some things and I thought for sure, this thing, we couldn't do it, we can't find it. And we gave it a try and low and behold, there it is, it's out there. So to be honest, I can't tell you at this point because I have not run into any limits. But that is because we are still a very young startup and we are not like Google, we're not trying to be all encompassing, looking for everything and looking over everything. We're just looking over everything that a salesperson want, that's it. So I'm going to make you jump up a couple levels such that you've been thinking about this and working on this for a long time. There's a lot of conversation about machines taking everybody's jobs. Then there's the whole side tranche to that, which is no, it's all about helping people do better jobs and helping people do more higher value work and less drudgery. I mean, that sounds so consistent with what you're talking about. I wonder as somebody down in the weeds of artificial intelligence, if you can kind of tell us your vision of how this is going to unfold over the next several years, is it just going to be many, many, many little applications that slowly before we know it are going to have moved along many fronts very far? Or do you still see, it's such a fundamental human thing in terms of the communication that these machines will get better at learning but ultimately they can kind of fulfill this promise of taking care of the drudgery and freeing up the people to make what are actually much harder decisions from a computer's point of view than maybe the things that we think about that a three-year-old could ascertain with very little extra effort. Yeah, if you take a look at what we do and hopefully didn't sound like with underselling our startup but a lot of it really is we're taking away the time consumer and also grant work process of the data collection and cleaning up the data. The humans, the real human intelligence should be focused on data analysis, to be able to derive lots of insights out of the data. So, and to be able to formulate a strategy how to win the account, how to win the deal. That's what the human intelligence should be focused on. The other part like struggling with doing the Google search and you return 300 entries with 30 different pages and you have to click through each one and then give up after the first three. That kind of data collection, data hunting work, we are really should not, I don't think it's worthy quite honestly for a very educated person to deal with. And we can invest it back in helping the human to do what the humans are really good at is that how do I talk to Jeff and I'm going to get a deal out of Jeff? How can I help and do helping him solving his problem? How can I take the burden of solving the problem from Jeff's head and solve the problem for him? That's what human intelligence, for me as a salesperson, I would prefer to do that instead of sitting in front of my desk and doing Googling. Right, right. So, so what, so then that what I'm trying to say using ourselves as an example is that we're not taking over the job of a salesperson. There's no way that we can close the deal for you. But what we're doing is that we're empowering you so that you look like you're on top of 300 accounts and you talk to pick any of those accounts, you'll be able to talk to the people, your customer, that particular customer, like you know them inside out. Right. And without you being the super human to be able to do all this stuff. But as far as that customer is concerned it sounds like you were on top of all this stuff all day and that's all you do. You have no other customers. They're the only customer. In fact, you were on top of 300 customers. So that's kind of the value that we see to provide to the human is to allow you to scale by removing these grunt work that are preventing you from scaling or living up to your potential of how you wanted to present yourself, how you wanted to deliver yourself. Right. Well that's a great thing. There's no way that we can be smarter than the human in a way. I just don't see it. Not in my lifetime. I just love, you know, we've had a lot of conversations over the years and you talking about the difficulty in training, you know, the computers on some really nuanced kind of humane things versus the things that they're very, very good at. And you know, keeping the AI in the right guard is probably just as important as keeping, you know, the user interface in the right lane as well to make sure that it's a mutually beneficial exchange and one doesn't go off and completely miss the benefit to the other. Well, Ben, it's a great story, really exciting place to dedicate yourself and we are just digging, watching the story and we're going to enjoy watching this one unfold. So thanks for taking a few minutes and sharing your insight on natural language processing and the supplied machine learning techniques. Thank you, Jeff. It's always a pleasure. Yep. All right. He's Ben. I'm Jeff. You're watching theCUBE. Thanks for watching. We'll see you next time.