Log a meal by photographing it

Conventional logging asks three questions per food and gets a shaky answer to each. Which chicken — the database has four hundred, and they differ by sixty calories. How much — you did not weigh it, so you guess, and the guess carries more error than the database choice did. What else is in it — a stir fry with eight ingredients is eight searches, or it is one entry called “stir fry” that is wrong.

Estimates for that loop run between one and three minutes per meal once you count the false starts. People do not abandon it because they stopped caring about their health. They abandon it because it is half an hour a week of data entry.

A photo or a sentence returns a full breakdown in seconds. Not one lumped total — an entry per ingredient, each with its own numbers, so a correction to the rice does not require re-entering the chicken.

Underneath it, the assumptions. Assumed oatmeal cooked in water, no added sugar. Assumed the same portions as your earlier bowl. No guacamole or sour cream included. These are not a disclaimer. They are the interface for correcting the estimate: an app that says “661 kcal” and nothing else gives you no way to know whether it counted the sour cream.

This is the part that decides whether any of the rest is worth having. AIM runs a frontier multimodal reasoning model — the same class of system used for hard general-purpose work, not a small image classifier trained on a fixed list of dishes and not a lookup against a food database.

It matters because a plate is not a classification problem. Working out that the bowl holds roughly 180 g of rice, that the chicken was grilled rather than fried, and that the sauce is adding more sugar than it looks like, is a judgment call about a specific photograph. A classifier can tell you “burrito bowl”. Getting from there to numbers you would act on takes reasoning.

So the model is given room to think before it answers, rather than being run at the cheapest, fastest setting. That costs a few seconds per meal and buys a breakdown that separates roughly seven components of a plate instead of collapsing it into three coarse rows — which is what makes per-ingredient correction possible at all.

The camera is the fast path, not the only one. “Chicken burrito bowl”, “two eggs and toast”, “large flat white” — a sentence gets the same treatment as a photograph, which matters for the meals you did not think to photograph until afterwards.

You can also log something you ate yesterday. Say when, and it goes on that day.