The Science of Flavor

November 2021

In November 2021, Gastrograph AI founder and CEO, Jason Cohen, was featured on the podcast "Through the Noise" discussing the science of flavor.

"The Science of Flavor" Transcript

Guest: Jason Cohen | Host: Ernesto Gluecksmann

Speaker 1:

Through the Noise with your host, Ernesto Gluecksmann.

Ernesto:

Jason, man, thank you for joining me on the show. Man, I was reading through your blog and every post, I started thinking, "Wait, I got a question about that." And then the next post, "I got a question about that, I got a question about that," it's just really interesting and, I would say, probably complex. No, complicated but maybe not complex world that you're in, right? Hey, everybody listening, thanks for jumping on. If you're on the YouTube channel, thank you.

Ernesto:

We don't have that many listeners. I know most of you have been listening to us over the years on the podcast but the YouTube channel could use some love, so thank you for that. This is CEO and founder of Gastrograph AI, you're Jason Cohen, thank you for jumping in. You got the word gastro in there and you got AI and that's what caught my attention to reach out to you. We're talking about flavors and you're using AI to figure some of this stuff out, right? How would you, Jason, break it down to somebody who's not in your world what your company does?

Jason Cohen:

Well, first, thank you for having me. I'm really excited about this conversation. So, what we do is, if you think about individuals around the world tasting food and beverage products and let's say that you and I were to taste the same product, there could be differences in what I perceive, there could be differences in what you perceive. And what our technology does is it allows companies, food and beverage companies, trying to make products, new products, reformulate products, it allows them to understand with different target consumer groups, different demographics around the world, what different people are going to perceive in that product from that predicted preferences. So, we don't work on a personalization level, we've done work on a personalization level, but most of the time, we're not predicting for you, we're predicting for you and people like you what's the distribution of what you're going to perceive, what's preferences are going to result from those types of flavor profiles.

Ernesto:

Okay, because you have some wonderful blog posts that are just using, I know they're not your clients, but they're examples of what the world could be. And you were writing or you're trying to figure out the predisposition, let's just say, of a certain set of potato chip products, Hot Cheetos in particular, let's say, in the Chinese market. Which one of a set of products might work best in a specific region of the country? So, okay, well, first of all, how do you know what folks in China, what are their particular flavor profile preferences relative to what ours typically are like? How do you know there's a difference?

Jason Cohen:

Great question. So, we actually do that the hard way. We have been collecting data since 2009, we run our own large scale consumer panels, consumer demographic surveys around the world, we've now done that in more than 26, soon to be 30 countries and we take up regional preferences as well, so we'll do testing in multiple cities as we enter that. And so, in 10 years now, 11 years, sorry, since 2009, we've built up the largest sensory data set of on market consumer products that's ever been assembled. So, what we can do with all that data, we're taking data on flavor, aroma and texture, we're taking data on a variety of different products, whether coffee, cheese, chips, scotch, cognac, salty snacks, baked goods, confections, chocolates, anything that's ready to eat or ready to drink and we push the boundaries on that quite a bit, for example, instant noodle soups.

Jason Cohen:

What we can do with that data is we can use it to revert to first parameterize the distributions of perceptions that exist in that market. So, we can look at a specific country or a specific part of a consumer demographic and say, "This is their sensitivity, they're going to respond to this underlying sensory stimuli." And then, once the AI learns that, we can reverse engineer the preferences that are driving that market. So, we can actually say, well, Hot Cheetos was the example, Hot Cheetos in China, that maybe that product is available, maybe that product is not available but we can take a flavor profile about anyone, translate that perception to the target Chinese consumer demographic and then use our Chinese preference model to say this is how that product would perform amongst those consumers.

Ernesto:

Man, okay. So, to try to summarize what I think I understood you say is you are surveying ahead of time particular regions of the country. So, let's say, you're in Shanghai, you have a good understanding because you did a bunch of surveyed flavors. Does it come down to you bring in a volunteer and you put a nutty flavor and then you put something else and like, "Do you like that or do you not like that?", on some scale or do you go through that step by step process?

Jason Cohen:

So, we have our own custom sensory system. It's a mix of scaling and skilled attribute intensity models with a descriptive analysis component. And to explain a little bit about what that means, there are 24 broad-spectrum flavor categories-

Ernesto:

Okay.

Jason Cohen:

... so 24 flavor categories. Those are things like fruits, earthy, herbaceous, roasted, nuts and seeds. And our original research, my original research back at Penn State University when I started this company, was how do you build a sensory system that totally encompasses gustatory flavor such that anything you taste, you can put onto the system without any offloading, without any covariance, without any cascading effects in between interactions between the terms.

Ernesto:

Are you saying distinct flavors, like how many distinct flavors we could select?

Jason Cohen:

Yeah, into categories, exactly.

Ernesto:

Okay.

