What if you could ask an AI to suggest what goes well with chicken, and it would give you completely different answers depending on how it was trained? That's exactly what’s happening with a new AI system called Epicure from a London startup, Kaikaku.AI. This system shows how the way an AI learns can dramatically change what it recommends — and why that matters for everything from cooking to health.
What is Taste Embedding?
Think of taste embedding like a map that AI systems use to understand flavors. Just as you might use a map to find your way around a city, AI systems use taste embedding to understand how different foods relate to each other. But there are two main ways to create this map:
- Recipe-based learning: The AI looks at thousands of recipes and sees which ingredients are often used together. For example, if chicken is often paired with garlic, tomatoes, and rice in many recipes, the AI learns that these ingredients go well together.
- Molecule-based learning: The AI looks at the actual chemical makeup of foods. For example, garlic and chicken both contain certain molecules that make them taste good together — even if they’re not in the same recipe.
These two methods can lead to very different results, as the new AI system demonstrates.
How Does Taste Embedding Work?
Imagine you're teaching a child to recognize animals. You could show them pictures of cats and dogs together (recipe-based learning) — and they learn that cats and dogs are both pets. Or, you could show them the actual physical features of cats and dogs (molecule-based learning) — and they learn that both have fur, four legs, and whiskers.
Similarly, AI systems can be trained in two ways:
- Recipe-based model: This model learns by looking at millions of recipes. It sees patterns like, "chicken + potatoes + onions" and builds a map of what ingredients go together in real-world cooking.
- Molecule-based model: This model looks at the chemical structure of foods. It knows that chicken contains certain amino acids and fats, and that garlic contains allicin. It then uses this chemical knowledge to predict flavor matches.
Both models can be trained on the same data, but they end up with different results — much like how a child might learn to group animals by appearance or by habitat.
Why Does This Matter?
This discovery matters because it shows that AI systems don't just learn from what we tell them — they also learn based on how we train them. In the case of Epicure:
- The recipe-based model suggests things like rice and vegetables because those ingredients are often used in the same recipes.
- The molecule-based model suggests ingredients like garlic or even raw meat because they share chemical features that make them taste good together — even if they're not in the same recipe.
Interestingly, the molecule-based model even outperforms the recipe-based models in predicting taste and nutrition, even though it was never explicitly trained on that data. This shows that the AI is not just copying recipes — it’s actually understanding how foods interact at a deeper level.
Key Takeaways
- AI can learn in different ways: It can learn from recipes or from the chemical makeup of ingredients.
- Different learning methods lead to different results: Recipe-based AI might suggest what's commonly used together, while molecule-based AI might suggest what tastes good based on chemical similarities.
- Chemical AI is powerful: Even without being trained on nutrition or taste data, the molecule-based AI still performs better at predicting flavor.
- Real-world impact: This kind of AI could help chefs, nutritionists, and food scientists make better decisions by understanding flavor and nutrition at a deeper level.
So next time you ask an AI what goes with chicken, remember — it might give you different answers depending on how it learned. And that’s a big deal for the future of food, science, and AI.
