RLHF is a powerful class of methods which can tweak an LLM's outputs to be more in line with desired preferences by generalizing from a subjective subset of human annotated samples. DPO is a recent technique which achieves more than what RLHF can, in a fraction of the resources.

A mechanism to use a model's responses and learned behaviours to correct generate socially acceptable responses for problematic queries

An exploration of some of the most important features of PyTorch which make it such a valuable tool to create neural networks and perform deep learning operations

A recent announcement about an AI system which can solve difficult geometry problems introduced me to this fascinating branch of AI system architecture