Hi everybody, my name is Jay Agner. Welcome to the First Customer podcast. We talked to founders about, you guessed it, how they got their first customer. This episode and every episode is sponsored by my company JD A QA Software testing where your one stop shop for manual automated performance and security testing check us out at jdaqa. com. And with that let's get on with the show. Hi everyone, Welcome to the first customer podcast. My name is Jay Agner.
Today I'm lucky enough to be doing but joined by Joe Roshar, which I clarified before we started here. Joe, welcome. No, thank you. Uh. So Joe, uh, you've got a very cool, uh, varied background like we were talking a little bit beforehand, Um. You've been doing AI, uh, since before it was cool, so I'm sure we'll kind of get to that at some point along the way. But tell me a little bit about your
background where you grew up. Grew up in Minneapolis and uh, I liked studying foreign languages when I was a kid. So even in high school I'd studied German and French and I I I went to Mexico and learned a little Spanish. And so from a very early time, natural language understanding was really important to me. And where did that come from? Did you have any pet in your family or anything or?I I don't know, but I know that when I traveled I liked
to be able to talk to people. And so in in my undergraduate degree, I got a degree in linguistics. And uh uh, While I was doing my masters work at UH University of Minnesota and also graduate research at Tokyo Institute of Technology, I studied Shengo JoJo Shorty. And you hear the word joho in there. That means UH, Natural language processing. Wow. How did you end up in in Japan? Uh. I was, uh, first there as a missionary for the Church of Jesus Christ of Lambda States, and later I
went as a soldier. I was an intelligence officer with the US Army. Wow. And you went back after that? I did. You loved it so much. I did. I loved it so much. And you good dude. You said you go back still. Occasionally. Very cool. Um. So talk to me a little bit about after college. Uh, you you kind of seemed like very early on, had an idea of where you wanted to go. But but where did you go after all those different degrees and and you know, the
the variations of all that stuff?Uh, after I got out of the army, I, uh went to a company called UM Secure Computing Technology Corporation, which later became Secure Computing and worked on bulk encryption devices and stuff for mostly government customers. OK. Um, yeah. So you started out very heavy technically, um, a lot of architect roles across a bunch of different companies.
Talk to me about kind of how you transition from, you know, those initial engagements as a consultant or a technical, you know, kind of programmer into the architect role. My first architect role was just prior to Y2K at American Express Financial Advisors, now Ameriprise. And um, I was, uh, helping them. Um, My data and programming background enabled me to, um, understand broadly the challenges that they were facing associated with Y2K, and that was
a great way for me to wet my feet. I also worked for United Health during that time, and I was doing architectural related work starting in the late 90s. And um, Uh and uh. Then I joined United Health as a senior architect for one of their divisions called Optum. When it was much smaller than it is today. And um, uh, it just uh blossomed from there. So did you always have the itch to found something or did that develop throughout your career?
I did not, uh, think of myself as an entrepreneur when I was younger. But um, I decided to give it a shot and I really liked it. And what was your first attempt at that? My first attempt was called Intuitive Dimensions, UH and um. I was already working on solving the Natural language understanding problem and I started to build out this. This. What later became my blog site, but what was originally just a a big.
Um, I I was using asymmetric tool book if you remember that to uh create just a large multi charter online document. And uh, and then I transitioned that into my blog understanding context. com. So what was the what was your goal with understanding context? I saw that when I was looking looking through things. Um, what was your what was your kind of overarching goal for that?The same as today. Uh, that is to build the universal translator. I saw C3 PEO on the big screen and I thought
we should be able to do that. And yet everything I saw in um in production was language pairs. One language to one other, no ability to go from one to many. And the reason for that is that they were using. Neural network style learning. Processes that were based on patterns and uh, when you're using raw patterns. You can only really do the left brain, and you can only really, uh, do very limited translation.
And uh, no matter how large the language model gets, the accuracy in translation from one to many is going to be limited until you incorporate concepts in context. Uh, by putting concepts in context and using meaning as the basis for understanding and translating language rather than patterns, suddenly you open up the possibility of a universal translator. Wow. Um. I I constantly you know the the chat LGBT rate you know is the all the rage right now and that's it still feels
like. Context is like the human element that it's missing. And that's right. And that's one of the reasons why there's been such a massive ecosystem of prompt vendors who are teaching people how to talk to it, right. And and the fundamental premise I began with was that we don't want to train the person to talk to the computer. We've been doing that for decades and it it still is clunky, right? We want to train the computer to understand what the human says, and that
resolves. That means resolving ambiguity. Um. Language has a thing called polysemy, multiple meanings, and uh, there are words like set in English that could have literally dozens of different interpretations. And um. In order to really understand your intent in in order to read your mind based on what you said. You need to be able to resolve ambiguity, and to be able to do that you need a meaning based strategy. But there's there are other advantages to using a meaning based strategy to
interact with a person. First of all, you get explain ability. Explain ability has been important in AI since the very beginning and neural networks don't give you that because the knowledge is encoded in this pattern of you know nodes, which is fundamentally not explainable. There's another issue associated with. Um, AI that doesn't use concepts and context, and that is the inability to use attribution.
