Digitizing taste.
A photograph became pixels. A song became samples. A dish still has to be cooked and put in someone's mouth.
- image
- sound
- taste
Taste is the last sense with no digital form. We are building one.
Every recipe is a positive sample.
Nobody writes down a dish they didn't like. A recipe exists because someone tasted something, wanted it again, and took the trouble to say how. That makes the world's recipes a labelled record of what tasted good, kept for centuries without anyone calling it data.
Cooking is irreversible.
A method changes an ingredient, and there is no way back. Seared is not raw with extra steps. It is a different thing, and the next method starts from there. So a dish is not a list of ingredients. It is an ordered composition of methods, each of which closes doors behind it.
Together these two facts are the whole thesis: a very large set of examples that people liked, each one a path through a space where every step is one-way. That structure is what we model.
231,637home recipes in the first corpus we work from
1,071,520ratings left by the people who cooked them
Three problems nobody has solved.
-
No negative samples.
Recipes record what worked. The dishes that failed were eaten in silence or thrown out, so the bad half of the evidence was never written down.
-
No objective ranking.
Two cooks disagree about the same plate and both are right. There is no test you can run on a recipe to find out whether it is delicious.
-
Skill is in the loop.
The same recipe in different hands gives different food. The written instructions were never the whole program. The cook supplied the rest.
Code has a cheap, fast, objective test: run it. Taste has a human mouth, which is slow, expensive, and disagrees with itself. Most of the work is in building around that gap.
What a digital sense of taste is for.
- New dishes that are actually good. Creativity in food is limited by how many attempts a kitchen can afford. Move the trials off the stove and the limit goes away.
- Seasonal and sustainable by default. When you know why a dish works, you can rebuild it from what is nearby and in season instead of what was shipped in.
- Restaurants against inflation. When an ingredient triples in price, a kitchen that understands its own recipes can substitute or change method and keep the plate on the menu.
Once taste has a form you can compute with, a recipe stops being a fixed text and becomes something you can adapt on purpose.
Work on this with us.
We want to hear from people who cook, people who study food, people who train models, and people who hold the data or the machines this needs.
- chefs and cooks
- food scientists
- machine-learning researchers
- recipe corpus owners
- compute partners
Tell us who you are and what you would bring. Methods and data are shared under agreement.
Three Porridges is a small independent studio in California. This is its main project.