A London startup trained an AI on 4.1 million recipes across seven languages
- KAIKAKU.AI published Epicure, a family of three ingredient AI models trained on 4.14 million multilingual recipes.
- The model doesn't store recipes—it stores what was learned from them, letting users navigate cooking knowledge mathematically.
- Three variants—Cooc, Chem, and Core—sit at different points on a recipe-context vs. flavor-chemistry spectrum, each answering a slightly different culinary question from the same 2MB file.
Josef Chen says he compressed all of human cooking into two megabytes. That's a bold claim. It also checks out.
Chen, co-founder and CEO of London food AI startup KAIKAKU.AI, published a paper on arXiv this week, alongside researcher Jakub Radzikowski, presenting Epicure—three AI models trained on 4.14 million recipes pulled from 11 datasets across seven languages. The result: a map of 1,790 ingredients, each described by 300 numbers, ...
[...]
Think of it as a map. Every ingredient gets a precise location based on how it behaves across millions of real dishes worldwide. The math is blunt: 1,790 ingredients × 300 numbers per ingredient × 4 bytes each ≈ 2.05 megabytes. Those numbers encode which ingredients appear together, which share flavor compounds, and which belong to the same culinary tradition. Once the model learns all that from the recipes, the recipes can go. The knowledge lives in the coordinates.
This is essentially the same trick word2vec pulled on language back in 2013, when Google researchers showed that you could do arithmetic with meaning. Epicure does that for food. Take beef, point it toward America and you'll get bread, lettuce, maybe beer. Point it toward South East Asia and the model stops thinking about burgers and grills and starts thinking about soy sauce, ginger, and sesame oil.
[...]
Epicure comes in three versions, and picking the right one depends on what you're actually asking. Cooc learns from recipe co-occurrence—what shows up together in real dishes. Chem learns from flavor chemistry—which ingredients share aroma compounds from the FlavorDB chemical database. Core is a mix between the previous two.
[...]
Why this isn't ChatGPT for food
Epicure has no general knowledge, no language generation, and no ability to hallucinate an ingredient it's never seen. It knows 1,790 ingredients. That's the whole world, as far as this model is concerned. What it gives up in breadth it gains in reliability—unlike recipe chatbots that will confidently suggest poison as a cooking ingredient if you push them the wrong way.
[...]
Practical uses aren't hard to picture. A chef asks what the East Asian equivalent of a Mediterranean ingredient looks like. A food product developer asks what minimally processed swap lands in the same flavor zone as an additive. A recipe app needs a coherent substitution when an ingredient is missing from the pantry.
The Epicure paper is a research release. The trained models are live on Hugging Face and an interactive ingredient map is publicly accessible at epicure.kaikaku.ai. They even released an MCP for your agents. Full training code is not released at this time.
I would clarify it to "All Modern Human Cooking", as the ingredients don't include woolly mammoth nor dodo. But it does have bison.
(Score: 5, Funny) by Anonymous Coward on Sunday May 31, @05:36PM (5 children)
To serve man
(Score: 2) by mhajicek on Sunday May 31, @06:19PM (4 children)
The ingredients are only part of a recipe.
The spacelike surfaces of time foliations can have a cusp at the surface of discontinuity. - P. Hajicek
(Score: 3, Informative) by pTamok on Sunday May 31, @06:25PM (2 children)
Ho yes. 'Sweating' onions or frying them. Big difference. Same for whale-meat in a stew as opposed to fried. And there are huge differences between potato types: some great for chips, some great for mash. Lokk at South American varieties for a mind-blowing experience, including ones that are 'impossible' to peel.
(Score: 4, Funny) by turgid on Sunday May 31, @06:58PM (1 child)
Potatoes? Golden Wonder is the Slackware of potatoes.
I refuse to engage in a battle of wits with an unarmed opponent [wikipedia.org].
(Score: 5, Interesting) by pTamok on Sunday May 31, @07:58PM
Or apples. A world of difference between Cox's Orange Pippins, Worcester Pearmains, Egremont Russets, Golden Delicious, Granny Smiths, Bramleys. And you want Seville Oranges for marmalade.
One of the many skills involved in cooking is knowing the properties of your ingredients, then making the best of them. A good cook can make excellent meals from mediocre ingredients.
(Score: 4, Funny) by istartedi on Sunday May 31, @07:10PM
No way. All you need is 10 minutes in the microwave. I tested it with the database and Pho. It worked great. Beef Wellington next.
Appended to the end of comments you post. Max: 120 chars.
(Score: 3, Funny) by SomeGuy on Sunday May 31, @06:43PM (2 children)
Yes, but can it preserve the fine effervescence of well fetid squirrel, finely fermented rat, and assorted bird droppings that go in to every Domino's pizza?