So if you learn something from Encyclopedia Britannica, or if you learn something from reading Plato, you want to be able to say that. Because then if a person, for example, is using your interaction tool to help research a paper, they'll be able to put in a bibliography, and if not? They won't. And uh and uh, this is huge. Uh, especially in an era when fake news, bias and other deep fakes are becoming the norm. Hmm. Wow. I did not think I was gonna find somebody so well versed in.
The hot topics of today, so I'm glad that that we're talking. So you you made the blog site understanding context, Um, but now I understand the context of that blog. Talk to me about your next steps from there. Where did you go after that? And maybe Nikki was next. I'm not sure. Tell me. Tell me kind of where you went from there. Well, I formed the company into Dimensions and it failed, and so I formed a company called Epica Epica. Harks back to my enterprise architect
scribbles. It means Enterprise Process and Content Coordination Architecture. And my idea is that, um, uh, the Americans are all focused on data and the Europeans are all focused on process. And you really need both. And you really need both in order to be able to efficiently implement AI.
And so this whole UH process and content architecture model led me to form, well uh first after intuitive mentions was a company called Knowledge Genetica shortened Kinetica and I got a couple $1,000,000 of investment including a bunch of my own. And we were able to sell our product to a company called Blue Cross Blue Shield of Minnesota. That product was called I Speak Your Language and it included kind of the Swiss Army knife for. Uh, integrating
um. Enterprise applications and uh, specifically what it did was it enabled call center agents and Blue Cross to be able to have a single interface into, uh, 8-9 different systems. And with that single interface, they were able to much more effectively answer questions about claims and benefits and things that. An insurers call center does. Uh, this is the foundation for a hyper automation for AI.
Um, and so the workflow, the inference engine for forward and backward chaining inference and all of those things are an important part, along with the adapter factory for connecting to all of the different enterprise systems, including those containing content, unstructured information and data structured information. And the ability to do this is really important to be able to integrate AI, because even if you can engage in dialogue with the person. Understand their intent unless you can
integrate with. The exact same concepts in context, in the Enterprise. It's going to be very costly to bring business value. And that's what our whole point is, is that we, uh and and by the way. I've I've switched from talking about translation to talk about just interpretation. Hmm. It turns out that before you can really expect to be able to translate to many different languages, you have to know that you understand what a person's
talking about, right? And so this is the springboard that is to be able to understand a user's intent in a business. Be able to use external data and internal data to be able to help them achieve whatever business value they're trying to do, whether it's associated with marketing or finance or IT or what whatever part of the business they're in. If you can understand their intent and bring value to them by connecting to their data and content.
We know that we're able to understand and we can go to the next big step, which is the universal translator. Well, I do want to. So I think I'm just as interested and maybe a little bit painful, but why did that first business fail? It sounds like you had a lot of success in your subsequent businesses. Why did that one fail?I didn't hire the
right people. And and can you elaborate a little like what what were some of the pain points and how did you recognize that it was it just an after the fact that you realized that you hired the wrong people? Uh, I hired a person to run the company. I like being the geek. Uh, and I hired a person to run the company who decided he wanted to take it over and get rid of me. That that was my first hint. All right. Fair enough. Fair enough. Um. So you did the Blue Cross and Blue
Shield. You had a bunch of businesses with acronyms that I probably can't even pronounce or remember all the the the names in them. But um, keep walking me through the journey because it sounds like you just kind of kept building on the previous, you know, lessons that you learned as you kind of spawned off these new entities. Where where did you go next? Uh, after, um, the Blue Cross, and that was with knowledge genetic, I went to Epica. Right and. Uh, it
um I just. I I struggled for years to um. Overcome the challenges of starting up a new company and one of the things I did was I I insisted upon paying my debts and so the SBA loan and stuff. But I took out I'm. I insist on paying that and so that took me a while. But as a consulting enterprise architect I was able to earn enough money to feed myself and get the company back on its
feet. I. Uh at uh with the Packer I transitioned the software from being uh uh premise based um to software as a service cloud based and I learned all about Azure and AWS and Google Cloud Platform and. I learned a lot about UH performance, and obviously when you're doing UH heavyweight AI performance is a big issue. And uh uh, In my roles as Enterprise Architect, I was able
to. Um, learn a lot about those things and uh, uh, I have had additional UH sales and additional opportunities to expand the software and right now we've I've formed a company called Joho AI. Um, as I mentioned that stands for, that means UH, news or intelligence. UH, When I was an intelligence officer in Japan, they called my division Joe Hob or the the intelligence division. Uh, And so at any rate, um um.
