Large language models deal with uncertainty and compounding it by continuously asking for hallucinatory answers can cause issues. Tests are being conducted to find ways to limit the length of video conversation to prevent manipulation of the model.