Does it contain health and allergy information needed to avoid Bixonimania?
What is its recipe for... cake?
That's right, with glorious all mighty AI, you too will be cooking the best dishes with hallucinatory mushrooms in no time! Guaranteed to send you running to the bathroom for a spewing diagonal farginal. 6 out of 73.4 experts agree Cooking With AI is for everybody. It's not a slicer or a dicer, it's not a chopper or a hopper, it's the AI-O-MATIC! To clean up, just throw it in to a vat of molten steel! Battery acid not included.
(Score: 3, Touché) by turgid on Sunday May 31, @06:56PM (1 child)
Yes, but can it preserve the fine effervescence of well fetid squirrel,
Ah! Lovely bit of squirrel [youtube.com]
I refuse to engage in a battle of wits with an unarmed opponent [wikipedia.org].
(Score: 2, Interesting) by pTamok on Sunday May 31, @08:01PM
You may laugh, but there was a pub in Rye (could even have been Spike Milligan's local) that served squirrel pie. I tried it, and was not impressed. But I am partial to flying rats (aka pigeons) as opposed to tree rats (squirrels).
(Score: 4, Interesting) by looorg on Sunday May 31, @08:34PM (5 children)
Two megabytes of just references and text is quite a lot of data. A lot of cooking is very similar, have similar ratios of ingredients or combinations of ingredients and preparation steps. So I don't really find it all that strange that it can be compressed. A large set of the data is probably different names for the same or very similar dishes.
I doubt barbecuing or frying some mammoth stakes is all that different from preparing meats today, I'm sure there is some African recipe for elephant or we could just see it as big beef. Making Dodo omelettes is probably not to different then making it with normal sized eggs, you just need to scale it up to the proper egg size.
So are their recipes for long pork in there? If it's just "human cooking" I doubt it will be a very long list. It's one dish.
(Score: 2) by looorg on Sunday May 31, @08:44PM (1 child)
Mammoth stakes ... sounds very woody and hard ... spelling ...
(Score: 3, Touché) by cereal_burpist on Monday June 01, @03:07AM
Mammoth stakes are used to kill giant vampires.
(Score: 3, Insightful) by Thexalon on Monday June 01, @02:27PM (2 children)
Also, it's an art, not a science, and there's lots of room for individual and cultural variations that mean we don't all expect identical food to be "good". Techbro thinking doesn't make food better or easier or healthier, it just leads us to such nonsense as some kind of regulatory standard for a green salad, when the right answer is "I'll put in it what I want to eat, dammit".
"Think of how stupid the average person is. Then realize half of 'em are stupider than that." - George Carlin
(Score: 3, Interesting) by looorg on Monday June 01, @04:41PM (1 child)
That is usually how I do cooking and baking. Go with the force. That seems good. As long as the ratios are within reason. Then as noted how do you like your food prepared. But that is execution. Recipes and ratios are somewhat standardized I would imagine. I don't think there is a big difference in how a lot of things are done, around the world. If you want to at least reach a somewhat similar result.
(Score: 3, Interesting) by Thexalon on Monday June 01, @05:20PM
Generally my rules are:
1. Understand the key principles of the type of thing I'm doing, e.g. leavening with yeast or chemicals, how egg whites work for air, that sort of thing.
2. Once you know that, the recipe is more of a l suggestion.
3. The more times you make it, the less you measure and the more you just do what you like.
Contrast that to how the modern food industry seems to want to do things, which is everything standardized down to the milligram and the exact number of fractional seconds something should be in an oven or on a griddle or in the deep frier. Which takes most of the fun out of it.
"Think of how stupid the average person is. Then realize half of 'em are stupider than that." - George Carlin
(Score: 3, Insightful) by jb on Monday June 01, @08:13AM
Then it cannot possibly cover "all human cooking".
Quite a few dishes do use poisonous ingredients. The skill of the cook lies in getting the proportions just right to maximise the flavour experience for the diners, without ever killing any of them and ideally without even making any of them just a little bit ill either. And because the vast majority of ingredients (whether poisonous or not) come from animals or plants (all of which are subject to inherent variation), that skill is more of an art than a science (although there's a bit of both involved).
(Score: 3, Insightful) by VLM on Monday June 01, @05:34PM
AI must be dying because now there's a desperate attempt to claim any time a computer calculates something, that's AI.
Essentially, what's the correlation coefficient of some cooking ingredients? This is "AI" because a computer did it in 2026, but most people would call it a rather boring statistical analysis.
It's not even tracking if it's "good" just is it common. Like the flies on poop analogy, clearly that must be delicious.
Maybe I can use AI to calculate my high school GPA since mere trivial statistical analysis is now "AI". Add 8 numbers and divide by 8 in my former school district is now doing "AI". Well I'm less likely to get it wrong than a chatbot LOL.