KAIKAKU.AI has launched Epicure, a set of AI models compressing 4.14 million recipes into just 2MB, revolutionizing culinary exploration.
KAIKAKU.AI Launches Epicure: A Revolutionary Culinary AI Model
KAIKAKU.AI has unveiled Epicure, a groundbreaking set of three AI models that distills extensive global culinary knowledge into a compact 2MB format. This innovative technology is based on a dataset comprising 4.14 million recipes from 11 languages.
Josef Chen, co-founder and CEO of KAIKAKU.AI, shared insights about the model through a paper published on arXiv, demonstrating how it captures the essence of 1,790 ingredients, each represented by 300 numerical descriptors. Chen emphasized the model’s efficiency, stating, "All of human cooking compressed into 2 megabytes."
Understanding the Epicure Model
Unlike traditional recipe storage systems, Epicure does not hold specific recipes. Instead, it functions as a sophisticated coordinate system that maps ingredients based on their culinary applications. The model can mathematically navigate the relationships between ingredients, making it akin to a culinary map rather than a cookbook.
Epicure employs a mathematical technique known as SLERP rotation to allow users to explore different culinary directions. By adjusting the focus on an ingredient, such as chicken, users can discover various flavor profiles linked to different cuisines. For example, shifting towards Tex-Mex cuisine reveals associations with corn tortillas and salsa.
The Epicure suite includes three models: Cooc, which analyzes ingredient co-occurrence; Chem, which focuses on flavor chemistry; and Core, which merges aspects of both. Each model provides tailored insights depending on the user's inquiries, enhancing the culinary exploration experience.
Epicure's Limitations and Future Applications
While Epicure represents a significant advancement in food AI, it is not without limitations. The model is restricted to its dataset of 1,790 ingredients and cannot generate new language or concepts outside its training. This specificity enhances its reliability compared to broader recipe bots that might suggest inappropriate ingredients.
Epicure surpasses previous models like FlavorGraph by incorporating a larger, multilingual dataset, offering potential applications for chefs, food developers, and recipe creators seeking substitutions or flavor compatibility.
The trained models are now accessible on Hugging Face, with an interactive ingredient map available at epicure.kaikaku.ai, marking a notable advancement in culinary AI research.
