Gemini 4 Carbon wrote its own CUDA path tracer and rendered a minute-long animation

A typical "AI wrote the code" demo looks like this: the model builds a landing page, a Snake game or a parser. This story is different. According to the demo's author (the fofrAI account on X, retold in the Telegram channel "Metaversishche i AIshche"), the Gemini 4 Carbon model went on its own from an empty folder to a finished video: with geometry, lighting, physics, camera work and music.

What exactly the model built

The list from the post looks like this.

First: a custom 3D renderer in C++ and CUDA. It uses path tracing, that is, an honest calculation of light, reflections and shadows by tracing rays on an NVIDIA GPU. This is not a call to a ready-made "Render" button in Blender, but code that itself calculates how light bounces off surfaces.

Second: physics via MuJoCo. This engine is familiar in robotics. It calculates gravity, collisions, friction and the interaction of rigid bodies.

Third: the scene. The model designed a Rube Goldberg machine in which balls and levers set each other off in a chain. It then built a 3D studio and a continuous camera flythrough.

Fourth: rhythm. The camera movement was synchronized with a track from Lyria 3.5, Google's music model.

Why this is not just a pretty picture

The key figure here is 60 seconds of simulation at 120,000 steps, or 2,000 steps per second. There are no keyframes. Objects move because simulated forces push them, not because someone drew a trajectory. For a Rube Goldberg machine this is essential: if even one link in the chain is calculated incorrectly, the ball won't reach its target and the whole reaction falls apart. You can't fake such a clip by eye; the physics either works or it doesn't.

The second point is the breadth of the task. Graphics programming, GPU code, physical simulation, scene composition, camera work and timing to sound all came together here. Such tasks are poorly checked by standard benchmarks: the model knows each part individually, and the difficulty lies in making the parts fit together and not break down at hundredths of a second.

There is also a separate irony in the NVIDIA Omniverse joke in the headline of the original post. Omniverse was sold as a platform for simulation and 3D pipelines. Here the model assembled a pipeline similar in purpose from open and ready-made components. And it still runs on CUDA, so a scenario like this is, if anything, good for NVIDIA hardware.

What we don't know

Caution is needed here. The source is a Telegram post retelling tweets, and it is a single demo clip, not a reproducible experiment. The material doesn't show what the prompt looked like, how many iterations there were, whether a human intervened or fixed errors, or how much the run cost. The image quality is also unclear: a path tracer written "from scratch" can produce noise and artifacts that a mature renderer wouldn't. Finally, "wrote it itself" doesn't mean "came up with it itself": MuJoCo and CUDA are ready-made tools, and the model put them together.

So the claim from the post that "3D artists are no longer needed" is a joke, not a conclusion. An artist still decides what is worth showing at all, and a studio pipeline has things the demo doesn't show: assets, animation revisions, client approvals.

What to do with this in practice

For developers and product managers, the takeaway is more modest but more useful. First, it's time to stop measuring models with small tasks. If a model assembles the combination of "GPU code plus physics plus visuals plus sound" in one go, prototypes of simulators, visualizations and training environments can be built much faster. Second, physics engines like MuJoCo are a convenient verifier: the result is either plausible or not, and that helps catch the model's errors. Third, if you build tools for 3D, games or robotics, it's worth running the same kind of long scenario on your own stack and seeing exactly where it breaks.

A demo is not proof, but it is a good reason to update your test suite. If Gemini 4 Carbon really does this on demand, and not through a lucky coincidence, what will change most noticeably is the barrier to entry for complex graphics prototypes.