you do need to be less agentic when wanting to learn a language, compared to when doing original research
I think the main difference is that for language you already have a good dataset that lets you learn words, and with them, concepts they invoke. This might even be enough to explain the Sapient Paradox, how it took humans ~100K years to get culture going: they didn’t have abstract concepts for clear thinking (deliberation, agency) not directly invoked by concrete objects already in environment, and it took that long to develop them without scholars. Scholars only became productive after enough concepts facilitating their work crystallized in language/culture, and for thousands of years now they are capable of crafting new concepts not already represented in culture, much faster than culture previously crystallized them on its own. Some of these new concepts, when given names (words) to invoke them, can survive in culture without scholars, as datasets of everyday speech.
So the difficulty with original research is that you also need to craft the datasets for new concepts, there are no existing datasets to learn the concepts from. With deliberative reasoning (System 2), it takes a long time and focused effort (agency), and ideally you seek a reflective equilibrium of post-rigorous knowledge of concepts, where the episodes you can generate for the datasets of these concepts (lemmas, simply-stated solutions to simply-stated problems you can imagine, justified by proofs/constructions) are already learned as intuition, and as intuition (System 1 reasoning) provide no more insight into generating more episodes that would change the intuition (model behavior) significantly.
I think the main difference is that for language you already have a good dataset that lets you learn words, and with them, concepts they invoke. This might even be enough to explain the Sapient Paradox, how it took humans ~100K years to get culture going: they didn’t have abstract concepts for clear thinking (deliberation, agency) not directly invoked by concrete objects already in environment, and it took that long to develop them without scholars. Scholars only became productive after enough concepts facilitating their work crystallized in language/culture, and for thousands of years now they are capable of crafting new concepts not already represented in culture, much faster than culture previously crystallized them on its own. Some of these new concepts, when given names (words) to invoke them, can survive in culture without scholars, as datasets of everyday speech.
So the difficulty with original research is that you also need to craft the datasets for new concepts, there are no existing datasets to learn the concepts from. With deliberative reasoning (System 2), it takes a long time and focused effort (agency), and ideally you seek a reflective equilibrium of post-rigorous knowledge of concepts, where the episodes you can generate for the datasets of these concepts (lemmas, simply-stated solutions to simply-stated problems you can imagine, justified by proofs/constructions) are already learned as intuition, and as intuition (System 1 reasoning) provide no more insight into generating more episodes that would change the intuition (model behavior) significantly.