The Future of Generative AI with Zapata AI Co-Founder and CEO Christopher Savoie
October 24, 2023 – On today’s show we welcome special guest, Zapata AI Co-Founder and CEO Christopher Savoie. Zapata AI is an industrial generative AI software company developing solutions and applications to solve enterprises’ hardest problems.On the show we discuss:
- How Christopher went from getting his PhD in a medical field to founding several tech startups
- A deep dive into what artificial intelligence is and what it can accomplish
- Some use cases for industrial generative AI
- How the company offers faster, cheaper, and more accurate generative AI
- And more
Welcome investors to The Absolute Return Podcast. Your source for stock market analysis, global macro musings and hedge fund investment strategies, your hosts, Julian Klymochko, and Michael Kesslering aim to bring you the knowledge and analysis you need to become a more intelligent and wealthier investor. This episode is brought to you by Accelerate Financial Technologies. Accelerate because performance matters. Find out more at accelerateshares.com.
Julian Klymochko: All right. Welcome to the show, Christopher from Zapata AI, Co-Founder, CEO. How are you doing today?
Christopher Savoie: Doing well. Thank you for having me here, Julian.
Julian Klymochko: Yeah, so generative ai is a, you know, massive development, especially over the past 12 months. Really taken the economy and the market by storm with a ton of applications. So, I’m super interested to get into your insights and what you’ve been working on in that segment. However, prior to getting into that, you have an incredible background. I mean, Biology, Law, Molecular, Medicine, where you got your PhD, your published scholar in Medicine, Biochemistry, and Computer Science. Now you’re doing tech and software startups. Can you talk about your background, how you went from Law to PhD in a medical field and now the founder of not just one, but several technology startups?
Christopher Savoie: I guess a charitable way to say it is that I didn’t know what I wanted to be when I grew up, and I ended up doing all the above. But no, no, I started programming computers when I was a kid. Got a VIC 20 very early back in the day, and I think I was a frustrated Computer Science Major that never got to do that. I actually grew up a little bit poor and couldn’t afford to go to the engineering schools that I had gotten into. And so I went, you know, the track that a lot of folks go to in being either pre-med or pre-law, and I chose pre-med. Med school was never going to work out great for me because I’m a hypochondriac.
So, that was probably a poor life decision to be pre-med and eventually med. Ended up in medical research and worked on immune molecules. Actually, I became basically a biophysicist working on structural biology back on the old Indigos back in the day in the nineties. And while I was a grad student a poor grad student, this thing called the internet happened. And I was asked at my institute, hey do you want to learn how to program these Cisco routers? I was in Japan, and they’re like, you understand English, so you can talk to the Cisco people about this, I raised my hand. I said, sure, yeah, I’ll do that. Wow, I get to do computers after all. So, I learned to do things like program routers and switches and learned to program HTML which at the time was a new language and Perl CGIs and all this kind of stuff.
And I used that actually on the side as a side gig to start a network consulting company that got funded by a VC. And Jeffco, one of a larger VCs in Japan. So that was my entree into entrepreneurship, I guess by running a side hustle as a web company that actually took off and was doing multimillion dollar deals while I was in grad school. And that was mid-nineties. And then Babak Hodjat who’s now running ai over at Cognizant and myself came up with this pretty cool way of using machine learning. I was doing machine learning on these molecules, these immune molecules, to try and find out what combination of amino acids would give you the right thing for a T-cell to want to attack a cell.
So, for that reason, for vaccine development, that kind of stuff. And Babak Hodjat had been doing work in artificial life, and together we came up with this technology to parse language, which was a small language model, I guess you could call it. And this was technology. We spun it out. We moved to California to the Bay Area and started a company in Mountain View called Dejima that was eventually acquired by Sybase that actually had the technology. Developed the technology that is behind Siri at the end of the day, AOSA. This agent oriented software architecture that we created. So that was kind of my second foray into entrepreneurship. And so, it’s been machine learning in a bunch of different areas since then.
Applying machine learning to that space, which was natural language interfaces and mostly in telco in that life. And then later applying Bayesian machine learning in another startup like GNI in the pharmaceutical industry applying this to regulatory networks using that for drug development. And that company went public in 2007 and then later came back to the States. That’s when I did my law degree at night. Because why not? Right? and while I was doing that, I was working eventually at Nissan, Verizon, and then Nissan doing machine learning applied to battery, predictive analytics, and a bunch of other things. Brought the first Hadoop stack into the automotive industry, and that was my first foray into doing mobility stuff. So, I guess, you know, the theme is, yeah, a lot of different things, but mostly machine learning across a number of different industries. And then you know, we end up doing generative ai here at the present company, which kind of ties it together because we’re doing this across a bunch of different verticals, a bunch of different industries.
