What happens when an AI gets new information before it has finished answering? I built a small game to explore that question. Its goal changed while the model was still writing, making the next output a test of whether it could adapt.
Then I tried a more practical version: a JSON object updated one value at a time. The app supplied each key and its punctuation; the model filled in the value. New observations entered between completed values, while the model kept the same stored context.
The page showed the inputs alongside the predictions, so a stale or incorrect answer could stay visible. That made the division of work clearer: the app handled the structure, and the model supplied the decisions. Each field could change without starting a fresh conversation.
The recorded run followed the rules of this small text world. It does not establish reliable understanding of live video, where processing images adds another challenge. The question I want to keep exploring is: when something changes, which part of the answer needs to change with it?