A language model learns one thing at each position: the next word. DeepSeek-V3 adds a small module that also learns the word after next, and at serving time the same module drafts words for the main model to check. Then predict both words in parallel, attach a module that was never trained, or skip the check. This lesson needs JavaScript.