Compressing the Culinary World: A Data-Driven Approach to Global Ingredients

A recent research paper published by Kaikaku, a London-based startup, has sparked a significant debate among the technical and culinary communities. The project aims to "compress" the vast complexity of human cooking into a compact 2-megabyte dataset. While the title suggests a totalizing effort, the actual work focuses on the normalization of ingredients across multiple languages and the identification of flavor pairings that transcend cultural boundaries.

The Core Concept: Ingredient Normalization

At its heart, the research is an exercise in data normalization. The authors have processed recipes from 11 sources spanning seven languages—including English, Chinese, Russian, Vietnamese, Spanish, Turkish, Indonesian, German, and Indian-English. By using LLMs (specifically Claude) for classification, they have reduced thousands of disparate ingredient names into approximately 1,800 "primitives."

This process allows the system to recognize that "scallion," "green onion," and "long onion" are essentially the same primitive. By mapping various regional terms to a single identifier, the researchers can analyze the co-occurrence of ingredients across different cuisines, creating a mathematical representation of what makes certain flavors work together regardless of the where they are cooked.

Technical Implementation and Utility

For developers and food-tech entrepreneurs, the utility of this dataset lies in its ability to act as a flavor-pairing engine. Because the data is structured around primitives, it can suggest substitutions or identify complementary ingredients that might be surprising but mathematically sound.

One user, @dhx, shared their experience with a demo of the system, noting that the model is surprisingly sophisticated in its handling of specific culinary contexts. For instance, the system knows to suggest chilling cooked jasmine rice before using it in fried rice, or soaking Basmati rice for a pilaf. It also intelligently selects cuts of meat—such as shoulder or shank—when paired with vegetables typically used for braising.

The "Clickbait" Controversy

Despite the technical achievement, the project has faced heavy criticism regarding its framing. Many in the Hacker News community argued that the title "All of human cooking compressed into 2 megabytes" is misleading. The primary arguments against the title include:

• Lack of Technique: The dataset focuses on what is in the dish, not how it is made. Preparation methods, proportions, and the "art" of cooking are entirely absent from the 2MB file.
• Cultural Gaps: Critics pointed out the glaring omission of French and Italian cuisines—two of the most influential culinary traditions in the world—as well as a lack of representation for African ingredients.
• Data Scale: With only 11 sources, the dataset is far from a comprehensive map of "all human cooking."

As @epsteingpt noted, a more accurate title would have been "all of human ingredients compressed into 1,800 primitives."

Insights and Counterpoints

The discussion revealed a fascinating tension between the desire for a clean, programmable dataset and the reality of culinary diversity. While some users saw this as a "word2vec moment" for flavor profiles, others argued that the essence of cooking cannot be reduced to a list of primitives.

"Making a nice lentil soup doesn't require any thought or description. I know that I, and millions of cooks in Asia will do it with just their hands."

This highlights a fundamental limitation of any data-driven approach to cooking: the tacit knowledge—the smells, the textures, and the intuitive adjustments—that exists outside of written recipes.

Conclusion

While the 2MB dataset may not represent the totality of human culinary knowledge, it provides a powerful tool for ingredient normalization and flavor discovery. By stripping away the linguistic noise of different languages and regional dialects, it creates a foundation for automated cooking systems and AI-driven recipe generation. The project serves as a a reminder that while data can be compressed, the cultural and sensory experience of cooking remains stubbornly uncompressible.