Epicure - All of Human Cooking Compressed Into a 2MB AI Embedding Model

Researchers Jakub Radzikowski and Josef Chen have created Epicure, a groundbreaking food embedding model that compresses the culinary knowledge of 4.1 million recipes into just 2 megabytes. The model draws from eleven different sources across seven languages, including English, Chinese, Russian, Vietnamese, Spanish, Turkish, Indonesian, German, and Indian English. The researchers used an LLM-augmented pipeline to normalise raw ingredient data, reducing thousands of ingredient variations down to 1,790 canonical entries. They then trained three complementary embeddings: one based on ingredient co-occurrence patterns in recipes, another derived from shared chemical compounds between ingredients, and a third that combines both approaches into a unified model. The resulting 300-dimensional vector space positions each ingredient near its closest culinary and chemical neighbours, creating a compact representation of global food knowledge. This approach enables practical applications like intelligent ingredient substitution, cross-cultural recipe recommendations, and systematic analysis of different cuisines' flavour profiles. The work demonstrates how embedding techniques originally designed for language processing can effectively model culinary relationships. Food applications have traditionally relied on hand-crafted rules for features like ingredient substitution, but Epicure's data-driven approach offers a more sophisticated foundation for understanding how ingredients relate across different cooking traditions and chemical properties.