The evolution of Giuseppe, an AI system that generates fake meat and dairy recipes

NotCo makes plant-based versions of beloved foods, like milk, meat, and mayo. Founded in 2015, the Chile-based startup has raised $360 million dollars to date—most of which came via a $235 million Series D in late July—and is now valued at $1.5 billion. Its sales grew by 3x annually over the last four years, though the company won’t disclose how much it pulls in.

At the heart of NotCo’s operations is Giuseppe, an AI system named after Giuseppe Arcimboldo, a 16th century Italian artist known for painting portraits of fruit- and vegetable-based humans. NotCo’s Giuseppe predicts which combinations of plants will add up to animal-like outputs.

A team of more than 20 food scientists and chefs work hand-in-oven-mitt with the system, testing out Giuseppe’s educated guesses in the real world, adjusting them as needed—a little more crunch here, a little more color there—and feeding their changes back into the system, creating an ever-improving loop of mock meat and dairy. NotCo’s food scientists and chefs test more than 100 recipes each month.

Giuseppe started as a modest model, trained on data scraped from the web and focused on a limited scope: recipe prediction. Since then, NotCo has bought a bunch of scientific equipment, like spectrometry and texturometer machines, in order to train Giuseppe on more precise, scientific data. Now, the team has Giuseppe working on bigger challenges, like analyzing meat protein structures to perfectly emulate, say, a ribeye.

We spoke with Karim Pichara, who created Giuseppe and is also cofounder and chief technical officer of NotCo, in order to learn more about how Giuseppe has evolved, how NotCo’s chefs bring the AI system’s predictions to life, and what the most challenging thing about replicating animal-based products with a bunch of plants is.

Tell me about the creation of Giuseppe, the AI system at the heart of NotCo’s plant-based products. How did it come to be?

To change the food industry, we realized that the major problem was that we keep exploiting the animals to produce our food. But then we also were discussing that, in the end, people are not going to stop eating what they eat, unless you offer them a really good alternative.

So we said, “Why don't we create food products that are plant-based, but also are really close to the animal version in texture and flavor?” Of course, it is not easy to come up with a solution, but we had the intuition that if we have an AI system that tells us how to combine ingredients—and what ingredients to combine—we could maybe find a match to the animal alternatives.

The first version of Giuseppe was an AI system that was generating, according to a given animal-based target, a plant-based formula that mimics the animal in flavor and texture. And the initial team of chefs was basically one chef that was taking the output from Giuseppe and actually cooking in the kitchen.

And so, in the initial version, was the chef taking every recommendation from Giuseppe and cooking it? Like one-to-one, if Giuseppe said try this, they would just test it out?

Yeah, yeah, exactly. I mean, we envisioned the technology as an assistive technology and an assistive technology must learn from its mistakes over time and also generate or teach something new to the user. So it's like an interaction where every party involved is getting something and also delivering something to the result.

And since the very beginning of NotCo, we started generating data. That means the chef was cooking the recipes, but also immediately suggesting improvements or alternatives with their sensorial feedback, such that we started populating a new database where we can connect recipes with sensorial feedback.

But that is only at the culinary level, in the kitchen. You cannot claim that you have a product until the product is able to be produced at scale. You can have a prototype that the chefs could cook in the kitchen, but then if you use the same prototype with the same recipe and the same ingredients at scale, it’s not going to work because the mechanics and dynamics involved in the production pipelines are totally different.

Today, our technology—besides assisting in the creation of the initial recipes—helps the food scientists and food engineers to scale up. Which means, like, how do you improve the texture of a given food in order to make it resistant to the industrial pipeline? How do you improve the color? How do you supplement the flavor with natural ingredients, such that you keep the same sensorial perception after the process is at scale?

How has the type of data you feed into Giuseppe changed over time?

We started also generating scientific data. We realized that it's not just about sensory perception in order to map from the ingredients to the final result. We needed to see the food from every single different perspective we could. So we started by acquiring spectrometry data, texturometer, viscosity, pH, all the types of compounds, and a bunch of different scientific equipment that scans the food from different perspectives.

So today we have all these data sources being integrated and used by Giuseppe to represent the food ingredients. Giuseppe understands by itself that, if you need to improve the texture, you're going to have to pay attention more to the textural analysis machine, and then you're going to have to be able to achieve a given texture profile, by recombining some ingredients, and also by paying attention to some of the variables defined by the food scientist.

Let me give you an example: Let's say if a scientist needs to make the texture more crunchy, and the food scientist uses, in the lab, the most suitable machines to scan the texture, and then asks Giuseppe to get closer to a given target. And also says, you know, I would like you to rebalance the amount of oil and starches, and the pea protein and so on, in order to achieve the desired texture profile.

And Giuseppe assists during this experimentation process to get to the desired target way faster than if the food scientist was doing this by trial and error.

So now you have a bunch of precise, verifiable data from these machines. You can scan a piece of pasta and feed the data directly into Giuseppe. But in the original model—before you had the fancy machines—what information was Giuseppe based on?

At the very beginning, just to start with something, we scraped data from the web. We gathered all the information that we could from the web—macro-, micronutrients, physical, chemical features, and a bunch of information that was out there, kind of spread in papers and on websites.

And we gave an initial structure to that data and we put it all together. We had the initial database to sort of kick off the first model, and it was actually working, but then we realized that we needed to supplement it with new information, quickly.

And with the introduction of better and more data over time, have things sped up?

Yes, definitely. That is one of the main aspects where you improve a lot. You improve orders of magnitude because you get more data. And sometimes it's not just about more entries, but we also need new variables, new scientific machines that add new variables that supplement with better perspectives.

Our new products come out to the market way faster compared to the initial products. So, yes, that is one of the natural results of having a technology that is improving over time—it’s the speed you gain.

[Ed. note: On its website, NotCo claims improvements in Giuseppe enabled the company to develop a milk product in 2 months, compared to 10 months the first time around.]

This conversation has been edited for brevity and clarity.