Everything I've learned up to this point regarding language and language-learning is that the human first has an initial phase of gathering lots of 'chunks' of information. Then, with time, they learn how those chunks fit together. Then, their brain processes it all in the background, during sleep and breaks and recreation and rest. Then, they can finally think about what they've learned, the 'grammar' of the language, per se, the layer of conceptualizing and thinking about the thing that they now know inherently to some extent.

In Japanese, they might realize は ('wa') is its own chunk of information-- it comes before context, but that understanding might not come for years, at first you simply recognize the sound and the fact that は is it's own distinct word/chunk like あなた or はい.

Then, the brain starts to group things semantically in a vector sort of space, like an LLM does. は isn't just a chunk on its own, it has certain attributes:

• it goes near the start of sentences
• it denotes context
• there's a little pause after people say it
• it sounds like 'wa' but it isn't わ
• it's written like は but it isn't pronounced 'ha'

At no point does the baby say these things in their mind, in words, they just exist as hazily-understood facts, neurons with hesitant, impartial definitions.

The majority of evidence with regards to language learning seems to imply that this is the only way people learn language. You can lead a baby to Chinese but you can't make him drink - the only thing that's going to teach the baby Chinese is giving him lots of exposure to data of people speaking Chinese words in different contexts. Yelling grammatical rules at the baby isn't going to help.

Adult language learning education on YouTube and other sources that center around people who simply want to learn a language the fastest, most effective way have started to trend towards a realism that apps like Duolingo which claim to teach languages like it's some bitter pill to swallow through a TTS-esque voice intoning 'Agua' and then waiting for you to click on the glass of water and making a happy ding noise and then serving you a 5 minute video advertisement are at best extremely slow, and at worst nearly completely ineffective.

All of the respectable polyglots at this point seem to point to Comprehensible Input as not just the best way but literally the only way. Their messaging boils down to the following:

• You must, of your own volition, comprehend a piece of information to some extent.
• The more exposure you have to the sorts of chunks this information comes in, the better. This means if you want to understand math, you need to know what numbers look like, and how they are pronounced, and the operators people use with them. If you want to learn Melee, you need to know what a wavedash is called, and then what it looks like, and then how to do it, and then when it ought to be done. If you want to learn Chinese you need to hear the tones enough to distinguish them.
• The more self-guided, fun, or necessary this learning is, the more you actually learn. If you have zero interest in actually learning the subject, your brain seems to basically just shovel the information directly into the garbage chute. You'll remember having studied it, you might remember some of the chunks of information, but at some point very little remains, I assume this is what happens when people cram: they have very little interest in the topic, they commit a bunch of words about it to their short term memory, and they learn ostensibly nothing. You can see this in the language learning community: people saying you have to do something fun in your target language, like watching soaps or anime or reading comics, but also in the programming community, where people always recommend learning by trying to complete a project that you actually care about rather than empty tutorials or fizzbuzz. I think there's a difference between Leetcode and Duolingo, though: Leetcode (hopefully) forces you to comprehend and demonstrate knowledge, whereas Duo simply wants you to press the right button or rearrange the word puzzle pieces in the right order based on rote memorization and context clues.

My question is: is this just how people learn everything, generally? What does it mean to learn something if not to comprehend an input? Maybe this is something everyone knows and I'm just now coming into awareness about it. But it seems like a big key to solving the problem of knowledge transfer and learning in general.

This post has been written on my phone by someone who is still ultimately a newbie to LW so I apologize if I haven't learned the language of LW fluently enough yet.