Seems like now that we have multimodal LLMs and a virtually infinite supply of digitized video, we could get it to prompt itself about what might happen in the

Seems like now that we have multimodal LLMs and a virtually infinite supply of digitized video, we could get it to prompt itself about what might happen in the video and then tie rewards to what actually happens later in the video.