Andrej Karpathy, a prominent artificial intelligence researcher, has demonstrated a paradigm shift in how large language models (LLMs) are evaluated—moving from abstract coding benchmarks to complex, real-world software engineering tasks. In a recent experiment, Karpathy challenged Anthropic’s Claude Opus 5 to transform the opening paragraph of J.R.R. Tolkien’s The Lord of the Rings into a fully interactive 3D website, using the Three.js JavaScript library.
Instead of requesting small code snippets or static images, Karpathy provided the model with a token limit of one million (approximately $10 in cost) and instructed it to generate a complete, browser-based 3D scene. This test goes far beyond conventional evaluation methods, which typically focus on isolated tasks like code generation or text completion. Here, the model was required to think, write thousands of lines of code, manage resources, and produce an interactive product—all in one go.
From Benchmarks to Real-World Engineering
Karpathy’s approach reflects an industry-wide trend: as AI models grow more capable, engineers want to assess their performance on end-to-end workflows rather than narrow tests. The experiment underscores that the frontier of AI evaluation is shifting toward holistic, production-like challenges that reveal both strengths and persistent weaknesses.
Visual Reasoning and Scene Composition Still a Challenge
While Claude Opus 5 succeeded in generating an interactive 3D scene, the experiment also exposed critical shortcomings. According to Karpathy, visual rendering and scene composition remain difficult for even the most advanced AI systems. Additionally, evaluating the quality of such outputs is problematic: while code can be automatically tested, visual data often requires human judgment to assess accurately.
Raising the Bar for AI Evaluation
Karpathy’s experiment illustrates that expectations for LLMs are expanding beyond answering questions or writing code. Developers now want AI to build complete applications, reason across multiple steps, and generate interactive digital experiences with minimal human intervention. As AI capabilities continue to improve, real-world engineering challenges like this one will play a bigger role in measuring model performance, highlighting not only what these systems can accomplish but also the technical gaps that remain before they can consistently deliver production-ready software.
