Coffee and Beef Share 102 Flavor Compounds. Shrimp and Lemon Share 9. | Edition 306

Two companies set out to do the same audacious thing: let a machine design food. One produced a blue cheese that blind-tasting judges named a winner. The other built an AI that invents, formulates and optimizes new products, and patented it.

Neither of them is trying to sell you food anymore.

That is not a hit piece and it is not a hype piece. It is the most useful frame available for every AI in your kitchen claim you will hear this year, because it forces the only question that matters: who is making the claim, and can you check it?

Most of them, it turns out, you cannot. But the single most checkable thing anyone has ever published about why certain foods taste good together is free, fifteen years old, and testable tonight with what is already in your kitchen.

The one AI food claim that got judged by someone else

Climax Foods was founded in Berkeley by Oliver Zahn, an astrophysicist who had worked as a data scientist at Google. The pitch was that cheese is a solvable data problem: measure what makes dairy behave the way it does, then search plant proteins and fats for a combination that lands in the same place.

In 2024 that pitch got tested by someone other than the company. Climax Blue, its plant-based blue cheese, was named a winner at the Good Food Awards — a competition decided by blind tasting, with a panel the foundation currently describes as nearly 300 industry leaders, technical experts, grocers, chefs and food writers. Blind means the entrant's identity is concealed from the panel.

Then, on 22 April 2024, a week before the 29 April ceremony, the Good Food Foundation disqualified it. Two reasons were given: the cheese contained kokum butter, which the foundation said may not be FDA GRAS-approved, and the product was not retail ready, being available only through foodservice.

Neither reason is about taste. Executive director Sarah Weiner told AgFunderNews: “We will then disqualify an entry if we've determined it doesn't meet the standards and/or rules, regardless of timing.” Zahn's response: “This could so easily have been rectified if we had been contacted earlier. But they never contacted us to remedy the situation.”

So the honest scoreboard reads: an AI-designed cheese cleared a blind-tasting panel, and then failed a paperwork rule. Both halves are true, and most coverage keeps only the half it prefers.

What happened to the cheese

Production of Climax Blue was paused during 2025. On 6 October 2025 the company rebranded as Bettani Farms, closed an initial $6.5 million Series A led by S2G Investments, and installed a new chief executive, Sandeep Patel, previously chief financial officer of Califia Farms. Patel confirmed that founder Oliver Zahn “is no longer a part of the team.”

The new company is not chasing award-winning blue cheese. It is selling a plant protein called Caseed, designed to do casein's job, to other manufacturers. “Mozzarella and other mass market cheeses like jack and cheddar are where the unique stretch, melt, and mouthfeel of our Caseed protein can truly shine,” Patel said.

Read that as a business decision rather than a verdict on the technology. A specialty blue cheese is a hard thing to scale; a functional protein sold into frozen pizza is not. But the practical consequence for you is blunt: the most independently validated AI-designed food product of recent years is not something you can buy.

The other one made the same move

NotCo, out of Santiago, built Giuseppe — an AI system for finding plant-based combinations that reproduce the taste, texture and behaviour of animal products. Its consumer products, NotMilk and the Not Cheese slices made with Kraft Heinz through their joint venture, did reach American shelves.

Then in 2024 NotCo handed its North American consumer operations to Kraft Heinz, as Bloomberg reported in February 2025. Today NotCo's own site sells Giuseppe as software. It is described there as “the first end-to-end product development platform,” aimed at “marketing & innovation at CPGs, retailers, private labels,” at food scientists in manufacturing, and at “engineers in synthetic biology, materials, or biopharma.” The headline claim is that it can “reduce your trial and error by up to 10x.”

We could not find any independent evaluation of that 10x figure. It appears on NotCo's marketing site and nowhere we could verify outside it, so treat it as a vendor claim rather than a measurement.

The pattern across both companies is the same, and it is not sinister. The AI is real, and it is being sold — to food manufacturers, not to you. What reaches your kitchen is an ordinary grocery product, and nothing on the package tells you whether a model helped design it.

The appliance aisle is more honest than the marketing

The literal AI in your kitchen — the kind with a plug — is easier to pin down, because the specifications get published.

