동일한 첫 번째 원칙 음식 프롬프트, 많은 LLM: 모델 랩 차트 동물 대 식물 칼로리 공유

I built a small static experiment: give several frontier models the same hard question about food, force a fixed output shape, and put the results side by side.

Live lab: https://nutrition.livo.community

What it is

Model Lab is a personal reasoning experiment, not a product and not medical advice.

Every model gets one core prompt. Sole objective: metabolic optimal for a healthy young adult (peak hypothesis, not “people do fine” or guideline aesthetics). Method is locked to first principles + logic only — studies and guidelines are out of bounds as authority so the comparison stays fair.

Mandatory answer shape:

1. ## Plate description (bullets first)
2. ## Substantiate plate choice (short)
3. CALORIE_SHARES footer (animal / plant % of calories + constituents)

How to read the site

• Carousel: each card is one model outcome (or the joint debate fable on the first card). The fable is not labeled as the correct answer — it is a Grok × Claude write-up under the same prompt contract.
• Chart: Artificial Analysis Intelligence score vs animal/plant % of calories. Animal share is drawn in red so you can scan the spectrum quickly.
• Discuss: optional browser-side OpenRouter key if you want to run more models yourself (key stays in localStorage only).

Why this is interesting for builders

Same system + user prompt, same format constraints, many vendors. You get a clean view of how models diverge when you strip citation theater and force a concrete plate + calorie split.

Static captures live in responses.json; the site is plain HTML/CSS/JS (no framework).

Again: hypothesis comparison only. Not a diet prescription.

If you try it, curious which models land nearest each other on the animal/plant axis — and whether that correlates with intelligence scores at all.