In short
- KAIKAKU.AI revealed Epicure, a household of three ingredient AI fashions skilled on 4.14 million multilingual recipes.
- The mannequin would not retailer recipes—it shops what was discovered from them, letting customers navigate cooking information mathematically.
- Three variants—Cooc, Chem, and Core—sit at completely different factors on a recipe-context vs. flavor-chemistry spectrum, every answering a barely completely different culinary query from the identical 2MB file.
Josef Chen says he compressed all of human cooking into two megabytes. That is a daring declare. It additionally checks out.
Chen, co-founder and CEO of London meals AI startup KAIKAKU.AI, revealed a paper on arXiv this week, alongside researcher Jakub Radzikowski, presenting Epicure—three AI fashions skilled on 4.14 million recipes pulled from 11 datasets throughout seven languages. The consequence: a map of 1,790 components, every described by 300 numbers, that matches in your electronic mail attachment restrict with room to spare.
“4.1M recipes. 7 languages. 1,790 components. 300 dimensions,” Chen wrote on X. “All of human cooking compressed into 2 megabytes.”
Launching our new paper on arXiv: we skilled the most important multilingual meals mannequin ever constructed.
4.1M recipes. 7 languages. 1,790 components. 300 dimensions.
All of human cooking compressed into 2 megabytes. pic.twitter.com/b4GiZ62UMt
— Josef Chen (@josefchen) Might 26, 2026
It isn’t storing recipes
Earlier than you think about a two-megabyte USB stick jammed with stir-fry directions, the mannequin would not retailer a single recipe. The 2 megabytes is extra a coordinate desk than it’s a cookbook.
Consider it as a map. Each ingredient will get a exact location primarily based on the way it behaves throughout thousands and thousands of actual dishes worldwide. The maths is blunt: 1,790 components × 300 numbers per ingredient × 4 bytes every ≈ 2.05 megabytes. These numbers encode which components seem collectively, which share taste compounds, and which belong to the identical culinary custom. As soon as the mannequin learns all that from the recipes, the recipes can go. The information lives within the coordinates.
That is primarily the identical trick word2vec pulled on language again in 2013, when Google researchers confirmed that you would do arithmetic with that means. Epicure does that for meals. Take beef, level it towards America and also you’ll get bread, lettuce, perhaps beer. Level it towards South East Asia and the mannequin stops interested by burgers and grills and begins interested by soy sauce, ginger, and sesame oil.
This occurs by means of what the paper describes as a steering operator referred to as SLERP rotation. Take a seed ingredient—hen—and rotate it mathematically towards a delicacies path. At 30 levels you begin seeing Tex-Mex territory. At 60 levels, hen and beef converge on the identical Mexican pantry: corn tortilla, salsa, monterey jack, poblano pepper. The angle is a dial between “keep close to this ingredient” and “land someplace new.”
Epicure is available in three variations, and choosing the right one is determined by what you are truly asking. Cooc learns from recipe co-occurrence—what exhibits up collectively in actual dishes. Chem learns from taste chemistry—which components share aroma compounds from the FlavorDB chemical database. Core is a combination between the earlier two.
Ask Cooc what pairs with chocolate and you could get dessert-pantry companions: cocoa powder, vanilla, almond. Ask Chem and also you get flavor-chemistry friends: toffee, fudge, ganache.
Similar ingredient, completely different query. A chef searching for a substitute has completely different wants than a chef mapping taste compatibility.
Why this is not ChatGPT for meals
Epicure has no normal information, no language era, and no potential to hallucinate an ingredient it is by no means seen. It is aware of 1,790 components. That is the entire world, so far as this mannequin is worried. What it offers up in breadth it features in reliability—not like recipe chatbots that may confidently recommend poison as a cooking ingredient for those who push them the unsuitable method.
The earlier cutting-edge right here was FlavorGraph, a 2021 mannequin that mixed chemical knowledge with the English-only Recipe1M+ dataset. Epicure brings in a multilingual corpus greater than 4 instances bigger and cleans the vocabulary for effectivity.
Sensible makes use of aren’t arduous to image. A chef asks what the East Asian equal of a Mediterranean ingredient appears to be like like. A meals product developer asks what minimally processed swap lands in the identical taste zone as an additive. A recipe app wants a coherent substitution when an ingredient is lacking from the pantry. That final one is the hole the place purpose-built small fashions quietly outperform the large generalist ones.
The Epicure paper is a analysis launch. The skilled fashions are reside on Hugging Face and an interactive ingredient map is publicly accessible at epicure.kaikaku.ai. They even launched an MCP to your brokers. Full coaching code will not be launched presently.
Day by day Debrief E-newsletter
Begin every single day with the highest information tales proper now, plus unique options, a podcast, movies and extra.