(Unresearched uncareful speculation.)

TL;DR: Maybe eating food that’s made of a bunch of different ingredients is bad because it prevents your gutbrain and headbrain from learning which foods correspond to which nutrients. Without knowing that, the brains can’t accurately say when to eat what.

My supporting data:

1. It stands to reason.
2. When I haven’t eaten enough protein, I feel something’s off and I’m compelled to eat more food; but I don’t automatically specifically eat protein.
3. Sometimes when I’ve exercised a lot, and have been drinking water, I still feel off. Then electrolytes tastes actively good, and I feel better after drinking them (and they stop tasting so good, or even taste a little bad). I had to learn this, I didn’t automatically know it.
4. One time I got sick and vomited after having eaten sweet potato fries with dijon mustard. A decade later, I still maybe wouldn’t want to eat sweet potato fries, and would probably avoid the combo. They probably had nothing to do with me getting sick and vomiting, though.
5. One can acquire tastes, and acquire non-distastes.

To slightly expand:

Your gut tracks various facts about what you’ve eaten. Your head tracks other aspects (e.g. how it looks, tastes, smells). Some other systems track stuff about what nutrients your body does and does not have enough of currently. These systems communicate with each other and learn, but imperfectly. So you make have intuitions about what to eat——a sense of taste——that embodies useful skill/information about what is good to eat in general, and what you should eat right now; but also, there will be missing information, and the systems can’t necessarily adapt to the situation (e.g. to large changes in your physiological state, or to novel foods becoming available).

Since a lot of this learning happens unconsciously, it won’t learn the way that you learn consciously. Conscious learning can squeeze information from a rock, so to speak: If you have something you want to understand, you can play around with it, test it, break it, put it back together. Intervening on the thingy makes it easier to see what causes what. E.g. knocking out supports from a structure until the whole structure collapses shows you which supports were crucial. Unconscious learning is (I imagine...) much broader and much shallower. It won’t squeeze anything out that’s too deeply buried, but it can make gradual progress through a lot of channels. But if the channels are too noisy, progress could be stopped. (Or maybe not; in a computer SGD works fine!)

Gutbrain-headbrain-tongue-nose-nutrient learning is unconscious, so maybe it gets confused by complicated food. If you eat a meal with 12 different ingredients, it’s hard to assign credit or anti-credit. If the system were really doing SGD this wouldn’t be a problem, but if you tell me it definitely works by SGD and your proof is that SGD is totally the vibe right now, I absolutely will judge you to be a fucking idiot. If the system does something else, such as waiting for some threshold to assign credit, then difficult credit assignment might be a problem——which food had enough of nutrient XYZ? Which food would fix it? Which of these other foods don’t have enough of nutrient XYZ?

Note also that different information comes in at different times. You see food, then you smell it, then you feel and taste it; then later, your gut gets info about the chemical makeup of the food (acidity, protein content, some other stuff I assume...); then even later you have some change in physiological state that gets registered by various senses (including direct monitoring, e.g. blood glucose levels). So the credit assignment problem is already complicated. Your gutbrain-headbrain system already has to correlate all these things if it’s going to learn anything.