Samsung's AI Vision, the camera system inside its Bespoke AI refrigerators, recognized 37 food items, as of April 2025,, which Samsung describes as fresh fruits and vegetables. Packaged foods it cannot identify on sight: up to 50 of those can be saved with names you assign yourself. In a December 2025 announcement ahead of CES 2026, Samsung said an upgraded version would recognize more items and unlock the existing limitations of a system that until then required processed foods to be pre-registered, and that Google Gemini was being built into a refrigerator for the first time.

Thirty-seven is a real number, and it is small. That is worth considerably more to you than the phrase “AI-powered kitchen,” because it tells you exactly where the edge of the capability sits. The hardware, meanwhile, carries hardware prices: a Bespoke slide-in electric range with the Smart Oven Camera, model NSE6DG8700SRAA, currently sells for $2,499, down from a $4,109 reference price, on Samsung's own store.

The part you can actually check

In 2011 four network scientists — Yong-Yeol Ahn, Sebastian Ahnert, James Bagrow and Albert-László Barabási — published Flavor network and the principles of food pairing in Scientific Reports. It is open access. Anyone can read it, including you, right now.

They took 56,498 recipes from epicurious.com, allrecipes.com and the Korean site menupan.com, and linked 381 ingredients to the 1,021 flavor compounds known to be present in them — each ingredient linked to about 51 compounds on average, since a single compound turns up in many different ingredients. The average recipe used around eight ingredients. Egg turned up in 20,951 of them, more than a third.

Then they asked whether recipes pair ingredients that share flavor compounds more often than chance would predict. The answer split the world in two. North American and Western European recipes use compound-sharing pairs significantly more than chance. East Asian and Southern European recipes use them significantly less.

The paper's own illustrations of the two extremes are the fun part:

• Coffee and beef share 102 flavor compounds.
• Shrimp and lemon share 9.
• Chocolate and blue cheese share at least 73 — which is the reasoning behind restaurants that serve them together.
• White chocolate and caviar share trimethylamine, the logic behind one of the more notorious pairings in modern fine dining.

Why the finding is smaller than it sounds

Here is the part that almost never survives the retelling, and it is the most interesting part.

The effect is not spread evenly across a cuisine. It is carried by a handful of ingredients. When the authors removed ingredients one at a time, in order of how much each contributed, the statistical signal collapsed after just 13 removals for North American cooking and 5 for East Asian. The North American drivers were milk, butter, cocoa, vanilla, cream, cream cheese and egg. The East Asian ones were beef, ginger, pork, cayenne, chicken and onion.

The 13 top-contributing North American ingredients - those seven plus peanut butter, strawberry, cheddar cheese, orange, lemon and coffee - appear in 74.4% of all recipes. So the finding is less Westerners follow a hidden pairing rule and more Western baking leans hard on dairy and vanilla, and dairy and vanilla happen to share a great deal of chemistry.

The authors also flag their own limit plainly: they could not account for how much of each compound is present, or the threshold at which a human nose detects it. A compound present in trace amounts and one you can smell across the room count the same in their data.

That is what a checkable claim looks like, including the parts that weaken it. Compare it with “reduce your trial and error by up to 10x.”

What to do this week

Three things, in ascending order of effort.

1. Cook one deliberate high-overlap pairing. Coffee and beef is the paper's own headline example at 102 shared compounds, and it is not exotic — a spoonful of finely ground coffee in a beef rub or a pot of chili is the cheapest possible version of the experiment. If you would rather take the dessert route, dark chocolate with a crumble of blue cheese is the 73-compound pairing.

2. Then cook a deliberate low-overlap one, and pay attention to the difference. Shrimp and lemon share nine compounds, and it is one of the best-loved combinations on earth. That is the honest lesson buried in the study: compound sharing is a real, measurable pattern in Western recipes, and it is not a rule for what tastes good. East Asian cooking has been going the other way, deliberately and successfully, for a very long time.

3. When you meet an AI food claim, ask the two-part question. Who is making it, and where can it be checked? “Blind-tasting judges named it a winner” is a fundamentally different kind of statement from “our platform reduces trial and error by up to 10x,” even when both appear in the same article, in the same confident tone.

The machines really are getting good at this. But right now the AI most likely to change what ends up on your plate is not in your kitchen — it is in a food manufacturer's R&D department, and you will never see its name on the box.