The current state of the art is RLHF (reinforcement learning with human feedback); initially trained to complete human utterances, plus fine-tuning to maximize

The current state of the art is RLHF (reinforcement learning with human feedback); initially trained to complete human utterances, plus fine-tuning to maximize human feedback on whether the completion was "helpful" etc. https://huggingface.co/blog/rlhf https://huggingface.co/blog/rlhf