This AI Compressed 'All Human Cooking' Into 2 Megabytes
Josef Chen and Jakub Radzikowski trained Epicure, an AI food model that encodes all human cooking knowledge into a two-megabyte file. Epicure was trained on 4.14 million recipes in seven languages, mapping 1,790 ingredients into a 300-dimensional vector space. The model doesn't store recipes but assigns each ingredient a unique location based on how it is used across global cuisines. These coordinates capture patterns of ingredient co-occurrence, shared flavor chemistry, and culinary tradition. Epicure offers three variants: one learns from ingredient co-occurrence, another from shared chemical compounds, and a mixed version. Unlike generative AI chatbots, Epicure is limited to its ingredient list, preventing unreliable suggestions but providing precise, context-aware answers—such as compatible ingredient swaps or cross-cultural analogues. Its efficiency and reliability improve on previous models like FlavorGraph by using a larger, multilingual dataset. Epicure is intended for research and developer use, with the models and an interface publicly released online, though full training code is not available.