Jason Cohen:

And so, those categories are broad categories, right? Fruits have many different flavors, roasted has many different flavors. And so, you can go in on our system and you can rank. You can say, "I taste fruit at a four out of five" or, "I taste fruits at a one out of five." That's the edge of perception, right? A hint of fruit or zero out of five, not present. And then within groups, you can then say I taste specific flavors. So, it's a modified form where you could say, "I taste apple, I taste peach, I taste kiwi," and in fact, you can give those sentiments. So, you can say, "I taste apple and it's neutral, I taste kiwi and it's positive and I taste the peach and it's negative," just with a few taps. And so, we can very quickly build up quite a high resolution picture of what's in that product.

Ernesto:

So, I taste this one, I taste a fruity note to it. Okay, well, is there a strength in any of these? Like it leans towards kiwi for whatever reason?

Jason Cohen:

Yeah, so we don't ask for strength of specifics and the reason is because consumers are good at ranking overall and they're good at some level of identification, misidentification rates are about 15 to 20%, even for common flavors, but they're not good at the ability to parse through, "Well, I'm tasting-"

Ernesto:

Interesting.

Jason Cohen:

... "raspberry, peach and kiwi. The raspberry is strongest and the kiwi is weakest," that becomes much, much harder. The cognitive load in order to go through all of the flavors is too high. And so, actually, we use the AI itself in order to actually determine, we call it a decomposition, in order to decompose the overall flavor profile into the relative attribute intensities. So, the AI can look at whatever you put in where, let's say, you taste that wine and you say, "I'm tasting a three out of five on fruits and a two out of five on dry and a one out of five on earthy," and everyone likes to talk about minerality in wine, so it's two out of five on mineral, the AI then goes in, and depending on what you marked, it's able to say, actually, you're tasting gooseberry and the gooseberry is 3% of the flavor profile and you're tasting calcium and calcium is 1% of the flavor profile. And it can do that very, very accurately.

Ernesto:

Oh, because you know the ingredients that were in the particular product that they were tasting, you're not asking them to find the ingredients or the intensity, you're more interested in do they like it or they don't, right? In a sense.

Jason Cohen:

In many products, we know the ingredients. But in a product like wine, the ingredients actually would not be helpful in that exercise because no one's adding a raspberry flavor or calcium flavors or anything like that, right? It's just those are compounds that are the grapes or results of the fermentation process. So, we actually don't need the ingredients in order to do the flavor profile decomposition. We can determine the actual relative magnitude intensities of specific signatures because the AI has seen now hundreds of thousands of products with or without those labels. So, we think there's two sides of the system. The first is being able to model what someone's going to taste because all preferences are built on perception, right?

Ernesto:

Yeah.

Jason Cohen:

If you and I taste the same product, we may or may not perceive the same things, we may not have the same sensitivities, we might have different vocabularies, we might have different preferences that come from our experiences growing up. So, we have to control for that before we can build a preference model. So, once we know it's able to perceive it, then we predict their preferences.

Ernesto:

Oh, my god. Okay, so you're using AI just on the data collection of all these panels, surveying it just to normalize the data into something that predicts well enough?

Jason Cohen:

Right, correct.

Ernesto:

Oh, my god. Okay. Then somebody else comes around and goes, "Hey, we're thinking we're going to try to sell Hot Cheetos." That's my love, I love Hot Cheetos. "We're trying to sell Hot Cheetos in China, do you think it's worth the effort?" Is that one of the goals of why we would contract for you? Can you tell me, with some sense of accuracy, whether this product in its current formulation is likely to be interesting enough for people to buy it?

Jason Cohen:

We can't. We're careful to say that we don't predict sales. And we don't predict sales because we don't know the distribution, the product-

Ernesto:

You're not in control of the deal making, right? Yeah, the cost of the bag of Hot Cheetos might be astronomical, I don't know. Who knows?

Jason Cohen:

Yeah.

Ernesto:

But whether they will actually like it.

Jason Cohen:

Exactly, exactly. Will it be competitive, will it be preferred. And we can go a step further than that and we could say what is the driving preference? Is it the spiciness? Spiciness preferences here in the United States are very different than the spiciness preferences in China. Is it the cheese flavor? Cheese is very much a new thing in China. Or is it something else and what needs to be modified to make this product more preferred in that market? And sometimes they make those recommendations, and sometimes they don't. Sometimes it's not worth the cost of modifying the product.

Ernesto:

Right. Yeah, that's what I noticed in all the blog posts. You said it could do even better if they'd added a little more this mineral or that thing or that type of flavoring based on what the data that you have collected for the region of the typical things that they particularly like, right?

Jason Cohen:

Exactly. Exactly.

Ernesto:

Oh, my god. And this is done for the customer, it's done programmatically. They're not actually having to go find the panels, do all the data collecting, this is like analyze your flavor profiles, essentially.

Jason Cohen:

Exactly, exactly. And that's really the gain of the system. In order to do this previously, you had to do a large scale consumer testing, you had to recruit panels from target markets, you had to design an experiment to test exactly what you wanted to know but we always considered that snapshot data. That data may have been true and correct if you had a good experimental design with enough statistical power and results that allowed you to reject the null hypothesis or whatever. But that data couldn't be reused, it couldn't be built on top of, it couldn't be used to inform and update and train models that can make predictions about the future.

