Thank you. Welcome to NVIDIA's AI podcast, where we explore the expanding world of artificial intelligence with the people who are developing and applying AI to the problems and opportunities that cut across our lives.
And also the joys, which include thanks to our guest, Jason Cohen.
Jason is the founder and CEO of analytical flavor systems based in New York city. which applies AI to the problem of analyzing flavors in beer and really all kinds of food and drink.
Jason, welcome. Thanks for having me. I'm happy to be here.
Jason, we all, most of us, I should say, I'll hazard to say, love beer and other things.
How did you get to this idea that you could apply AI to the issue of flavor and what we eat and drink?
Yeah, so I started as a professional tea taster and I ran a interdisciplinary research institute at Penn State University called the Tea Institute at Penn State.
And I was working on the way that people think and talk about flavor.
And what I realized is that there weren't any models, there wasn't any way to understand this, that flavor had historically been written off as a subjective taste unlike sight, unlike sound. and I didn't believe it and Mark Zuckerberg had a great quote recently where he said in the next few years artificial intelligence is going to be able to model all of our senses.
He talks about sight, he talks about sound, and he talks about using that for music or for art. that's what we're doing for flavor and it and it came about because there wasn't anything that existed before and we needed a way to be able to measure and understand what people were tasting in order to be able to make improvements and measurements on food and beverage products.
Well, okay, so let's back up. You say you were a professional tea taster, which I didn't know that that existed.
Let's start there, but... I can imagine you sipping cups of tea and spitting it into buckets, but I can't imagine you then saying, okay, let's write a program to do what I'm doing.
Yeah, and it's not so much about programmatically... tasting tea in like an electric tongue or in a technology kind of way.
It's about developing systems to record what I think or I believe that I'm tasting.
And then to be able to say, well, this is a snapshot of this tea in time.
In a year from now, when this harvest is over, the next harvest exists, Is it, does it taste the same?
Does it taste better? Does it taste worse?
If you don't have a system, then you have to rely on your memory and your memory of flavor is incredibly faulty.
If you ask someone to try to explain what they tasted a week ago or a month ago or a year ago, they're not going to give you an accurate answer.
And then when you look at that in the industry, when that comes to beer flavor profiles, when that comes to chocolate flavor profiles, things that people expect to remain consistent within a brand. that's a real problem.
It means that you can have spec drift. It means that you can have dispreferences developing.
It means that consumers, the people who like your brand, may not like it in the future.
They may not trust you to produce the right flavor in the future.
What you were trying to do was take this ephemeral thing, this taste and this memory of it, and... and kind of operationalize it or programatize it and make it last, it sounds like.
And how did it occur to you that AI was the way to do it?
Well, people have been trying to understand flavor for years.
Sensory science, the science of tasting, started in earnest in the early 60s, some people might say a little earlier.
And it revolved around frequentist statistical hypothesis testing.
And it hasn't really grown from there. They're using some things that they think of as advanced techniques like MANOVAs or PCA or ICA. but it was barely explanatory.
And if the measure of any other science is its abilities to make predictions about the future, then sensory science wasn't working.
And so when I got involved in it and I started to measure and predict tea flavors and then beer flavors and then coffee flavors, it became obvious that I couldn't make predictions without actually modeling human sensory perception.
And as we started to build those models, we realized we could not only model what people were tasting, but what people like, people's preferences.
And that's where the real value comes from.
Understanding if you have these preferences for these flavors today, that might evolve or that might change in the future.
So when you look out at the industry, whether it's the tea industry or whether it's any other food and beverage industry, something crazy like 95% of new products fail, either in the conception phase or the R&D phase, or in the go-to-market phase.
That's because these companies can't predict consumer preference because they don't have a measurement of what consumers are going to taste in that product.
What you guys do, you have something called the gastrograph then.
Let's back up and explain to us how this is an AI a smart system or a platform?
It sounds to me like the input, the data is still coming from people at least initially.
And so you have all this feedback from people and what they've experienced and how they've described it.
But I still have a hard time kind of syncing up what I do when I take a sip of beer and I think about how it tastes to then predicting how this beer will taste to other people.
Yeah, so there's two sides to the system.