Michael Kesslering: So, you mentioned the machine learning really tied everything together with your previous startups, successful startups, I would add. Could you talk about what else were some of the key lessons that you learned through these multiple startups that you’re carrying forward to today?
Christopher Savoie: Yeah, I guess one key thing I learned with Dejima, of course, was, you know, we had a demo of what could have been Siri back in 1998, working with a VCR and saying, hey, record this and record that. But we didn’t have certain technologies that were necessary to make that into a product yet. We didn’t have 4G Networks, 5G Networks. We didn’t even have 3G networks yet. And so, you know, you couldn’t use this in a mobile type of format in 1998. There were other technologies that just weren’t there. So, I guess one of the lessons is don’t be too early in some ways, but you can be prescient and have an idea, but you need the surrounding technologies and the environment to be there to turn that really into a business, into a product. So yeah, later on 2007, it gets acquired, acquire acquired, the technology eventually ends up at Apple. And there you go and off to the races, but that was a good almost 10 years for that to happen. So, you know, you really have to, I think as an entrepreneur be realistic about when the computing platforms and if you don’t have cloud, you can’t do certain things, right? and that’s something I think that has been a lesson for me that I hopefully learned a little bit from.
Julian Klymochko: And with respect to your various entrepreneurial ventures and roles at different companies, machine language has seemingly always been the common thread. And machine learning, AI has become very popular. However, it seems like probably the average person, while I have heard of it, they probably don’t know exactly how it works. So, from a high level, can you explain to our listeners, you know, what machine learning is and more importantly, what it can accomplish, why it’s such a phenomenal new technology?
Christopher Savoie: Sure. So, machine learning is a machine learning stuff. Like, again. So, what is learning? Well, there’s kind of rote learning, right? Where you just, you know, memorize a lot of stuff, right? Like, you memorize a lot of music you become a musician, well, maybe but there’s also different techniques to learn different types of this technology. But the basic idea is that you take data and you feed it into the machine, and the machine learns that data and it learns statistics about that data and how that data interrelates in different ways. So that allows you to recall the information, to find associations in that information. And, you know, humans can memorize lots, for example, music, like that. But seeing the correlations between Bach and Eddie Van Halen guitar riffs. Computers are a little bit better at that because you can’t hold all of it in your head at one time.
So, the really cool thing about machine learning as applied to really large data sets is that, you know, the machine learning is able to really understand trends in massive amounts of data that one human really can’t do. So that’s one piece of magic. Now, generative ai is adding a whole different level to our capabilities. And so, machine learning, if I were to give an analogy of this was kind of rote memorization. Generative ai, what’s I think really surprising to a lot of people is this kind of human-like behavior that’s coming out of it, right? You look at Chat GPT, and it’s kind of surprising. It’s not like the machine learning that we’ve had up until now. The machine learning that we’ve had up until now really is rote memorization. You show it a thousand pictures of cats and it learns those thousand pictures of cats.
And if you tell that machine learning model, hey, give me a picture of a cat. It’ll take a piece of that, one piece of that, one piece of that, collage it all together, and here’s your new cat. But if you showed it a picture of a purple polka dotted cat, it would say that’s not a cat. Generally speaking, right? With some exceptions because it’s just rote memorized what a cat is. It doesn’t have an idea of really cat as a concept, right? The human thing that’s really cool about these generative models now is that it takes it to a different level and does something that’s creepily a little bit human-ish that we see as being human, where it says, okay, I know a cat, so, and I know Picasso, so I can actually draw a Picasso cat.
You can tell one of these models, give me a Picasso cat, and we’ll recognize as a human that that’s a Picasso-ish cat, right? because It’s got a model for what is Picasso, and what is cat, right? So that generalization capability is really new. This is a new type of thing. And that’s what I think is so surprising is the behaviors that you get when you can generalize. And generalization means that you have this big statistical model of what a cat is. It’s not just that it has four legs and this and that. It can have a probability density of different tails, different tail sizes, different paw sizes, different colors, and all of this stuff. And now the second thing that you’re able to do once you have what is a cat, is that you can create things that are coloring outside of the lines a little bit, right?
