Two ways to ask what goes with an ingredient
When someone asks what pairs with chicken, there are at least two valid answers. One is culinary: which ingredients commonly appear alongside chicken in real recipes. The other is chemical: which ingredients share a similar flavor profile at the molecular level. New research highlighted by Kaikaku.AI argues that many AI systems blur those answers together, and that doing so hides an important distinction.
The company’s new work introduces three closely related models under the name Epicure. One model, Cooc, is trained only on recipe co-occurrence. Another, Chem, is trained only on shared flavor molecules using the FlavorDB chemistry database. A third, Core, blends both approaches.
Why the distinction matters
The differences become obvious in simple prompts. According to the source text, Cooc responds to “chicken” with ingredients like garlic, onion and black pepper, reflecting what cooks commonly combine in recipes. Chem instead returns ingredients such as beef or pork, which are not necessarily the most common recipe companions but are closer in molecular flavor profile.
The same pattern appears with herbs. For “basil,” Cooc suggests ingredients associated with familiar use cases, including parsley, olive oil and parmesan. Chem groups basil with flavor relatives such as oregano, tarragon and rosemary. In other words, one model behaves more like a cookbook, the other more like a chemistry map.
Data scale and multilingual scope
Epicure was trained on 4.14 million recipes from eleven sources in seven languages, including Chinese, Russian, Vietnamese, Turkish, Indonesian and German. That multilingual breadth is a major part of the project’s claim to relevance. Many food datasets skew heavily toward English-language sources, which can flatten regional cuisines and overrepresent Western cooking patterns.
The source text says the pipeline used embeddings from Claude and Gemini to help translate and normalize about 200,000 raw ingredient terms into 1,790 cleaned ingredient labels. That kind of data preparation is less glamorous than model design, but it is often the difference between a system that captures real structure and one that amplifies noise.
Unexpected performance from chemistry-first learning
One of the more interesting claims in the research is that the chemistry-driven model performs well even on properties that were not directly encoded in its training data. The source text says Chem more clearly classifies ingredients along dimensions such as sweet, sour or bitter, and also on nutritional axes like protein and fat content.
If that result holds up, it suggests that molecular relationships may act as a compact representation for broader culinary knowledge. A model built from chemistry alone may still learn something meaningful about how humans perceive ingredients, organize flavor and even infer adjacent properties.
What this could change
Food AI has tended to focus on recommendation, substitution and content generation. But these systems often collapse very different questions into one generic notion of similarity. Epicure’s framing suggests that future tools may need to be explicit about what kind of similarity they are optimizing for.
That distinction could matter in product design. A recipe assistant should probably privilege co-occurrence and cuisine context. A formulation or R&D tool might care more about molecular similarity. A creative system for new dishes may need a tunable balance between both.
Just as important, the work shows that even narrow-seeming domains can expose larger model design issues. Training data does not merely fill in facts. It determines what kind of relationship the system believes the world contains.
A more precise food intelligence stack
The broader value of the project is conceptual clarity. “What goes with this?” is not one problem. It is several. By separating recipe behavior from flavor chemistry, Kaikaku.AI is making a case that ingredient intelligence should be decomposed rather than averaged together.
That may sound niche, but it maps onto a wider pattern in AI research. Models become more useful when they distinguish between different structures in the same dataset instead of compressing them into a single score. In this case, the result is a cleaner way to think about culinary knowledge itself: habits, molecules and the space where they overlap.
This article is based on reporting by The Decoder. Read the original article.
Originally published on the-decoder.com
