HackerTube 中文?
Field guide

Learning signals and model behavior

Follow the evidence from a training update to a deployment decision. Separate what was learned from what was measured.

Choose a question

Where does learning get a signal without an answer key? →Start with an unlabeled patch of an image, then switch to a game that requires a sequence of actions. Neither has an answer key, but finding structure and choosing a move need different signals.What improved when a game agent got better? →A rising score might reflect a better network, deeper search or a loophole. Trace what the agent observes, what it tries and why the resulting move earns a reward.How far is next-token prediction from answering a person? →Separate three questions: how an input becomes a representation, how training changes a network and how a capable text predictor becomes an assistant. An explanation at one level does not settle the others.How much does a successful demonstration establish? →A program gives the expected output. Between “it did” and “it will” lie the test distribution, formal assumptions and deployment context. Those boundaries are easy to miss under pressure to ship.