You can have a Picasso cat, you can have a blue cat with red stripes, you can have a Picasso cat, you can have a Monet cat, you can have a Monet cat. You can do these kind of creative things that we find to be kind of human. Once you have a model of what cat is, you can have Doraemon, which is this Japanese anime character that has no ears, and you still recognize it as a cat, a blue cat with no ears, right? So that’s the kind of surprisingly human thing. And yeah, it’s not just about pictures of cats. We can use that kind of technology to be creative in all kinds of human areas, like engineering, a new bridge, engineering, a new financial product, all these kind of things. And that’s what’s really surprising and really powerful about this new generation of machine learning compared to what we’ve had before.
Julian Klymochko: I agree. Generative at is such a cool technology, and many folks would be familiar with certain aspects of that, whether it’s text-based applications such as Chat GPT, which is fairly popular, and, you know, you type something in and something really cool pops up most of the time or the generative image applications such as Mid Journey and others that create some really cool graphics. Now, what you’re working on specifically is industrial generative ai. What is that?
Christopher Savoie: So this is taking generative ai not to just do language tasks, and not to just do pictures, but to actually use the same kind of creative capability of this new machine learning to do creative tasks in industrial applications, industrial meaning enterprise business applications, things like developing a new financial product or using it to detect things in, an IOT environment, in a factory, this kind of way. To do predictive analytics using this new capability.
Julian Klymochko: That’s great. Now focusing on your company, Zapata Ai, which was actually spun out of Harvard University in 2017. Can you provide some of the background and what you’ve accomplished since that spin out?
Christopher Savoie: Yeah. We’ve actually, we started working on applying the models that we had. We come out of a lab that did linear algebra. I know that’s a kind of a mathematical term that is not really a category. And so, I tell people we’re a linear algebra company. People kind of get a little bit weirded out by that, but basically, we came out of a lab that did quantum algorithms. What are quantum algorithms? Well, it’s a kind of math that does these linear algebra problems, these high dimensional space problems. And from the very beginning looking at places where we could use this to do industrial applications. So where can we use linear algebra, this high dimensional space, really difficult math stuff to do practical stuff in industry, in business.
And the places that we found very immediately, so this is 2017, our first patent filing and generative modeling was actually in 2018, less than a year after we were formed. And we realized very quickly that generative modeling because of what we described with how these things work, you know, you model of a tail for cat and a paw for a cat. These are all things that are very much like the math that we do using in physics and chemistry in the lab that we were in. And so, we realized that we can apply this to things like large language models, which are these large statistical models. We can do what we do in particle physics and in chemistry and apply it to this generative ai paradigm and a lot of other industrial mathematics problems. So, we spun the technology out, did a lot of work on the Ip, did a lot of development, developed a platform that allows us to train these models and compare these models. And then we got engaged with industry with a bunch of different vertical companies in energy and other sectors that had these problems and wanted to apply these algorithms to their actual business problems.
Julian Klymochko: And in terms of practical applications, I’m sure there are many, I went cruising through your website and I saw pretty interesting video of you guys working with Andretti Motorsports and applying, you know, your software to all the analytics that goes into car racing, which was super cool. Can you discuss, you know, whatever would be the most sort of practical applications that resonate with the lay person in terms of understanding how the rubber hits the road, so to speak, as this technology applies to real world applications?
Christopher Savoie: Well, let’s talk very specifically, pun intended about the rubber hitting the road, because that’s actually one of the use cases that we’re able to talk about which is the slip angle of a tire. What’s a slip angle of a tire? Well, the direction that tire is moving on a track, when you go around and you’re racing at 200 miles an hour that’s really important because to win a race in racing, you have to do pit stops and change your tires at certain times because the tires degrade and then they’re slower over time. But you got to choose your marks about when you’re going to do that, because it takes time to pit a tire and change the tire. So, when you do that, how you do that is how you win and lose the race strategy wise.
So, to figure out when that tire is degrading or how it’s degrading or when that speed is falling off, you need really good physical models of this stuff. And it turns out we can use generative ai to be predictive in that way. Why, because we can learn what the slip angle is of that tire. What’s the slip angle? Where it’s sliding on the track. Well, we’ve got about a hundred some, sensors on that IndyCar when it’s going around the track. So, we have a lot of data. The problem is there are certain physical properties like the slip angle of that tire that we don’t have a sensor for because you can’t put accelerometers on the outside of an open wheel racing car because if you rub the wall, it’s gone. And that’s a pretty expensive day.
