We're missing (at least one) major paradigm for LLM learning