Andrej Karpathy, former head of AI at OpenAI and Tesla, advocates moving from simple benchmark testing to evaluating real complex tasks, such as generating a complete JS rendering program from "The Lord of the Rings" text, emphasizing model endurance and autonomous orchestration capabilities. On the capital front, a long context budget of around $10 makes one-time generation of hyper-customized interactive worlds possible, shifting resources from manual design to model autonomous construction and iteration, with the motivation to unlock a new application layer for "on-demand generation of GTA-like experiences" and reduce fixed costs in game and simulation development. Analogies include the evolution from early GPT generating simple game code to complete programs, and the current gap in multimodal models in video understanding. The industry is transitioning from "text/code generation" to "perceptible and interactive world building," a structural judgment of technological substitution: as generation costs approach zero, the bottleneck shifts from "can it be created" to "can it self-verify and optimize," with models possessing native game/video perception replacing processes requiring manual screenshot feedback, reconstructing the content production chain. As generation costs approach zero, verification becomes the new bottleneck; endurance unlocks customized worlds more than IQ, and if one cannot perceive their own output, true iteration is impossible.