Food AI is about to have its ChatGPT moment.
Our first paper is now on arXiv: Epicure.
For decades, food has been treated as too human, too sensory, too cultural, and too computationally intensive to model properly.
We broke that assumption. 🧵
Our first paper is now on arXiv: Epicure.
For decades, food has been treated as too human, too sensory, too cultural, and too computationally intensive to model properly.
We broke that assumption. 🧵
1/ We prove that LLMs can encode the deep structure of flavour. Taste, texture, geography, processing, culture. All emergent and none of it explicitly taught.
The next foundation-model domain is on your plate.
The next foundation-model domain is on your plate.
2/ Using FlavorGraph's 300-D ingredient embeddings, we built an LLM-assisted curation pipeline that consolidates 6,653 raw ingredients into 1,032 canonical entities, and reveals at least 15 independently classifiable dimensions of food meaning.
All linear.
All linear.
3/ Every cuisine has a fingerprint.
E.g Japanese owns umami (+0.45). South Asian uniquely owns bitter (+0.31) AND hardness (+0.40). Mediterranean and N. Atlantic claim fattiness (+0.23 / +0.30). SE Asian owns heat (+0.20)
The model derived this from co-cooking patterns alone.
E.g Japanese owns umami (+0.45). South Asian uniquely owns bitter (+0.31) AND hardness (+0.40). Mediterranean and N. Atlantic claim fattiness (+0.23 / +0.30). SE Asian owns heat (+0.20)
The model derived this from co-cooking patterns alone.
4/ The skeptic's objection: maybe the LLM that labelled tastes also leaks into the embedding via shared web text.
Counter-test: take the USDA nutrient database, define axes from top vs bottom terciles of measured fat / protein / carbs / fibre / calories, project the embeddings.
ρ = 0.44–0.47 across all five macros. p ranging from 1.6×10⁻³³ to 6.8×10⁻⁴¹.
Counter-test: take the USDA nutrient database, define axes from top vs bottom terciles of measured fat / protein / carbs / fibre / calories, project the embeddings.
ρ = 0.44–0.47 across all five macros. p ranging from 1.6×10⁻³³ to 6.8×10⁻⁴¹.
5/ Example: pungency.
The model never saw a chili. It was never told what Scoville is.
Yet a single linear direction in latent space orders peppers from bell → habanero with ρ = 0.76 against measured Scoville units.
The model never saw a chili. It was never told what Scoville is.
Yet a single linear direction in latent space orders peppers from bell → habanero with ρ = 0.76 against measured Scoville units.
6/ And it's not just heat.
All five basic tastes: sweet, salty, umami, sour and bitter emerge as clean linear probes. So do texture, regional cuisine, processing method, and cultural role.
Machines are beginning to understand food. Not recipes but actual latent culinary structure.
All five basic tastes: sweet, salty, umami, sour and bitter emerge as clean linear probes. So do texture, regional cuisine, processing method, and cultural role.
Machines are beginning to understand food. Not recipes but actual latent culinary structure.
7/ The global geometry is the part I find most beautiful.
Flavour space is a torus. Two dense poles: sweet and savoury, connected by a ring of dual-citizens: chili, citrus, miso, vanilla. The ingredients chefs reach for to bridge worlds sit, literally, on the bridge.
Flavour is not a line. It's a loop.
Flavour space is a torus. Two dense poles: sweet and savoury, connected by a ring of dual-citizens: chili, citrus, miso, vanilla. The ingredients chefs reach for to bridge worlds sit, literally, on the bridge.
Flavour is not a line. It's a loop.
8/ The kicker:
"Glutamate" appears 0 times in the corpus. "Sucrose" 0 times. The model never sees molecules.
Yet umami and sweetness are cleanly recoverable.
Co-occurrence between ingredients in recipes is a sufficient statistic for chef cognition. Chefs are the labellers. Recipes are the labels.
"Glutamate" appears 0 times in the corpus. "Sucrose" 0 times. The model never sees molecules.
Yet umami and sweetness are cleanly recoverable.
Co-occurrence between ingredients in recipes is a sufficient statistic for chef cognition. Chefs are the labellers. Recipes are the labels.
9/ Food has always been treated as a messy, human, tacit domain.
We are here to change that @KAIKAKU_AI
Robotics gives us execution. AI gives us understanding.
Together, they create a new category: autonomous food infrastructure.
We are here to change that @KAIKAKU_AI
Robotics gives us execution. AI gives us understanding.
Together, they create a new category: autonomous food infrastructure.
@KAIKAKU_AI 10/ Huge credit to Dr Jakub Radzikowski @jaklerad, who brought decades of culinary intuition.
📄 Paper: arxiv.org/abs/2604.22776
📷 3D explorer: epicure-data.kaikaku.ai
If you work on foundation models and care about new domains, DMs open.
📄 Paper: arxiv.org/abs/2604.22776
📷 3D explorer: epicure-data.kaikaku.ai
If you work on foundation models and care about new domains, DMs open.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
