Edit prediction suggests the next change based on the sequence of edits you have made, not just the next token. It triggers inline as you type and renders as a ghost overlay you can accept with Tab or dismiss with Escape.
Predictions are model-backed and respect the provider you have configured for the assistant. They are scoped to the active buffer and nearby files so suggestions stay relevant to what you are working on right now.
Disable it in settings if you prefer plain completions, or scope it to specific languages via settings.json.