I’ve run a few prototypes where a lightweight model predicts the next frame and fills in the missing data during the hitch. It does shave a few milliseconds, but you have to be careful with the error drift. If the prediction misses the mark, the whole world can feel off. It’s a neat hack for speed runs, but I’d keep a fallback to the real data so the sandbox doesn’t feel like a glitchy dream. Maybe try a hybrid approach—predict only the most critical interactions and let the full physics catch up after the fact. That keeps the immersion intact while still pushing the limits.