One use-case for large language models that I'm especially eager to see evolve is in assisting humans in learning languages. Once one has achieved roughly A2 in a new language, progress becomes much more rapid by holding spoken conversations in the new language. It is something which an LLM should be reasonably good at:
- Large corpus of idiomatic written and spoken language available on general topics.
- Generate prompts for a few thousand hypothetical situations to navigate in the new language, as part of the curriculum.
- Doesn't require the LLM to be an expert in a factual area. Just converse.
One could offer 1:1 instruction with the human student paired with an LLM instance. The prompt could gradually adjust the sophistication of the language as the student progresses. The LLM could evaluate the human student's pronunciation and word choice, and offer tips for how to be more idiomatic.