One is what we call predictive manufacturing.
So that's training the models and the AI, the models that the AI builds to understand where flavors are created, modified, and destroyed in a manufacturing process.
So in beer, for example, breweries have a bunch of levers that they can pull to change the flavor profile.
They may want to change the flavor profile to make the product better or to develop a new product.
Or they might want to fine-tune that control to make sure that they're producing the same flavor every time, even as their raw ingredients like... hops or like malt change.
So, That's the first side of the system.
The second side of the system is the ability to do that, which is... understanding what is the artificial intelligence model of human sensory perception.
So on that side, what the AI does is it understands the physiological and perceptorial differences across age, sex, race, socioeconomic status, tasting experiences.
And what that allows the AI to do is, given any one or two or three individuals who taste the product, is to project that out, to translate it and project that to other demographics and other preferences and other individuals.
I see, so you have me as a sort of a type or a demographic and then I give you my feedback.
But it's still, it's not as if, or at least maybe not yet, that, you know, You're pouring a cup of beer into some machine and it's analyzing it and tasting itself.
Or is that where we're headed? No. So we are not bullish on electric tongue technology or electric nose technology. because for pretend for a moment that those things worked and they could actually give you an accurate representation of, say, the chemical composition of a product, like what a GCMS gas chromatography mass spec or a high-performance liquid chromatography machine could do.
That does not predict flavor. You cannot right now predict the flavor of a product from chemistry. because the number of interactions, the halo effects, the masking effects, the time intensity effects, are just exponential, right?
The flavor is not a monotonic linear combination of normally distributed attributes that can be predicted from a chemical composition. wait say that again flavor is not a monotonically a monotonic linear combination of normally distributed attributes Okay, so when you don't like something at the dinner table next time, that's what you get to say to whomever has served it to you.
Yay. you get to say, you know, if you had the right model for what I was going to taste, you would have known to...
You only get to say that once, I think, and then you're never going to get fed again.
So just be warned. So when that happens, when you're looking at the chemical composition right now, and I don't believe it ever will be predictive for fluorescence, flavor and those tests are are expensive and even if the price falls let's say that suddenly you can say okay this is what they taste You still need humans to tell you if there is a preference or dispreference for that.
So the machines right now can't model flavor.
And if they can't model flavor, they can't model preferences.
And assuming that they could eventually model flavor, we do not believe that they could model preferences because what would the feedback loop be?
A human would still have to taste it. And if you're having a human taste it, then preferences are a moving target anyway, because it changes every flavor experience that you have. changes your perception, which changes your future preferences.
So we are not bullish on removing humans from the loop.
From a tasting perspective, Humans are already working on the production lines, they already have opinions, and the human palette has million years of evolution to be perfectly predictive for what humans like today and in the future. you are building models that kind of, like you say, are predictive that encompass or, take that human input and put it into a model so therefore in the manufacturing side, I get to turn the knobs to ideally appeal to that particular product taste profile, is that fair to say?
That is fair to say. Okay, I see. So where does it go?
And how is your gastrograph and what you guys are doing being used today?
So it's being used for new product development.
That's a huge one. So... like i said 95 of new food and beverage products fail it's because food and beverage companies don't have a map of consumer preference consumer preference is constantly changing for what people want to taste.
In beer, if you think back 10 years ago, No one drank IPAs.
IPAs were not a market-winning style. And then suddenly there was a global preference shift.
And IPAs went from a negligible, not even 1% of beer sales, to now they're the leading style category.
And when you look at how the large CPGs responded, they all shrugged in the early days and they said, it's just a fad.
And now, obviously, it wasn't a fad. They missed out on being market leaders in IPA, and they gave that up to the craft industry, which has produced...
Now over 1,000 IPAs are available in national distribution on the market.
So being able to help large and small companies Brewers or chocolate makers or coffee makers alike predict where preferences are going to go can lead to new, successful, better tasting products for everyone.
So that's one of the biggest applications.
Will this get to a point where I get to have my, you know, we talk about personalized medicine or precision medicine.
Is there going to be precision beer or personalized beer or chocolate for that matter, or tea?
Will I get to sort of have my own profile and then dial in exactly what I want or what I think I want.