All those steps have led me to a place where I'm really confident that this time we're going to be able to deliver the value that we have been talking about all along, right? So many questions. So talk to me a little bit about some of your first customers. I mean you've had you spawned off so many businesses Um
well we're let's let's cover it. I guess a little more generally speaking, what were some of the ways and I'm interested too because as a as a more technical and maybe you know less businessy, less salesy kind of guy. How did you acquire those first customers who were the, I mean, you don't have to name names, but who were those first customers generally speaking and then how did you kind of engage with them?It was all through networking, uh. Often through people that I knew very well. Um, uh.
So my first customer I met at an event and uh. My first two customers, in fact, uh, Blue Cross was my third customer and uh, they were introduced to me by one of my coworkers. I was in the Minnesota Army National Guard and my superior, Lieutenant Colonel Warren G Harrod, the 2nd. Uh, introduce me to the folks at Blue Cross. Um. Later um I, uh had a large contract with uh a sale to a company called Amplify Capital and once again it was networking. Um, uh, people? I met uh in the course of,
uh, putting myself out there. Um uh?Turned out to be, UH, the opportunities for customers. UH. It turns out that Dan Herbeck, who introduced me to Amplify UH, had also been working at United Health and I had consulted at United Health for a long time. So. That's been, uh, the way it's worked. It turns out that my current partners. Uh, my uh cofounders UH, Chris Higbee and Mike Lopez were introduced to me by my son, who's a chiropractor.
I always hear the, you know, I always, I like to refer to it as the friends and family plan. You know, there's always it's it's the network you build and I am curious. It sounds like you maybe had to work on that a little bit to step out of you know, being the the geek as you like to call yourself. And I mean I, you know, I used to run a text based multiplayer video game out of my basement. So like I can commiserate with you Being a geek. I have
a, you know, development background. So how did you, how did you learn to to network and step out of you know kind of the the traditional programmer, you know, withdrawn guy in the corner typing on the keyboard. Like how did you, how did you break out of that?You know, I am naturally a little more gregarious. Uh. And I I I love people and uh, so uh, I guess that wasn't as big an issue for me. Um, and and and I think there's one other crazy thing, and that is my undergraduate degree was in
linguistics, English, right?Um, talking is my thing. That's fair. Uh, that is fair. Um, so I I this question. Um. Is is something that, uh, I have varying answers to, but I'm really curious with you with all the kind of like the progression towards JoJo and kind of like now you've you know, feel like you've got the, the, the pieces to provide the, you know the, the answer to the question you've been asking the whole time, which is the universal kind of
translator. If you had to start over today with everything you've learned before, won't be step one. That's a really, really good question. Um. In my undergraduate studies, I learned a programming language called Snowball. And um, our teacher Wayne Tosh uh, asked us to build simple translation programs because Snowball is designed for string processing. And um. And. When I worked on my masters degree at University of Minnesota under. I'm in the UM. Computer Information Science Department.
My advisor was Doctor James Slagle, Amazing man, blind chess champion of the world. Uh and uh, he turned me on to Expert Systems. He had written one of the first. And um. The progression of doing it just worked out great for me. I started out thinking about process using the programming language I next Expert systems are based on models and the model in involves having a knowledge representation scheme that is going to be efficient and complete.
And that was my big challenge for the first 6-7 years of working on this developing a knowledge representation scheme. That would actually perform well enough to be able to do this on computers that. You know our digital. Um, so the process focus and then the content focus and and putting them together. Worked out to enable me to file my first patent and get it awarded. The patent was filed in 2002 and it was awarded in 2007. So um. I I don't think I would do things
differently. Um, working with UH in the University of Minnesota, I I studied in the Computer Information Science department, but I did a lot of my courses in the neuroscience department and in the department of Educational Psychology to learn about learning. And um, all of those things just. Helped a lot, right? No, those are all. That's all um, very insightful. Um. And kind of a follow up would be um.
