You know what an LLM is. Consider how it got that way: exposure, at staggering volume, to a large language corpus. Nobody taught it grammar rules. Nobody gave it a streak badge.
There is a plan many learners run, usually silently, sometimes out loud: "I'll just do Duolingo passively for now. I know it's not that great — but I'll have a passive base for when I actually want to learn the language."
Notice what's honest about it: the admission that nothing active is happening, and the intuition that a passive base is still worth building. Both correct. Passive knowledge is real capital — comprehension precedes production everywhere in language acquisition, and a learner who arrives at their "active phase" already understanding a lot starts halfway up the hill.
The plan fails at exactly one point: the quality of the passive base it builds.
A gamified app's passive base is a small curated garden: a few thousand sentences, written for learners, read slowly, recycled endlessly, organized into topic trees that exist nowhere in real speech. Spend two passive years there and what you own is familiarity with that garden — its voice, its pace, its vocabulary about ordering apples. The first native speaker you meet is not from the garden. Nothing they say sounds like it.
Now consider the machine's diet. A language model becomes fluent-shaped by ingesting the whole distribution — every register, every domain, every speaker quirk, at volume no curriculum designer would dare assemble. That's the entire trick. Not better rules: a bigger, realer corpus. And the corpora exist for human ears too: tens of thousands to hundreds of thousands of native-recorded sentences per language, real voices at real speed, with verified transcripts. Already assembled — you don't even get to procrastinate by collecting it.
Same passive intention, same effort budget, radically different asset. Two years of a curated garden buys you the garden. Two years of exposure to a large language corpus buys you the beginning of what natives actually have: a statistical feel for the language as it is genuinely used — the ear that returns serves, not the one that recognizes flashcards.
This is the whole argument, and it fits in a sentence: if your knowledge is only going to be passive for now, why not have good passive knowledge before you start speaking?
Dictation on a large corpus is still compatible with the low-commitment spirit of the passive plan. You are not conjugating. You are not performing in front of anyone. You listen to a real sentence you've never heard, type what your ear caught, see the answer, continue. Twenty minutes. The spaced repetition quietly turns exposure into retention — the part of the plan Duolingo's garden was always going to skimp on. And when the active phase finally arrives — the trip, the partner, the job — you step into it with a trained ear instead of a completed topic tree.
The model was not embarrassed to train on a billion ordinary sentences. Neither should you be.