Jason Cohen:

And that's really what gave us the idea about this opportunity is how do we go from descriptive, frequentist statistical hypothesis testing to how do we go to model building where data collection and running experiments actually has a positive ROI? Not just the ROI from answering a single question, but an ROI in that you can create better and more accurate information and predictive models over time so that, when you actually want to know something about the future, you have a trusted model that has proven to be true, has proven to be accurate about predicting what's going to happen next. If I take this action, if I launch this product, if I make this change in my formulation, what's going to happen in the consumer experience?

Ernesto:

Right. And probably, oh, the realization that flavors are evolving in and of itself, right? I mean, 15 years ago, walking into a 7-Eleven, you wouldn't really find that many super spicy snacks. Now, in the US, at least in some parts, at least where I live, there's a lot of really hot, super spicy stuff that's available to people. And so, essentially, because those products are there and because people have been buying them over time, the flavor profiles of preferences have suddenly evolved. So, coming in with an extra blend or something that counter to that might not do as well, right? You have to be thinking through that for these-

Jason Cohen:

That's exactly right. Repeating flavor yields perceptual differences which real yields differences in preferences over time. And so, people can become adapted to flavor, people are now more accepting of very spicy flavors. People can move towards or away from flavors because of repeat exposure. And just as a very contemporary example, COVID-19 quarantines was the largest systemic shock to consumer behavior since rationing in World War II, it changed what people were buying and consuming and drinking. If you think about sweetened, carbonated soft drinks, which frequently do upwards of 25% of their volume on on-premise sales, well, there was no on-premise. There were no stadiums or theaters or ballgames or concerts and people weren't buying it. And we published a study in EuroSense 2020, or maybe that was 2021, time blends together throughout the pandemic now.

Ernesto:

Yeah, of course.

Jason Cohen:

EuroSense is a conference for sensory science and we looked at whether or not COVID-19 was an accelerator of existing trends or a reversion to the previous preference state. And what we actually found was that it was both... There was all this talk that people were going back to childhood foods, looking for comfort foods, looking for snacks from their childhood, that was true, although a lot of it was actually a removal of exploration.

Jason Cohen:

For pre pandemic, everyone was in this exploratory mood, they were looking for new experiences, they were trying new products, companies were coming out with many new flavors, there was this big flywheel of individuals trying new things, gaining new experiences, gaining new preferences, all of that stopped when the pandemic hit. When people reverted to product preference state, they fell out of that exploratory mindset, and so that reversed some of the flavors. On the other hand, stuff like sour acidity which was already a growing trend, became more prevalent. And when you hear about people moving to things like sourdough, that is part of that larger trend towards more sour products.

Ernesto:

Man. And so, then, that changes some of the data sets to test against. So, do you get customers that come back and go, "We've been doing pretty well but are we going to continue to do well with this? Can we run another test or can we run a slight variation because we think we need to adjust the acidity in the product or something?" Right?

Jason Cohen:

Yes. So, on the macro level, we can predict usually about two years out, that's just for major flavor profiles, so that's things like sourness and acidity or fruitiness. For micro level trends, I have a really controversial view in the industry. A lot of companies will say, "What's the trend, what's the trend, what's the trend?" talking about like, "Well, is it mango or is it pineapple this summer?" And my take on it is that everyone has it backwards and trends go back, right? If you create a good product that people like, you'll create trends.

Jason Cohen:

This is talking specifically about micro level trends, specifically, because obviously, if sourness and acidity needs to be increased and the sugar is decreased in order to create a balanced flavor profile, that's a macro level, that's a real change in preference state. But these micro level trends, if we predict that you should create a product that tastes like lychee or that lychee is a nascent preference statement and there's no products that are meeting that preference, we're not predicting that that's a trend because we're not predicting that anyone will do that. And if no one does it, it will not trend.

Jason Cohen:

What we're predicting is that there is an opportunity, unmet preference need that you can take advantage of and [crosstalk 00:18:56]-

Ernesto:

Right. Where's the niche that you could-

Jason Cohen:

Yeah.

Ernesto:

The gap in the marketplace for the micro stuff.

Jason Cohen:

Exactly.

Ernesto:

The grand macro, the big trends have to be monitored, but at the smallest level, you're really looking for gaps, something that keeps that's interesting because people get bored with the same stuff over time, right?

Jason Cohen:

Yeah, people get bored, different flavors come to the forefront, exposure can happen through really interesting channels. If you look at the rise of Thai cooking in the United States, lemongrass didn't exist in the United States before Thai cooking. And now, you can find Thai restaurants even in small towns in the Midwest and that can be traced to immigration in the mid to late '90s, grants by the Thai government. They said we're going to promote Thai culinary food to increase tourism and to increase the stature of Thailand internationally and it worked. Thai people moved to the United States, opened up restaurants, promoted Thai food and Thai food is now one of the most popular types of Asian cuisines. People go to Thailand specifically on culinary tours and that was solely responsible for the introdcution of lemongrass into the United States and a lot of products that now taste like lemongrass.

