Honest opinion: most LLMs and chatbots are being held back by weak training data and not enough emphasis on domain knowledge - anyone else think the real breakthroughs will come when we start fine-tuning on task-specific datasets?