People were impressed by the raw fact that a model can grind through this at all. A year or two ago, asking a language model to turn prose into a coherent 3D scene with motion, camera changes, and some physical consistency would have failed much earlier. That said, almost nobody thought the result was actually good. The recurring verdict was “dancing bear.” It is notable that the bear is dancing, not that the dance deserves praise. The output looked like social-media demo bait, not something useful, beautiful, or fun.
The sharpest practical point was about benchmarks. The pelican test was never valuable because pelicans matter. It was valuable because it was cheap, quick, reproducible, and easy to compare across models. You can glance at one SVG and immediately spot failures in spatial layout, mechanical plausibility, and code generation. A long custom Three.js render burns far more tokens, takes longer to inspect, and introduces too many confounders. A benchmark should either stay simple enough to compare model-to-model, or be tied directly to your own real tasks. Several people said the right replacement for pelicans is not “bigger spectacle” but
evals on your own traces with a cost, latency, and quality profile.
A second theme was that the demo leaned heavily on prior cultural material. Using Lord of the Rings makes the result look smarter than it is because the model has likely absorbed endless text, stills, adaptations, and fan art. Bag End looked like the Peter Jackson films rather than a fresh interpretation of a sparse paragraph. That does not make the coding feat fake, but it does make the example a poor read on general scene understanding.
The other strong thread was that these visual coding demos are exposing the same limitation people see in code: models can place all the right ingredients on the page while still failing basic functional relationships. That was compared to pinball layouts with unusable flippers, pelican bikes with impossible steering geometry, and AI-generated games that look presentable in clips but fall apart when you actually play them. In other words, the frontier has moved from “can it emit recognizable pieces” to “can it arrange them into a working system.” Many commenters think that is the more honest benchmark now, whether the medium is SVG, 3D scenes, websites, or games.
twitter.com