Ernesto:

How do you-

Jason Cohen:

Those things don't have that clean of a story.

Ernesto:

No, a messy process that you have humans figuring out something or accidentally stumbling on a flavor or just finally convincing leadership to take a chance on adding some more spice to your Fritos. Who knew Hot Cheetos are delicious? Anyway, I am a former addict of Hot Cheetos, let's just say. So, I thought it was hilarious to come across that post. So, can we talk a little bit about the actual ... You've been at this for, you said about 10 years, working on the platform. A lot of AI startups that are flourishing right now might get a couple of million dollars here and there in early seed funding to get their feet going. But you've been at this for 10 years, what has been ... Let me ask this in two parts. What has been your greatest realization about AI as it pertains to the work that you do? And what is the platform structure that you think about that would have been better if you knew that back in the day? So, that's two questions.

Jason Cohen:

Those are great questions. That'll make a little more context in the story of the company. So, I started as a professional tea taster, that is a real job, spent a whole bunch of time in mainland China, Taiwan, Korea and Japan.

Ernesto:

You said tea taster?

Jason Cohen:

Yup.

Ernesto:

Okay, all right. Awesome.

Jason Cohen:

If anyone's watching on YouTube, that's a full tea room behind me.

Ernesto:

Oh, wow.

Jason Cohen:

Prior career but very much a personal interest. So, I then went to Penn State University where I started a tea research group that, like everything I touched, spiraled quickly out of control and became an interdisciplinary tea research institute call the Tea Institute of Penn State. That was about exactly as popular as it sounds, just really competed with the football team for funding and turnout.

Ernesto:

Yeah, right.

Jason Cohen:

Yeah. But despite that, I had 30 plus students in my field of study, I did my research originally in sensory science, science of taste and then moved to machine learning and artificial intelligence and did that for just about four years. And when everything started to work, we'd actually make predictions around what people taste and like and dislike in food and beverage products, realized it didn't belong in academia, spun it out of the university, I hired off the top three researchers from my research institute and we all started analytical flavor systems together. So, that research started way back in 2009, became a company in June 2013, then took about two years to go from research to a usable enterprise product.

Jason Cohen:

Officially, we received institutional capital in October 2016, moved the company to New York City and we've been off to the races since. So, yes, we've been doing this for a long time but it was research for the first four years and then product development, turning it in from research into a product for an additional two years, so we've really only been on markets since 2016 and we really didn't hit our growth curve, we didn't really hit our stride until 2018.

Jason Cohen:

And the two biggest things, I would say, that we learned is, one, is that bootstrapping is awful, bootstrapping a company. It worked out fine for us but I will never do that again. Sell earlier, market stuff earlier, get investors earlier. It's not easy advice to follow, it's not an easy thing to do but it probably would have accelerated our development, it probably would have brought that forward by a few years, it probably would have been able to higher up faster, to stack up faster and we wouldn't probably have an iteration time.

Jason Cohen:

On the other hand, I always say that we never pivoted, we've never changed what we set out to do, modeling perceptions and preferences to predict consumer responses and products. And some of the applications have changed. When we first started, we did a lot of work in the beer industry, particularly in the craft beer industry, and we did some work on quality monitoring, which we don't do anymore, but as identical technology. When we officially launched, we got much more into product development than any type of manufacturing processes with the same technology.

Jason Cohen:

So, would that process of discovery have happened faster if we had sold the product earlier, if we had taken investment money earlier? I think so but it's something that I have no way of testing and I have no way of knowing if that's true. But I do know that if I ever start something again, if this company does well, exits and I start something again, I will not stop bootstrap.

Ernesto:

Right, right. Most technology companies, pretty hardcore tech companies at this point, they're absorbing five to $10 million in initial. A $2 million investment, which is crazy to even say, is bootstrapping-

Jason Cohen:

Yeah.

Ernesto:

... in some capacities, especially when you're dealing with AI and material. So, it's just the amount of technical expertise you'd need to be able to design and model and put this thing together, just getting the data into the system in a structured manner that you can then work with is-

Jason Cohen:

Exactly.

Ernesto:

... is a huge challenge, right?

Jason Cohen:

Exactly. And then, our first round, our first real round was 4 million in 2018, we had raised quite a bit less than that in 2016. But I would also say that, when we have started, it was different time. When I started this research in 2009, no one was interested in food tech, very few investors were interested in niche applications of AI. I don't really think of food as that niche, everyone eats. But the thing about that-

Ernesto:

They do, right?

Jason Cohen:

They do, right. And back then, 2009 through 2016, nine years, it took nine to 10 years of additional growth in the market, the majority of investment funds in AI were flowing to things that were either slim layers or things that were generative build ups that can be used across multiple markets, they were not highly specific targeted applications of AI, that happened later and the expertise was still much less than demand. You didn't have individuals being poached from research labs or from universities with half a million dollar plus, if not more, salaries.

