Kaikaku.AI's new research introduces three nearly identical AI models trained on different data sources to explore how AI understands ingredient relationships. The models—Cooc, Chem, and Core—differ only in their training data. Cooc learns from real recipes, Chem from flavor molecules in the FlavorDB database, and Core from both. The models cluster ingredients based on similarity without being told which cuisine they belong to, sorting them into distinct regional groups. | Image: Radzikowski & Chen When queried with 'chicken,' Cooc returns garlic, onion, and black pepper, common recipe companions, while Chem suggests beef or pork, ingredients with similar flavor profiles. For 'basil,' Cooc provides parsley, olive oil, and parmesan, whereas Chem offers oregano, tarragon, and rosemary. The test evaluates how accurately models can read properties like fruity or bitter content, with Chem leading across the board. | Image: Radzikowski & Chen The chemistry-based Chem model also excels in areas where it shouldn't have direct information, such as classifying ingredients along axes like sweet, sour, or bitter. The model's performance suggests that chemical relationships act as a shortcut, tuning it to broader culinary concepts. Epicure processes 4.14 million recipes from eleven sources in seven languages, including Chinese, Russian, Vietnamese, Turkish, Indonesian, and German. A pipeline using Claude and Gemini embeddings translates and cleans 200,000 raw terms into 1,790 clean ingredients. | Image: Radzikowski & Chen The corpus remains unevenly distributed, with about half the material from East Asian sources. Only about a third of the ingredients are directly linked to the chemical database, with the rest picking up chemical signals indirectly. The model offers two modes: a simple neighbor search and a directional shift toward a target. Labels like 'dessert ingredients' or 'Chinese wok cooking essentials' are generated by Claude. | Image: Radzikowski & Chen The model's choice can influence the cultural context of the answer. Turning 'chocolate' toward 'sweet pastries' leads Cooc and Core to Western baking ingredients, while Chem associates it with an East Asian dessert cluster. The authors are building robot restaurants, with Kaikaku.AI's robotic restaurant, Common Room, in London. The company's machine, 'Fusion,' can dispense 360 bowls per hour. Model weights and datasets are now available on Hugging Face for independent verification. *Source: [thedecoder](https://the-decoder.com/ask-ai-what-goes-with-chicken-and-the-answer-depends-on-whether-it-learned-from-recipes-or-molecules/)*