Tagged RLHF · show all articles
Constitutional AI keeps RLHF's reward-model-and-reinforcement-learning core intact and swaps the source of one training dataset, a document standing in for a person's judgment. That swap is narrower, and its limits more specific, than the name suggests.
A deployed Claude model is a pretrained prior, a post-training pass partly shaped by written principles, a handful of per-request settings, and a documented evaluation gate — four separable things, not one object, and each changes the answer differently.
Reinforcement learning from human feedback does not install correctness. It fits a scalar model of what raters approved of and then pushes a policy up that surface — which is a different thing, in identifiable places.