Jason Cohen:

So, we started, really, a forefront of niche applications of AI, building an AI that only does one thing. It's flavor, aroma and texture, perception and preference, that's it. The AI does not know anything about marketing, branding, supply chain distribution, price point, it only knows consumer sensory perception and response. And now that's not so shocking, but overnight success is 10 years in the making is what they say.

Ernesto:

Yes, yeah, it's interesting. The research phase that you were in, what were you really researching at that point? Is it how do we actually categorize all this stuff?

Jason Cohen:

It went through different phases.

Ernesto:

Mm-hmm (affirmative).

Jason Cohen:

So, very early days, it was how do you build the sensory system, how do you collect the data, how do you make the data, how do you structure the data and tons of testing just went into seeing can we predict verifiable ground truths on the data that we collect. Can I predict where this coffee is from, what country it's from? Can I predict the cultivar of coffee? Can I predict the age of the consumer? Can I predict gender of the consumer who gave us this observation? Those are all verifiable ground tests [crosstalk 00:29:00]-

Ernesto:

Wow, you can predict the age of the consumer on a data set?

Jason Cohen:

Yes.

Ernesto:

Just based on how they respond to the certain flavorings?

Jason Cohen:

Yes.

Ernesto:

Oh, my god. And that's your proof of concept for it. The system is so good that it can do this. It's not necessarily that they're at ... Well, maybe there are particular products, flavor products that are targeting certain age groups, right?

Jason Cohen:

Oh, certainly.

Ernesto:

Okay.

Jason Cohen:

And a lot of the work that we do in new product development, it could be, "We want to target post yoga millennial women in the United States." That's crazy, right?

Ernesto:

We have this vegan product thing that's like, oh, [inaudible 00:29:39]. This health and wellness, new.

Jason Cohen:

So, that's a brief [inaudible 00:29:45], yeah.

Ernesto:

That's amazing. Let me ask you this. You've been 10 years into this, have you always had this vision of this world? I mean, you certainly went to Penn State, you did your research so you were in this general direction. But did you envision you'd be here where you are right now with AI and with the work that you have?

Jason Cohen:

No, it's been slower and faster than we imagined. It's been slower in that the uptake and understanding of technology is slower. It's still very new, every company wants to learn about it, every company wants to run it through its tests and its paces. The adoption curve in the western world for technology that changes processes is shocking and stuff. And to compare that, we do a lot of work internationally. Actually we do about 25 to 50% of our work in Asia, depending on the product and-

Ernesto:

They're like, "Of course this stuff should work." They're more trusting that you're giving them the right information.

Jason Cohen:

Exactly. So, when we go into-

Ernesto:

Wow.

Jason Cohen:

... a company in the United States, particularly the United States, US and Europe, it's very much how do I know it works? Can you prove to me that you can make these predictions? Can we run a pilot test? What other publications that you have? What other companies are using this? Now, at this point, these are all questions we have answers to, it's fine. It might eat anywhere from a few weeks to a few months in order to get started.

Ernesto:

The sale cycle. Right, okay.

Jason Cohen:

The sale cycle, right. And just running the test to prove to them that we can operate on their product and their product category. When we operate in China, we do not get those questions. In China, it’s very much, they're like, "Okay, great. We're not concerned about the AI, we're not concerned about the predictability. If we go from 60 products today to 600 products next year, will you be able to handle that? What's the operational load if, instead of five predictions per week, we want to make 50 predictions per week or 500 predictions per week?

Ernesto:

Oh, my god. That's the din you're getting from the clientele in China?

Jason Cohen:

Yeah.

Ernesto:

Oh, my god.

Jason Cohen:

And so, the conversations are very, very different.

Ernesto:

What the hell? I mean, that's surprising to hear. It's probably just on the sheer bulk that everybody gets their stuff made in China, the whole world does, that they have to consider just [inaudible 00:32:33] time, they're just scaling and issues there for what it sounds like. But it also means that they just are much more trusting that the technology does as advertised than we are, which means deal-making is happening faster and you start to question the United States' ability to innovate, in a sense, right?

Jason Cohen:

Yes. In the United States, the average consumers interaction with AI is very hidden, right?

Ernesto:

Mm-hmm (affirmative).

Jason Cohen:

They interact with it using some things on Facebook and Google, is it real AI, is important AI, that's a different debate. But the interaction with the majority of the AI in the US and in the western world is very hidden. In China, you walk into a building and the cameras scan your face and it determines whether you're allowed to enter. It takes me forever to get into a building in China because my face isn't in the database, I have to go and show my passport and stuff. My Chinese colleagues walk-

Ernesto:

Right through.

Jason Cohen:

... right in. They use AI for much more public facing things. The average consumer in China has more interaction with AI than the average consumer in the West and that means that not just the young people but the executives, the managers, from the-

Ernesto:

They're used to it.

