Atlas › What it knows

Case study 95 of 198

Jev knows calories and fat, not vitamin C or incubation

Comparing two foods by a nutrient, or two animals by lifespan, gestation or clutch size, which quantities does Jev know and which does it guess?

result12,000 questions

Even when one value is at least twice the other, how often Jev picks the larger one depends a lot on what's being compared: calories 99%, fat 98% and protein 97% for foods, but vitamin C and calcium only 88%, and for animals, egg incubation time 84%.

60%80%100%
incubationanimal
vitamin Cfood
calciumfood
cholesterolfood
potassiumfood
sugarfood
vitamin Afood
maturityanimal
sodiumfood
longevityanimal
gestationanimal
ironfood
fiberfood
caffeinefood
litteranimal
proteinfood
fatfood
caloriesfood

Jev

How to read this: Each row is one quantity (a nutrient in food, or a life-history trait in animals), with how often Jev picks the larger of two items that differ at least twofold, and a 90% interval.

10,717 pairs at least 2x apart over 18 quantities; food-group split over 3,240 cross-group pairs (1543 against the usual group). Not a food-group stereotype: between foods from different groups it is right 93% when the answer comes from the group usually richer in that nutrient and 94% when it doesn't (90% intervals [0.917, 0.937] and [0.928, 0.948]).

In short

  • Even on easy comparisons, Jev's accuracy depends on the quantity: 99% for calories, but 88% for vitamin C and calcium.
  • For animals it knows litter size (97%) much better than egg incubation time (84%).
  • The misses aren't food-group stereotypes: it is right 93% when the richer food is from the usual group and 94% when it isn't.

What the data shows

what it knows
Jev looking for the food with more vitamin C (88% vs 99% for calories)
Confused John Travolta Pulp Fiction meme: Jev looking for the food with more vitamin C (88% vs 99% for calories)
How funny is this meme? Jev: 2/5, slightly funny13%260%336%41%50%
  • Headline nutrients are near perfect: calories 99%, fat 98%, protein 97%, fiber 95%.
  • Micronutrients are shakier: vitamin C 88%, calcium 88%, cholesterol 89%, potassium 89%. Raw garlic versus jarred salsa is typical: Jev picks the salsa, whose vitamin C is actually lower.
  • Animals: litter size 97%, gestation and lifespan 93% and 92%, but incubation only 84%.
  • Not a food-group stereotype: between foods from different groups, it's right 93% when the answer is the group usually richer in that nutrient, and 94% when it isn't. So it isn't just guessing "fruit has vitamin C".

What it means, and what it doesn't

Jev knows the numbers people talk about (calories, protein) much better than the ones they don't (minerals, incubation times). Even on easy comparisons, about one in eight answers about vitamin C or calcium is wrong. That matters for anyone asking a model nutrition questions.

These are comparisons, not exact values. Being right 88% of the time on "which has more" says nothing about how close its estimate of the actual amount would be.

Caveats

  • Fortified and processed foods. The USDA table includes cured, fortified and processed foods (cured ham carries added vitamin C). Jev has to know those too, which is fair, but some "misses" are surprising for a reason, not ignorance.
  • Record values for animals. AnAge lists maximum recorded values (the longest-lived individual, not the typical one), which can surprise even experts.
  • Clear-cut cases only. Only pairs at least twice apart are scored, so every quantity is judged on easy comparisons; the gaps between quantities would likely widen on close calls.

Jev on this experiment

Would a person find it interesting to read?
Yes68%
Does it describe you?
No52%
Would you have predicted it?
No53%
How fair is the comparison?
The comparison is reasonable
How much should a reader rely on it?
Moderately
Which caveat matters most?
Clear-cut cases only71%

Why ask this

Numbers about food and animals are everyday knowledge: does a banana or a potato have more potassium, does a parrot or a dog live longer. A model might know the headline numbers (calories, protein) well and be vague about the rest (minerals, vitamins, how long an egg takes to hatch).

People ask models exactly these questions when planning meals or checking a fact, and a confident wrong answer looks the same as a right one. Keeping every comparison easy (one value at least twice the other) shows which kinds of numbers Jev has actually absorbed, rather than how well it splits hairs.

How this was done

The people and the data

No people; the answers come from two public databases. USDA FoodData Central (CC0) gives 13 nutrients per 100 grams for thousands of foods; AnAge (CC BY 3.0) gives five life-history traits for thousands of animal species: maximum lifespan, gestation, incubation, age at maturity, litter size. That's 18 quantities in all, and 10,717 pairs at least twice apart.

What Jev was asked

Two-option questions:

Gram for gram, which has more vitamin C: raw garlic or jarred salsa?

raw garlic · jarred salsa

Each was also asked with the two options in the other order.

How it was measured

For each quantity, the share of pairs where Jev picks the larger one, with a 90% interval.

Where these questions live

12,000 questions across 26 topics of the map; the 16 biggest are shown. Each opens on the map with every question in it.

Every question

All 12,000 questions behind this result, the telling ones first: the examples the analysis points to, then the ones where Jev misses, biggest gap first.

Jev’s own answer
    Showing 0 of 12,000