Andrej Karpathy: LLMs Can Generate Complex LoTR Program Rendering
Andrej Karpathy stated that testing Opus 5 with a budget of 1 million tokens generated 5,500 lines of JS code for procedural rendering based on the first segment of "The Lord of the Rings".
The model took about 2 hours to orchestrate polygon asset coordinates and write animation code, resulting in a rough but complete presentation of the story; he noted that LLMs have infinite endurance, making custom tasks that "no one would complete manually" feasible, and he envisions generating instantly accessible hyper-customized game worlds.
In the AI and content creation market, generative models lower the development threshold for complex interactive worlds, with funding flowing towards long context and multimodal capabilities; traditional game studios are pressured by the efficiency gap in customization, while models with native video and game perception will seize advantages in self-auditing and iteration.
Source: Public Information
ABAB AI Insight
As a former head of AI at OpenAI and Tesla, Andrej Karpathy has long advocated for moving from simple benchmark testing to evaluating real complex tasks; this time, generating a complete JS rendering program from "The Lord of the Rings" text continues his emphasis on model endurance and autonomous orchestration capabilities, having previously showcased breakthroughs of LLMs in coding and visual tasks.
On the capital front, the long context budget (around $10) makes one-time generation of hyper-customized interactive worlds possible, shifting resources from manual design to model autonomous construction and iteration; the motivation is to unlock a new application layer for "on-demand generation of GTA-like experiences," reducing fixed costs in game and simulation development.
Analogies can be seen in the evolution from early GPT generating simple game code to complete programs, as well as the current gap in multimodal models in video understanding. The industry is in a transition phase from "text/code generation" to "perceptible and interactive world building."
Structural judgments belong to 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 that require manual screenshot feedback, reconstructing the content production chain.
ABAB News · Cognitive Laws
As generation costs approach zero, verification becomes the new bottleneck.
Endurance unlocks customized worlds more than IQ.
If one cannot perceive their own output, true iteration is impossible.