And they also throw off the balance. You just can’t do it. You can’t use GPS because you’re going 200 miles an hour. And at the last Indy 500, there was a tie down to one 10000th of a second, so, GPS too slow. Cameras are not accurate enough. So, what do you do? You don’t have a sensor for slip angle. Well, what about the tire direction? Isn’t that the slip angle? No, because when you’re going 200 miles an hour and the car and you turn the car left the car wants to go right, because of centrifugal force. So that is not the slip angle. So, what do you do? Well, we have 20 years of historical data with the Indy team of cars going around the track with all these hundreds or so sensors. And what we can do is, like we learn with generative ai, what a CAD is, we can learn what a steering trace is, we can model that kind of behavior, and then we can actually synthesize what that inferred channel is based on the correlations that we learn. So, we’re able to use generative ai to very accurately generate an inferred sensor that allows us then to feed into our tire degradation models and our race strategy models to figure out when we should pit the car, which is a win or lose decision oftentimes in the racing world.
Michael Kesslering: From a competitive dynamic in your industry. How are you able to offer faster, cheaper, and more accurate generative ai? When looking at your website and the investor presentation that comes up a lot, how are you able to do that versus the competitors?
Christopher Savoie: The simple trite answer, pithy answer would be math. But it’s a different kind of math that we’re using. So, a lot of what we’ve heard of is these neural networks, but deep learning is really based on neural networks, which work really well. We see it, I mean, chat, GPT is great, you know, GPT-4, GPT-5 eventually would be better. But what are we doing? When we went from GPT-3 to GPT-4 we grew the model from, you know, one point something billion parameters to a trillion parameters. You know, it’s like 10 x, 10 x, 10 x, which is great for Jensen and the Nvidia guys. Really great for Nvidia stock. But at some point, it becomes a real mess because, you know, these GPUs are expensive.
They throw out a lot of carbon and heat. So, it’s not a sustainable thing. It’s really not sustainable even from a business perspective. Because these models get more and more and more expensive to run. So, what do you do? Well, the problem is that the math that’s running them is based on these neural networks, which worked great. And when the solution was just throw more and more and more GPUs that worked, but we’re kind of getting to a point where that math doesn’t work anymore. And the math that we have coming out of a quantum world that works for quantum physics is designed to do that math more efficiently. We deal with these statistical models like the model of a cat or what a cat is or the steering trace in a racing car all the time in quantum physics and quantum chemistry.
And in those cases, you know, the models are immense as far as the variables and the variable space. These are huge math problems, you know, well, what is a collider doing in high energy particle physics at Oakridge? Right? So, for us you know, we know how to do that. We’ve been doing that kind of math for decades with a different kind of model, a different kind of math. And that comes out of the physics world. So, the machine learning people were over in their world saying, more GPUs, more GPUs because it worked, and it worked really well. Here we were in the physical and the chemistry world saying, well, there are other mathematical models that can do this. And now we’re applying that to machine learning with a completely new model that is actually quicker, faster, better. It’s these quantum algorithms that’ll eventually run on quantum hardware, but they can run actually on GPUs today much faster than the neural networks can run.
Julian Klymochko: And from an investor’s perspective, thus far, you look no further than NVIDIA’s share price and clearly, you know, they’re going to be making a lot of money with all the GPUs being acquired and snapped up. Now with respect to offering artificial intelligence and machine learning to enterprises, that’s kind of an unknown, you know, the chat GPTs of the world, the Mid Journeys, and are they going to be able to extract profits from that, have great business models. Now, with respect to your company, can you talk about the business model and how the technology will translate to profits for investors?
Christopher Savoie: Sure. I think there are a couple of different categories. It’s still unknown whether, you know, the language model thing is going to be, you know, a big enterprise business.
Julian Klymochko: Right?
Christopher Savoie: There are definitely though certain areas where language models are going to be helpful for interfacing with data in this kind of a thing. But the idea that you’re going to take a model that’s trained on the entire internet and, you know that has been trained on Ozzy Osborne heavy metal lyrics and Aesop’s Fables and children’s novels and apply that to filling in a customs form for a company is a little bit silly. You don’t need that big a model to do that stuff. Maybe you can use parts of that model, but it needs to be retrained. It also needs to be run securely inside a company because if you’re going to use people’s private data if you’re a healthcare company and regulated industries and in finance and banks and that kind of a thing, you can’t just like take a model in there, hoover in all your data and let everyone access it.
