How Ai is quietly refining recipes

A written recipe is a broadcast: one author, fixed ingredients, an implied kitchen you probably don't have. Every cook already improvises around it, swapping, scaling, skipping steps. Language models happen to be excellent at exactly that kind of constrained rewriting, which is why the recipe is turning from a document into a conversation.

Adaptation beats invention

Watch what adaptation actually looks like on a Tuesday. The tikka masala recipe wants double cream; you have yogurt, and a model that knows heat will tell you to take the pan off the boil before it splits. The gratin wants an oven you don't have; it becomes a stovetop version with a lid, and the model can explain what happens to the crust. The stir fry wants five vegetables; you have two and a bag of frozen peas, and dinner still happens. None of these are inventions. They're translations, and translation is precisely the job models do best.

• Substitution with judgement. Not just "yogurt can replace cream" but whether it should, in this dish, at this heat.
• Constraint solving. Twenty minutes, one pan, no dairy, use up the spinach first. Four constraints at once is a small search problem, and models are decent search guides.
• Scaling that understands physics. Doubling a stew is linear. Doubling a tray bake is not.
• Goal tuning. The same dish, nudged toward your goals: less oil, more protein, an allergy respected without the dish losing its soul.

The fridge is the real input

Recipe sites start from the dish and send you shopping. A kitchen starts from what's already there. Inverting that direction, ingredients first, dish second, is the single most practical thing Ai has brought to home cooking. It also does the most against waste: households produce 60% of the world's food waste, and most of it dies in the fridge as good intentions. The courgette that would have gone off on Thursday becomes the reason Thursday's dinner exists.

This is Snapeto's cook-from-fridge mode, one of the app's two home buttons. The pantry and fridge are the query, the model proposes what tonight can be inside your guardrails, and every suggestion is cookable right now because it started from what you actually own.

The healthy twin: adaptation pointed at your goals

Snapeto runs the same adaptation engine in reverse, too. Every meal you snap comes back with a healthier twin: the same dish, reworked to fit your goals, allergies and pantry. Ate a takeaway korma? The twin is the version you could cook at home with half the oil and the same comfort. This is what information that changes behaviour looks like: not a score about last night, but a recipe for next time.

Where the models still fall over

Honesty section. Models occasionally invent confident nonsense, and in cooking that means a ruined pan or a flavour pairing no human would repeat. They've read about taste but never tasted. The fix is architectural: keep generations anchored to known dish patterns, constrain substitutions to tested families, and let user feedback prune what should never be suggested again. The other failure is blandness by averaging: ask an unanchored model for "a healthy dinner" and you get the statistical centre of the internet's food writing, which is always a chicken bowl. Personality has to come from your cuisines, your history, your repeats. The model refines. It must not homogenise.