AI Models vs Optimal Human Diet
Same first-principles prompt. Many models. Side-by-side metabolically optimal plates.
Reasoning experiment only. Not medical advice. Every plate is a hypothesis from an AI run or joint debate — not a prescription.
Model Lab asks frontier language models one hard question: if the only goal is metabolic optimum for a healthy young adult, what is the default plate? Answers must lead with a clear plate description, then substantiate the choice, then report animal vs plant calorie shares. Compare OpenRouter model outcomes, a joint debate fable, and an intelligence-vs-composition graph.
What is this?
This site is a small experiment: give several frontier language models the same hard question about food — what should a healthy young adult eat by default if the only goal is metabolic optimum? Each answer must lead with a clear Plate description, then Substantiate plate choice — that two-part outcome is the core of the lab.
It is not a product, a medical service, or a popularity contest. Models must reason from logic and first principles only. Scientific studies, guidelines, and authority quotes are out of bounds: they are associative, confounded, and endlessly re-interpretable in debate.
Metabolic optimal is the sole outcome this lab optimizes for: the intake pattern that best supports a healthy young adult’s core metabolic function — fuel handling, cellular repair, hormone and mineral balance, low chronic inflammatory / toxic noise — treated as the master hypothesis from which many other wellness goods tend to follow. It is a peak claim, not a survival diet that happens to work well.
Every answer is a hypothesis. A good hypothesis is not instantly wrecked by healthy people on another pattern; nor may it demand that those people must be deficient when observation allows them not to be.
The first card is a joint debate fable (Grok 4.5 + Claude Sonnet Fable 5) — one combined write-up, not the answer key. Other cards are single-model outcomes. Intelligence scores link to Artificial Analysis; they measure model capability, not nutrition truth.
How the comparison works
- Fixed core prompt — every model gets the same system + user task: design a metabolically optimal default plate from first principles only.
-
Mandatory shape — Plate description bullets first, then Substantiate
(≤280 words), then a
CALORIE_SHARESfooter (animal % / plant % kcal). - Outcomes carousel — read model cards side by side, including a joint debate fable that is not labeled as the correct answer.
- Intelligence graph — plot Artificial Analysis intelligence against estimated animal vs plant calorie share from each capture.
- Discuss — bring your own OpenRouter key and any model into the same core prompt in-browser.
Core prompt
Loading the fixed system + user prompt used for every model capture…
Model outcomes
Loading AI model plate outcomes…
Intelligence vs plate composition
Artificial Analysis Intelligence against estimated calorie share from animal vs plant foods across model answers.
No calorie shares yet. Prefill model answers (or open a discussion that returns the
required CALORIE_SHARES block) to populate this chart.
FAQ
What is Metabolic Optimal Model Lab?
A public experiment that gives many AI models the same first-principles prompt: design the metabolically optimal default plate for a healthy young adult. Answers are compared side by side with estimated animal vs plant calorie shares and model intelligence scores.
Is this medical advice?
No. Model Lab is a reasoning experiment only. Every plate is a hypothesis from an AI run or joint debate — not a prescription, diagnosis, or treatment plan.
What does “metabolic optimal” mean here?
The sole objective is peak metabolic function for a healthy young adult: fuel handling, repair, hormone and mineral balance, and low needless metabolic load. It is a peak claim, not mere survival or absence of frank deficiency.
Why ban studies and guidelines in the prompt?
The lab forces models to reason from food chemistry, physiology, and internal consistency only. Studies and guidelines are associative, confounded, and endlessly re-interpretable; they are out of bounds so answers stay comparable under one method.
How can I run my own model?
In the Discuss section, create a free OpenRouter API key, paste it (stored only in your browser), pick a model, and start discussion. The lab sends the same core prompt first.
Discuss with your own models
Bring any OpenRouter model into the same core prompt, then keep talking.
- Create a free account and API key at openrouter.ai/keys.
- Paste the key below — it stays in this browser only (localStorage).
- Pick a model from the live OpenRouter list, then Start discussion.
- That sends the lab’s core prompt first. After the model answers, use the box to challenge, refine, or dig into calorie shares.