So two sides to that question. On the first side, absolutely.
Where mass manufacturing will not be a thing in...
I don't know, 50, 25, 50, or 100 years from now.
In food and beverage, I think it's going to come sooner.
Everyone has their own special preferences, right?
We can model flavors. People like to think that they're a unique snowflake and that everyone tastes differently.
That's not quite true. people fall into demographic-tasting populations and they fall into preference archetypes.
That's why you can have market-winning products like Coca-Cola for soda or Budweiser for beer.
And those are great products. but when you get into it that doesn't mean that it's the best product for you and what's happening is that this technology was going to allow for personalized products at the individual level you'll be able to store your profile on your cell phone or on a wearable device and walk up to a machine in the future, maybe something like the uh, the Coke freestyle or the Pepsi next machine and get a product that's perfectly engineered for you.
Um, The other side of your question is really interesting.
You said, can I tune it myself? Or will I be able to tune in what I want?
The problem there is that Thank you. even flavor experts, right?
Even I couldn't walk up to a machine with all the knobs and know this is what's going to taste good.
This is what I'm going to want. Yeah, it's hard because, I mean, I...
Like you say, I think we all think we have these unique wants and desires and taste profiles, but I don't know.
Do we really know what we like until we taste it?
Absolutely not. And that's why you need a predictive AI in order to be able to figure that out. maybe the AI will ask you one or two questions.
Are you in the mood for something crisp or something heavy?
Are you in the mood for something fruity or something sour?
And from there, it will be able to craft you the perfect product, and it'll continuously learn your preferences, and it'll predict how your preferences will change as you try these new products.
So as soon as we get that into a reinforcement loop, and as soon as you start purchasing products from the machines that our AI enables in the future, you'll be having the best product for that category that you've ever had. basically every time you walk up to the machine.
And that's an amazing place to be because we're talking about pushing out the manufacturing or the blending of these products to create flavor profiles unique for an individual right to the edges, right to where they're always available.
And that's an amazing place to be because that type of personalization is really the future of food and beverage.
Does it get rid of this tyranny of adjectives?
I mean, I sort of can't stand... we all describe things somewhat differently.
Like you say, we kind of have different tastes, but whether it's tea or beer or wine, there's this like, oh, I feel it tastes, it's butterscotchy or hoppy or, you know, and sometimes those things mean nothing because we all have different ways of describing things. personalized and kind of unique as things get, does it also at some level standardize I don't know the components or the flavors so we can all understand what we're talking about.
So we do work on flavor profile communication.
So for example, right now with a bunch of our larger clients, we run something called cognitive marketing.
And cognitive marketing is... Instead of saying things like hints of elderberry and pine, which is chosen often by a marketing team, it actually understands what consumers will taste, and so it can pick out the most positive attributes.
Let's say that you were marketing this product or this product was targeted at white 25-30
30-year-old college-educated males who don't smoke.
That's a fairly large population that makes up a lot of the craft beer drinking population.
If you were going to market your IPA, you could choose any sort of terms.
You could say this has notes of pine, this has notes of oak, this has notes of hops, right?
But are any of those things meaningful? Will any of those prime a consumer to focus on what's good about that product? if you're doing this at random if you're allowing your marketing team to choose it if you're allowing your brewmaster to choose it then probably not.
That's probably not the right terms to prime someone to focus on.
So what the AI can do is it can pull out the terms that are most positive, that are going to be most enjoyed.
And just like when people go to a guided tasting, the reason that they go is so that someone who is more knowledgeable than them can point out the subtleties of the nuances and what's interesting about that product to focus on.
Being able to do that on a label, being able to do that in a written description is a very new thing that's never been done before.
And so for our AI to be able to do that and say, hey, to the producers, describe this product as citrus and... say, peach, right?
And that will actually be the positive attributes of this product that people will like.
The producers love that because that increases repeat in in-brand purchases because properly primed consumers will enjoy the product more.
And on the other side, the consumer is having a better experience.
They know what's special about that product.
And I think also have a better sense of what they're about to get, too.
Exactly. I just don't know. I know what peach and citrus means, I think.
Elderberry, I have a little harder time.