Jason Cohen:

... [crosstalk 00:34:01] all the way up to the C-Suite are used to it. And they've seen it in other industries, they've seen it in their industries and they're looking for any competitive advantage. There's none of the skepticism around will AI affect my industry, will it affect what I need to do day to day. And so, it's very much not about prove to me that it works, it's very much how quickly can we adopt it and how well will it work?

Ernesto:

Man, that sounds like there's profound implications there for our role in the world with technology in general. But maybe, on the other hand, our hesitancy, a little bit, I don't necessarily want to get scanned to walk into every building, right?

Jason Cohen:

Right.

Ernesto:

But that's a very typical Americanized perspective on privacy concerns and so forth, that's a common one. But on the flip side of all this, it is surprising and concerning to hear that your time to get to the ... We run a business here, so much of the business effort goes into the proposal, the discussions beforehand and so forth. And I've worked with folks in Mexico and in their part, typical, you have to go to dinner if you're going to make a big deal, you have to sit down, you have to have breakfast, there's a lot of more feeling things out before actually engaging. And they like to say that in America, it's very much like, "Oh, you can do this? Okay, yeah, check. Here's the order, here's the proposal," but I don't think that. What you're telling about China, it's just like let's go with AI, period. Full stop right there. Wow.

Jason Cohen:

Yeah. That's not to say that all contracting and that all operations and all sales are easy in China, certainly.

Ernesto:

No, of course not.

Jason Cohen:

to say that the questions are very different and the areas of concern are very different.

Ernesto:

Right. Wow. Okay, so let me think here. Where do you see this business that you're in? Clearly, most of them, your clientele is coming from Asia, in a sense, the growing portion of it or do you feel like you're going to eventually get into much bigger clientele here in the US? It's just there's just a lag for them to catch up.

Jason Cohen:

I'd say neither. So, where we work and where we operate is pretty complicated. We could have an American headquartered company that has development and research centers in Europe that wants us to work on a new product for Thailand. So, is that an American company? Is it-

Ernesto:

I see.

Jason Cohen:

... a UK company? Is it a Thai company?

Ernesto:

These are giant companies you pulled into whoever hears about your product.

Jason Cohen:

Exactly. So, our revenue, like I said, is probably 40, 50% US, maybe anywhere 15 to 20, South America, 15 to 20 Europe and the rest is Asia. So, Asia is not the majority but it could be the largest area competitive with United States, depending on the quarter. We are a US company, we started here first but we have one full time team member in China, we have a team member in Hong Kong, a team member in Singapore, a partner company in Japan, we've also had representation in Thailand previously. So, we do quite a bit of deployment work, even for European and American companies, in Asia.

Jason Cohen:

Where I think the business is going, and I think that this very much good thing, is that I think that it used to be that companies would try to sell one product locally. You can think of massive soda brands, massive beer brands, same product available around the world. And in some countries, there used to be local competitors. Peru had Inca Kola, India has Thumbs Up, which is a Coke competitor. And so, what we're seeing now though is we're seeing new products, either by the multinationals or by local companies, that are creating highly targeted products that are fit for a specific group or a specific area. Even in the United States, you can see that.

Jason Cohen:

For a long time, we've had really, really niche products that are regional products like Cheerwine in the Carolinas, goes great with Carolina barbecue. But that was a legacy product that remained in a very constricted geographic area. Now, what we're seeing is we're seeing products that are geared towards California, products that are geared towards Texas, products that are geared towards the Midwest and Northeast and Southeast. That type of regionalism is very new. And even beyond the regionalism, the attempt to appeal to specific consumer demographics, whether it's Hispanic-Americans in the United States or African-Americans in the United States or even certain groups of Asian-Americans in the United States, that's very new and I very much see that as more diverse products for a more diverse world. How do we get products into a consumer's hands that they're going to love? There's no reason that I need to be eating or drinking the same things as everyone else if those products don't speak to me or don't appeal to me, right?

Jason Cohen:

So, historically, the entire CPG world, the entire food world was built on creating mass market products that everyone likes. Now, it's about how do you build portfolios of products and brands that specific target consumers are going to love and they're going to love it more than any of the other products they choose at a similar price.

Ernesto:

Yeah. So, you do well, you do super well in a specific region that you can control or get the supply chains on the product delivered there, it gets super complicated at that point, right? It's the complexification of everything.

Jason Cohen:

Yeah.

Ernesto:

Can I get a bespoke product for this particular region based on the ingredient access that we have and the manufacturing process that we have versus like, "Hey, we made this thing that sells really well here in this part of the world, let's see where else we can sell it." I find that the most fascinating part of the food world evolution. You don't go into the milk section anymore and just buy the whole milk or the low fat milk, there's a bunch of other milks that are in there. And when you look at the ingredients of all that other stuff, it's super complex.