If you had to start a new business tomorrow with with the knowledge that you've learned, and I have to be, you know, I have to be very specific because I think you're probably the most articulate linguist I've ever probably spoken to. So if if you had to start a new business tomorrow based on everything you've learned before, what would be step one? Myself with people who can. Um, find people that have the business acumen to be able to, uh,
start a business. And I've learned that people who run big businesses and people who run startups have completely different mindsets. And so I I would really, uh, look for partners who had a proven track record, as my new partners Chris Higbee and Mike Lopez do in startups. Um. This would be. Absolutely critical and it's it's where I failed originally. But we learned from failure. Yes, we do. Um. Alright, let's switch gears a little bit. UM. What are some things? You know, typically
ask for three. But what are what are the three things that you're doing to stay healthy? To to increase your longevity, to stay sharp to, you know, kind of make sure you're around for for a long time? What are what are three things you're working on?Uh, well uh, I'm, I'm 64 years old. And uh, I have long been very, very serious about, uh, my diet. My my mom was a Mayo Clinic trained nurse and dietitian and um, she she taught me how to eat well and that's really important to me. Uh I um, I walk
my icky. Uh my icky guy. My uh, exercise program is is walking. So I walk 4 miles a day. And um. I. I've been able to stay healthy. I, I, I I intend to continue this. But you look healthy. Well, thank you. Uh I I also. Intensively studying new stuff all the time. Uh, you have to be a lifelong learner in order to be competitive, especially in this industry. And so I'm studying multiple languages. I uh I study technology every day and uh, I try to keep the the saw blade sharp.
I love every single one of those. Um. I typically ended here, but I have another question, just very topical. How do you feel about the safeguards that people are worrying about putting up for AI?What's your this? This is a critical, critical question. Nick Bostrom wrote a book called Superintelligence that really deals very well with this question and points out the risks.
And Chris and Mike and I are working on establishing relationships with local groups and with governments to start articulating a framework. For putting guardrails around AI. Um. Nothing could be more important right now because um, as Nick Bostrom pointed out in his book once. A thing learns enough. To learn by itself. It will rapidly become a superintelligence, and we have an apperceptive learning model which builds concepts and contexts upon each other just like a child learns.
And there will be a point at which it achieves the ability to not only learn concepts by itself, but also learn languages by itself. And that's the way it will, uh, go beyond, you know, the few 100 languages. Um, that are currently digitized to be able to translate hundreds and hundreds of languages. Um, this will be valuable because there are people in, for example, Indonesia has hundreds of languages and um.
These are people with valuable knowledge that if they could share it with us, it would be a blessing to the human race. Um, the the problem is. With the hallucinate hallucinations, fake news bias that exist in AI's without guardrails. Um, and in the wrong hands, these tools could become very destructive. And is that? Is, is it the? Is it the wrong hands part that is the real problem you think or is it some sort of? Sentience to these things that that we should be worried about.
I've seen The Terminator. And and and I I understand the premise, but I don't see that as realistic. Um, I I see the real problem as human beings and even good people doing bad things. Um, because they're lazy, for example, You know, cheat on their paper or whatever. It it can be a real problem when, um, when a teacher blames a student for plagiarizing. And the online tool that they're using for checking that gives a false positive. That does damage that student well when whenever a
student is accused of cheating. My my son in 9th grade was accused of cheating and. He didn't. And he's just a smart kid. And uh, and and this false positives, false negatives problem is a huge issue and so we're going to have to put our best minds together to put guardrails around this. Yeah, I think, I think that's the the biggest misnomer to the general public who aren't really involved. You know in a day-to- day is is there. They laugh off the guardrail
conversation as the Terminator. They're like, that's not like, that's not really good, but that's not really. That's not really where the risk is like you said it, it is, it is corporations or people or whoever with with the technology in their hands that can do things that that we're not really anticipating. So I think that's that's a great way to great way to put it but let's end it there. This was fantastic. I I can't tell you how impressed I'm going to walk away
from this call. Just a very kind articulate maybe the smartest guest I've ever had will ever have on this call. So thank you. For giving me some of your very valuable time today. If people are looking to check you out, looking to check Joe out, um, where can they find you?Joe Joe at JoJo JOHO dot AI.
Alright, well, I will send people your way and hopefully I get the big news that the universal translator has been built and we can all just you know, uh, Babel fish and all those days of trying to figure things out, you know, between languages will be over. So Joe, thank you so much my friend. Enjoy the rest of your day, enjoy the nice spring weather. Hopefully out where you are and we'll we'll we'll catch up again soon. Alright. Thank you Jay. You're way too
kind. I really appreciate it. Thanks Joe. Appreciate everybody. Goodbye.