I start thinking of like Renaissance fairs, elderberry wine or something anyway.
Who's pushing hardest? Who's adopting this and who's most interested in pushing this forward?
Yeah, I would say that about 25% of our clients are in the craft industry.
So smaller, more upstart... producers of beer, of chocolate or of coffee that want to maintain a competitive advantage Their goal is to produce things at the highest quality, even if it's at a higher price point.
And they love being able to see their consistency or they love being able to see... how products and preferences are changing.
On the other side, the companies that are making the most use of our predictive manufacturing and our deep market insights and cognitive marketing are much, much larger industries.
So we right now work in beer, we work in chocolate, we work in coffee.
We're just getting started in wine and then spirits.
Spirits is a big one because global spirit preferences We went from having no craft distilling to lots of craft distilling.
American bourbon and American whiskey are on the rise.
Vodka share is falling as people turn to a more flavorful, more nuanced, higher price point alcohols.
I mean, I can see this cutting across everything we pretty much put in our mouths.
At some level, even... anything that we smell for that matter too.
Anything that's very sensory, let's put it that way.
Exactly. Let's say that you guys, what you're doing becomes the norm.
How does this then change our experience when we walk into either a store to buy something, a restaurant for that matter, it might show up on a menu, or even a bar or our own shopping carts.
How does this change our experience and what will it start to be like, do you think?
So I think that in prepared foods or in cooked foods, I think that there's still a lot of room for chefs to experiment and to make new and innovative and interesting dishes. that people will like and seek out.
But when it comes to packaged foods, when it comes to branded foods, I think that there's a lot of work that needs to be done first around consistency. so that consumers know what to expect and so that they can build an affinity and a preference for that brand.
As new products come out that better and better and better meet consumer preference, that's going to become harder to do.
So the value proposition is going to have to change.
I don't think in two or three years from now anyone's going to walk into a wine or beer store and stare at a shelf blankly. and wonder, I wonder which one of these I should choose.
I wonder which IPA I will like. I wonder which red wine I will like.
Because we'll have more information on that shelf or because we'll have a better sense of our own preferences.
Two sides. I think one of the first things that's going to happen is that Some companies are going to get smart and they're going to use not just a recommendation engine because These recommendation engines have been around for a long time and they haven't done much to the food and beverage industry, but there will be an actual personalized recommendation engine that feeds back. data to the producers.
And so if I'm a beer company and I can run this recommendation engine, recommend things to consumers, consumers tell me how much they like it, I can develop better beers based off of it, you're going to see the fit the product market fit for consumer preference get much, much tighter.
And that's going to trigger a much more competitive market marketplace where consumers are going to be able to find products that they truly love.
And that's what's going to trigger the conversion over to product personalization. to the creation of personalized products at the edges, not anymore in branded bottles, but coming directly out of machines.
And so I might walk in with my phone and it's sort of my profile is obviously... you know, contained at some level, whether it's on my phone locally or in the cloud, but then all of these things kind of light up to say that like this fits your profile.
Is that, is that, Fair to say? Yeah, that might be the first stage.
And then eventually you'll walk up to a machine, hold up your phone, Maybe it'll ask you your price point for the day or for that night, and it will craft a product that's perfect for you in that moment.
And I can even see that being a social thing where if you're out with three or four friends and you want to have a shared experience.
It can find the commonality amongst your preferences so that it can craft a beer or craft a wine or craft a chocolate that's perfect for all of you to share.
And that's a pretty amazing thing because now you don't have to say, well, I like this one and I like this one and wind up bringing a six-pack to the party that only one... one or two people like.
Well, and that, as we all know, is a horrible party foul.
So if you can do just that and... Wipe the earth of that kind of mistake.
You will have done us all a good service.
Well, I can't wait for this future to roll out more broadly and, and, um, We'll be keeping an eye on it.
And Jason, thanks for your work. I would toast you with a beer, but it's a little bit early here.
And so I will do that later. Jason, thank you so much.
This has been awesome. Really fun podcast.
Postscript here. If you are listening to this podcast on iTunes, Google Play Music, or whatever your favorite podcast player is, do us a favor and leave us a review.
Thanks for listening.