Ernesto:

It's not so simple, I would assume, for a lot of these big food manufacturers to deliver on such a variety of products. This is the evolution of computer animated, just in time sales through the production process, all of it being put into it. And I guess it's companies like you that are now have been positioned to be able to almost real time test to see if we're making an assumption, putting together a product for particular region, go to you, Jason and be like, "What do you think? Will this do well if we'd made this?" Right?

Jason Cohen:

Exactly.

Ernesto:

Wow. What's the AI technology stack that you guys work? In 2009, when you started or-

Jason Cohen:

Yeah, 2009.

Ernesto:

AWS wasn't quite there with their stuff. What do you anchor on with your system? Just curious.

Jason Cohen:

Yeah. Well, all the original research, back in the university and stuff, that was all model building and so very little was done on hosting or software development or anything. So, by the time we started 2013 with software development, that was all on AWS.

Ernesto:

Okay.

Jason Cohen:

The models, a lot of companies, most of the companies that claim to be machine learning or AI companies aren't. And of the ones that are, the majority are using either pre-built models or pre-built model architecture. Things like Keras, TensorFlow, Spark, Nifi. And there are companies out there and they can pump their data into logistic regression and get meaningful results, amazing, more power to you. There are companies out there that have terabytes of data and need deep neural networks trained on GPUs by TensorFlow and served up on Tensor serve, amazing, more power to you.

Ernesto:

Great.

Jason Cohen:

That's not us.

Ernesto:

That's not you guys.

Jason Cohen:

Yeah.

Ernesto:

You have a unique model that's been mapped from the start.

Jason Cohen:

Exactly. So, we build all of our own models from scratch. We run the open-source metric learning package and the Arc programming language. The majority of our work is topological in nature, building meaningful pairwise distance metrics in high dimensional Non-Euclidean space.

Ernesto:

This is where you're going to lose some of my audience, but that's okay. We don't have a ton of people watching.

Jason Cohen:

And it's actually very, very intuitive because what that means, when someone's making a lemon soda, no one is squeezing lemons into that soda. So, you can't talk about how much lemon juice and you can't talk about how many lemons or how many lemon juices. So, then you say, "Okay. Well, it's lemon flavor, what about the intensity of the lemon compounds?" Well, actually, you're going to have firmenich lemon two, juvenile lemon five, tons of different lemon flavors that could achieve that. And those lemon flavors are going to have different intensities at the same molarity, which means you can't talk about amount of lemon compound. So, what we do is we actually say, "Okay, well, lemon is a vector in flavor space and the movement along with the lemon vector means that consumers taste more or less lemon. And that lemon vector could be parallel or covariant or slightly correlated with the citrus vector or the grapefruit vector or the lime vector." And so, by mapping out on thousands of these vector spaces. We can-

Ernesto:

The thing that people call lemon are several different things, is what you're saying?

Jason Cohen:

Exactly, exactly. Right.

Ernesto:

But you're mapping it out so you're capturing their preferences, they're likely to react in any of those areas?

Jason Cohen:

Yeah. And then we place each product in that high dimensional space, in that flavor space and we can learn the relationship between products.

Ernesto:

Holy crap. Would you license this model out? Would you then go to the Tensor world and be like, "Here's the model that works really well and"-

Jason Cohen:

No, that's [crosstalk 00:45:11].

Ernesto:

That so much proprietary. That's your nine years of effort, lived effort.

Jason Cohen:

[crosstalk 00:45:16]

Ernesto:

Right.

Jason Cohen:

It's the data that led to that, that led all of the models that we build on top of it. So, that's-

Ernesto:

And that's your competitive edge than anybody else, perhaps, doing something similar in your particular space, this is the research?

Jason Cohen:

Yeah, exactly. They would need both the longitudinal data, they would need the experience of working with the data and then they would need to build models on top of the transformed data. So, we would say that we have a pretty good, competitive moat right now, not a perfect competitive moat, there's, of course, areas under which we're very keenly aware that groups can attempt to compete with us. Some of the most common ones are hardware. I'm not believer in hardware for human sensory perception. We've had HPLC and GCMS, we've known exactly what's inside products since the mid-1950s or early 1960s, it's never been predictive of perception. There's groups, they're trying to do biosensors, response with ganglia for perceptual work, gene protein receptor sites, those types of things.

Ernesto:

That's different though. That's a different direction for-

Jason Cohen:

Yeah, it's a different direction and we don't really believe it's competitive with us. We joke, someone has asked us in the past, "What is the AI?" Whenever someone asks that, [inaudible 00:46:53], "It's brains, we're growing brains in vats and giving them" ... But the groups that are doing biosensors, that's quite literally what they would need to do in order to be able to get full spectrum perceptual responses. So, I don't really quite believe in that but there are other things. We don't do work on the behavioral side, we don't do work on people who watch certain TV shows or attend certain sporting events or go to certain activities like yoga, we don't do behavioral side work right now. And so, that's an area of open competition with us that-

Ernesto:

Well, maybe they could license your part of it and then they do the actual behavior part, adding to it, right?

Jason Cohen:

Yeah.

