I'm in the same boat. When I started thinking about LLM development as multi-year iterative training adjustments, it felt like the majority of shortcomings can be addressed.