Podcasts

In big data and artificial intelligence, one of the most well-recognized challenges to success is the “cold start problem.”

The cold start problem refers to when a lack of data hobbles recommender systems in machine learning models. Much like a cold car engine that causes a car to sputter and jerk along as a driver starts their journey, an algorithm built to discover and make accurate recommendations can’t perform well when it starts cold with a foundation of little to no good data.

And it’s this problem – a lack of foundational data around which to build a machine learning model – that often deters scientists, entrepreneurs, and companies across various fields from adopting new technology such as artificial intelligence.

The cold start problem is something Climax CEO Oliver Zahn was well-familiar with. As a world-recognized astrophysicist who worked for Google and SpaceX building complex data science models, Zahn knew that getting over this initial hurdle was one of the reasons established companies didn’t embrace machine learning and continue using the status quo – whatever that may be – to build new products.

So when Zahn decided he wanted to build a future food company using AI, he knew the initial challenge of building a dataset that could be mined to find new and promising building blocks in the world of plants would be his biggest hurdle. Still, it was a challenge he knew was worth taking.

“Traditionally, a lot of the big food companies around today pursue sort of a trial and error approach,” Zahn told me recently when we sat down for our conversation on The Spoon Podcast. “They use human intuition to guess what might work. But that often misses things that are less obvious.”

Zahn knew that the less obvious things could be the key to unlocking food building blocks that could power new types of food. Those building blocks, which come from the hundreds of thousands of different plants – many of them inedible – could then be combined in millions of different ways to provide new functional or sensory features to create something like a plant-based cheese. The only way to get there was to use machine learning, cold start problem or not.

“It’s a huge combinatorial screening problem,” said Zahn. “Even the largest food labs on Earth, if they all joined forces, would not be able to explore all combinations and millions of years.”

He knew AI could if he could get past those initial hurdles. But to do that, he knew Climax would have to begin not by gathering lots of data first on plants but on animal products.

“We started by interrogating animal products really deeply to try and understand what makes animal products tick the way they do,” said Zahn. “Why do they have their unique flavor profile texture profiles? Their mouthfeel? Why do they sizzle? Why do they melt and stretch when you eat them?”

You’d think that a lot of that data would already exist, but according to Zahn, it didn’t. The reason for that, he explained, was there had never been a business reason to build those datasets. But as the environmental impact of animal-based products became more apparent in recent years, there was a business motivation to start understanding how these products ticked so they could then be replicated using more sustainable inputs.

The data the company gathered by interrogating animal products allowed them to create labels for their machine-learning models to describe and characterize a food product accurately. With that in hand, Zahn said the company set about building data sets around plant-based building blocks.

“We built a lot of data sets on plant ingredient functionalities and the different ways of combining them. We then found these trends that can recreate animal products more closely, and sometimes in very non-obvious ways.”

Zahn says the process of creating accurate models can often take a very long time – up to 20 years – particularly if those building them don’t have the good intuition that comes with experience in machine learning.

“From the perspective of somebody starting a food company, that (long time horizons) can be scary, right? Because you need to get to market at some point. And so unless you have a very good intuition and have a lot of experience, in my case, a couple of decades, of trying to derive meaning from messy, large data sets, people don’t even start.”

For Zahn and Climax, the models they have built have already started yielding impressive results, enough to help them begin making what will be their first product – cheese – using artificial intelligence. What helped them get there so quickly was Zahn’s experience in building these models that told him to start with trying to understand and describe certain features of animal products – be it taste, mouthfeel, or nutritional benefit – and then find combinations of plant-based building blocks that achieved the same result.

“To look in the plant kingdom for something that is chemically identical to the animal ingredient, like a protein that you might be after, is a little bit of a red herring,” said Zahn. “Because it doesn’t need to look identical microscopically, or the sequence doesn’t need to be identical, for it to behave the same. There could be other ways to accomplish the same functionality.”

Now, after just two and a half years, Climax is ready to start rolling out its first products, a lineup of cheese that includes brie, blue cheese, feta, and chèvre (goat cheese) made from plant-based inputs. It’s an impressive feat, partly because, as a first-time entrepreneur, Zahn also faced the challenge of learning how to build a company, in itself another “cold start problem.”

If you’d like to hear the full story of Zahn and Climax Foods’ journey to building plant-based dairy products, you can do so by listening to our conversation on this week’s episode of The Spoon podcast. Click play below or find it on Apple Podcasts, Spotify, or wherever you get your podcasts.