The pelican cyclist’s feet are clipping, but the claim that video models are doomed has already spread

Chetaslua posted a Three.js animation of a pelican riding a bicycle, with a bold caption: one shot, output by the Fable 5.5/Opus-Next router, proof attached.

The pelican stands on a red bicycle, with bread and green vegetables in the basket, against a background of seaside railings and boats. The feathers, lighting, and chain details are all there. The tweet also included an even stranger test: ask where Karpathy went in 2026, without internet access or memory tools, claim that only the new Fable knows the answer, and then use that as a prompt.

It is not just an isolated bit of showing off. The pelican riding a bicycle is often used in AI circles to test a model's ability to generate 3D scenes and code, and some in the comments call it the best test going. Chetaslua then posted a comparison with the same prompt: another output was a four-panel cartoon with a scarf, a lighthouse, and a beach, in a warm art style. His judgment is that Fable automatically goes for hyperrealism, detail, and physics without extra requirements, and that at this rate of updates, video or image models won't be needed in the future.

That claim spread faster than the animation. The comments were not all praise. Some asked where to test it, how much it costs, and whether a reference image was provided; some said the style did not match the instruction; the two most specific were that the pelican's feet clip through the pedals and that there is a pause in the middle of the pedaling motion. Some users pointed out that if you exclude the cartoon option in the prompt, other models might also achieve this level of realism. The author responded by posting proof screenshots, but the clipping and pause were not covered up by the video's lighting.

What is worth examining is where the difference comes from. Under the same prompt, one model defaults to cartoon and another to realism; this may just be a default style preference, not proof of crushing superiority in understanding physics. A Three.js scene is code generation plus real-time rendering; a video model generates pixels. The two overlap, but they are not the same contest. Saying video models are doomed is a judgment from the tweet, not a fact already proven by this demo.

Replies also included screenshots of other checkpoints; their authenticity has not been verified.

The comments section is more honest than a launch event. No matter how fine the feathers are, the detail of the feet clipping through the pedals still carries more information than the term AGI.

Links: main post https://x.com/chetaslua/status/2116882694026269148 quote https://x.com/chetaslua/status/2116795545205543398