Data · Garbage in, garbage out Six real runs, identical diets by weight

You are
what you eat.

Six models, same size, same number of steps, and — this is what makes the comparison mean anything — exactly the same number of characters to learn from. Only the quality of those characters changes. Every one is then marked on the same clean paper it never saw. Pick what to feed it.

score on the clean paper

characters fed

identical for every run

of which genuinely different

What that model writes — greedy, so this is the training and nothing else

All six, side by side — longer is worse

Why every diet weighs the same. It would have been easy, and wrong, to feed the bad runs less text — they would have scored worse for a reason that has nothing to do with quality. Every recipe here is the same number of characters, checked before training starts, so the only thing separating these six numbers is what was in the food. Next: what happens when it has already seen the exam.