On Wednesday, a video from OpenAI's newly launched Sora AI video generator went viral on social media, showing a gymnast sprouting extra limbs and briefly losing her head during what appears to be an Olympic-style floor routine.
It turns out that the nonsensical synthesis errors in the video – what we like to call “jabberwockies” – point to technical details about how AI video generators work and how they can get better in the future.
But before we get into the details, let's take a look at the video itself.
In the video we see a view of what appears to be a floor gymnastics taking place during an Olympic sporting event. The subject of the video twists and thrashes as new legs and arms quickly and fluidly emerge and morph from her twisting and transforming body. At some point, about 9 seconds later, she loses her head, and it spontaneously reattaches to her body.
“As cool as the new Sora is, gymnastics is still the Turing Test for AI video,” wrote venture capitalist Deedy Das when he first shared the video on message on Bluesky: “Hi, gymnastics expert here! This isn't funny, gymnasts only do this when they are in extreme distress.”
We contacted Das and he confirmed that he had generated the video with Sora. He also provided the prompt, which was very long and split into four parts, generated by Claude from Anthropic, using complex instructions such as “The gymnast initiates from the back right corner and takes position with her right foot pointed backward in a B- plus attitude.”
“I know for six months after playing with text-to-video models that they have difficulty with complex physics movements like gymnastics,” Das told us in a conversation. “I had to try it [in Sora] as character consistency seemed improved. Overall it was an improvement, as previously… the gymnast would just teleport away or change outfits mid-flip, but overall it still looks downright gruesome. We hoped that AI video would teach basic physics, but that hasn't happened yet!”