That’s not a good model. So, I think we’re still in the early days of figuring out how we’re going to do that. In our case, at Zapata because of the way we’ve made our models smaller and more pliable and more trainable, we’ve made it cost effective where you can take your model and some of these open-source models, bring them into an enterprise and be able to retrain them at a reasonable cost for that kind of an application. So, you can train it on the data that you’re actually looking at not on the entire internet, which leads to, you know, all kinds of, we’ve call it hallucinations, where, you know, the model just makes shit up.
Julian Klymochko: Yeah.
Christopher Savoie: You know, that’s not good. You can’t do that with somebody’s, you know, clinical trial data that you’re trying to fill in for an FDA submission or something like that, or people’s financial data. That just wouldn’t be cool. So, the language model’s one thing, but there’s also this other side of using generative ai for the kind of use case we talked about with Andretti. You know, you can use that kind of virtual sensor thing, yeah. For cars, for automotive, but also for things like new insurance products and different things like that. We’re getting creative about the products and generating new data, generating new scenarios and this kind of a thing. So, the numerical side of the business has pretty immediate effect. And there are a lot of companies now looking into how do we use not just the language models, but these numerical models to do really interesting, important things in data analysis inside companies. And so, I think it’s not even a question that new data analysis techniques and these kind of things in this world of big data and data-driven company decision making are going to be important and people are going to invest in and spend money there.
Julian Klymochko: Yeah, it seems like we’re so early stage with respect to generative ai, so it’s great to see what you guys are able to accomplish, and I’m sure there’s so many applications that people have not addressed yet that are going to come out in the future. And so, we’re really looking forward to your future progress at Zapata ai. And now, Christopher, prior to letting you go and wrapping up the show. I thought it was very interesting and we chatted about before the show how you have a black belt in judo, very experienced in jiujitsu and a lifelong martial artist. Did your success in martial arts contribute to your success in business and academics? And if so, how? What’s your viewpoint on that?
Christopher Savoie: Well, I think, you know, judo, grappling, you know, in a wider sense really teaches you resilience and grit. It teaches you humility, you know, you have to problem solve. And you lose a lot actually, which teaches you humility that you’re not always going to be right. But it also teaches you how to get up. There’s a judo saying from Kano, the guy invented judo. It’s like, fall down seven, get up eight. And really that resilience I think is a lifelong lesson. It’s great. And why I recommend, you know, some form of disciplined martial arts to everyone because you know, really in business, you’re going to fail. You’re going to make mistakes, you know, read the biography of anyone who’s really made it.
And it is that social media meme of, okay, here’s what success really looks like. It’s pretty messy. And, really, I think martial arts trains you to really keep a cool head to understand your environment and, okay, I lost, now what did I do wrong? And be introspective about that and play it back. I’m sure you have a jujitsu background. I’m sure you’ve done that where like, all night you’re playing back, oh, I had just done that. How did I get into that choke, how did I get out of it? I could have got out of that. And you problem solve, and hopefully after a lot of experience, you learn how to deal with various problems. So yeah, I think that there are really good lessons in there for how we deal with business problems and, you know, environments and working environments that I think are really valuable.
Julian Klymochko: Oh man, you’re totally right. Last weekend I got trapped in a buggy choke, and so [laugh], I had to go home and watch some YouTube videos on escapes that I could practice so that it doesn’t happen again. So, lessons learned, and it’s certainly one of the most humbling experiences to get tapped out and certainly checks your ego. So, I definitely understand those life lessons. And Christopher is great having you on the show. Love what you’re being able to accomplish at Zapata ai. You guys are taking the company public, which is super exciting, and we wish you the best of luck. We’ll be watching you guys as you go.
Christopher Savoie: Thank you, Julian. Thank you, Michael. Appreciate it.
Julian Klymochko: All right, take care. Bye everyone.
Thanks for tuning in to the Absolute Return Podcast. This episode was brought to you by Accelerate Financial Technologies. Accelerate, because performance matters. Find out more at www.AccelerateShares.com. The views expressed in this podcast to the personal views of the participants and do not reflect the views of Accelerate. No aspect of this podcast constitutes investment legal or tax advice. Opinions expressed in this podcast should not be viewed as a recommendation or solicitation of an offer to buy or sell any securities or investment strategies. The information and opinions in this podcast are based on current market conditions and may fluctuate and change in the future. No representation or warranty expressed or implied is made on behalf of Accelerate as to the accuracy or completeness of the information contained in this podcast. Accelerate does not accept any liability for any direct indirect or consequential loss or damage suffered by any person as a result relying on all or any part of this podcast and any liability is expressly disclaimed.