Ernesto:

Maybe that could be a future.

Jason Cohen:

It definitely could be. But no, we wouldn't license the core model.

Ernesto:

The core model, you wouldn't actually create a copy of the model into it. You wouldn't do that.

Jason Cohen:

Yeah, yeah.

Ernesto:

Right. It's come talk to us for [inaudible 00:48:01]. Look, forget the exit, I mean, why not head towards potential future IPO? Why not get to this level? Because if this is working and you're 10 years ahead, I don't know anything about your space, I wouldn't even know who is a competitor, but is that door a possibility out there for you guys? Is that what you're-

Jason Cohen:

It could be when we were certainly not thinking of exits right now. We're doing well, we're in the earliest stages of our gigantic hockey stick inflection, things are, right now, quite good. So, we're certainly not thinking of an exit. When I think of something like an IPO, which is certainly a few years in the future, who knows what the market looks like at that time. But looking at it right now, I think what we do is relatively difficult to understand, particularly from an institutional investor side, a public company, [inaudible 00:49:10] market, institutional investor side. And-

Ernesto:

There's a bottleneck there, in a sense.

Jason Cohen:

Yeah.

Ernesto:

Okay.

Jason Cohen:

And I would assume that we would trade a lower PE ratio than similar AI tech companies or AI tech companies in other industries.

Ernesto:

So, you're looking? You're looking? You're keeping an eye on what's happening with Tesla, you're looking at what happens with maybe 23andme, you're looking at Palantir.

Jason Cohen:

Yeah. And those companies, if we could achieve what those companies are achieving, that would be very attractive, right?

Ernesto:

Right.

Jason Cohen:

But realistically, we look at other companies in the space, whether they are flavor house companies, whether they are suppliers or producers of ingredients or other industry service providers for CPG, those tend to trade at a price discount to standard PE ratios partially because it's not an area where the analysts and big banks and investment firms spend a lot of time and effort.

Jason Cohen:

So, if we could change that, then an IPO can be very attractive.

Ernesto:

Mm-hmm (affirmative).

Jason Cohen:

Maybe running a public company less so but-

Ernesto:

Right. That's a whole other conversation on that one. But yeah, I just think that for most, I think we're getting past the stage of like, "Oh, AI, that's an amazing thing. Do you do AI, too?" And now really digging under the levels of what that actually incorporates and it's just much more complex and people start to peel back the onion on what that all means, that's when they quickly run aground into like, "I don't know enough about what all this is." But it's just do I know enough to be able to distinguish the product value that you have, the model that you have, from what somebody else is telling me like, "Yeah, yeah, we do AI too on our system, what do you need?"

Ernesto:

That kind of approach, right? That part, perhaps, it's taking time, but maybe in the next few years, maybe in the next generation of decision makers that are filling in these executive spots at some of these food manufacturing places are going to be more prepared to be able to understand the value. So, man, thank you so much. That was an hour, that just flew by, Jason. If somebody out there in the food industry, I know there's a couple, for sure, that listen to the show that are just curious about this, want to find out more, where should they go about you?

Jason Cohen:

Yeah, our website is gastrograph.com. Company is Analytical Flavor Systems, product is Gastrograph AI, all of our email addresses and everything are at gastrograph.com. That would be great, very happy for them to reach out. We have tons of public-facing technical white papers, we have tons of internal technical white papers for when conversations become more serious. We take a very strong stance that the AI must be explainable. A lot of companies say, "Look, we have this magical black box, AI, trust us," and I don't and I don't recommend that you do. We-

Ernesto:

But they have to set aside the time to talk to you or they have to truly be somebody that has to be there to sit down with you to go through it, if they want, right?

Jason Cohen:

If they want to.

Ernesto:

If they want to.

Jason Cohen:

If they want to, yeah.

Ernesto:

Right, right.

Jason Cohen:

Because we take the stance that it has to be understood in order to be trusted and it has to be trusted in order to be used. So, we're not going to give away the algorithm but we will explain exactly what's going on, how it works, why it works, how it's possible, what our proof is, what tests we run to validate it.

Ernesto:

And the potential outcomes for them to get it right, it could mean millions of dollars. So, getting a product into one locality, especially if it's something overseas, I can't even imagine what the costs are just to put it all together and then deliver shipments and then the distribution and all of that. So, why not first check with the company that's got an AI to see if people even like what you put together, right?

Jason Cohen:

Yeah.

Ernesto:

Run it by you guys. So, thank you so much. Best of luck. Hang on, I want to talk to you off mic a little bit. But those of you guys listening to it, hopefully we ... Go to the blog because it's fun to read through some of that and it's not deep in the jargon of trying to figure out, just very nicely written explainers to just see just how sophisticated Gastrograph really is. Commend you and best of luck, Jason.

Jason Cohen:

Thank you.

Speaker 1:

For more episodes of Through the Noise, go to throughthenoise.us.

 

Create a better-tasting world

Get in touch to speak with our team so we can learn about the challenges you're facing.