I'm curious about it too. I think there are two bottlenecks, one is that training a relatively large LLM can be resource-intensive (so people go for RAGs and other shortcuts), and making it finetuned to your use cases might make it dumber overall.