Kaikaku.AI and Epicure: Redefining Culinary AI

Kaikaku.AI and Culinary Innovation through AI

The startup Kaikaku.AI, specializing in restaurant technologies, recently unveiled a significant advancement in the use of artificial intelligence for culinary analysis. Their project, named "Epicure," consists of three distinct AI models that explore the relationships between ingredients from two perspectives: recipes and flavor molecules.

The Epicure Models: Cooc, Chem, and Core

Researchers Jakub Radzikowski and Josef Chen developed three nearly identical AI models, differentiated by their training data. The "Cooc" model focuses on ingredients that appear together in recipes. "Chem" relies on shared flavor molecules, using the FlavorDB database. Finally, "Core" combines both approaches.

Each model represents ingredients as points, clustering those that are similar. Although the models receive no information about the culinary origin of the ingredients, they manage to classify them into distinct regional cuisine groups.

Varied Responses to the Same Question

When querying these models about an ingredient like chicken, the responses vary. "Cooc" suggests ingredients commonly associated in recipes, such as garlic and onion, while "Chem" recommends meats with similar flavor profiles, like beef. This distinction illustrates how each model perceives culinary relationships.

For basil, "Cooc" recommends parsley, olive oil, and parmesan, typical of pasta dishes, while "Chem" suggests oregano, tarragon, and rosemary, related herbs.

The Superiority of the Chemical Model

The "Chem" model stands out for its ability to classify ingredients based on taste and nutritional properties, even though this data is not explicitly coded in its training. This demonstrates the effectiveness of chemical relationships in adapting the model to various culinary concepts.

According to the authors, "Chem" excels in areas where it should not have information, such as sweet, sour, or bitter flavors, and nutritional values like protein or fat content. Chemical relationships appear to act as a shortcut that also adjusts the model to other culinary concepts.

A Multilingual Corpus for Global Analysis

Epicure is based on a vast corpus of 4.14 million recipes from eleven sources in seven languages, including Chinese and Russian. This corpus is processed by Claude and Gemini embeddings, which translate and clean the data to yield 1,790 distinct ingredients.

Features and Practical Applications

The final model offers two modes of operation: a neighbor search to identify similar ingredients, and an adjustable ingredient movement towards a target direction. These features allow for the discovery of coherent ingredient groups, labeled by systems like Claude.

Without predefined categories, the analysis finds groups of ingredients that go well together. The groups are then assigned labels generated by Claude, such as "dessert ingredients" or "essentials of Chinese wok cooking."

Future Perspectives

Founded in London in 2023, Kaikaku is leveraging these innovations in its robotic restaurant Common Room, located in the Brunswick Centre. With automated systems capable of preparing 360 bowls per hour, the startup envisions chain expansion. The Epicure model, available on Hugging Face, could transform how global cuisines are understood and utilized by AI.

The company employs its own machine learning systems to weigh and portion ingredients. Its machine, called "Fusion," can theoretically distribute 360 bowls per hour. The system also includes ML-powered inventory management and safe 3D-printed food components. The company raised approximately $1.8 million in a pre-seed funding round in 2024.

Conclusion

The practical viability of these applications remains to be seen. The model weights and datasets are now available on Hugging Face, allowing for independent verification in theory. However, the examples presented in the article are carefully selected. In underrepresented regions like South Asia or Latin America, the responses are likely much less stable than for the dominant cuisines of East and West Asia.

The cleaning of vocabulary also depends on the output of linguistic models, which carry their own cultural biases. The fact that chocolate appears near matcha in the "sweet pastries" direction of one variant of the model is an interesting effect. But it says little about the reliability of such rotations beyond the carefully chosen examples.

Co-author Josef Chen promotes the model on X as "the largest multilingual food model ever built," claiming they have "compressed all of human cuisine into 2 megabytes." An older version of the model is available for demonstration at epicure.kaikaku.ai.