The problem is training. To train an AI assistant, you have to use its LLM (Large Language Model, previously trained on Terabytes of text) to generate a LOT of

The problem is training. To train an AI assistant, you have to use its LLM (Large Language Model, previously trained on Terabytes of text) to generate a LOT of output, which is then ranked/labeled/evaluated by humans as part of the training loop. If you want competence in a given area, I guestimate you'll need a century or more of evaluation time from humans qualified in that area, because of the inefficiencies of current training algorithms. If you don't have competent experts, then the tendency will be towards BS that can fool the non-expert.