“Roy Lee predicts that within two to three years, AI models will be multimodal, have large context windows, and be capable of reasoning over a three-hour screen recording.”